2026-11-15
MES and ERP Integration: Why Done Means Something Different in Each System

Show notes
Done can mean something different in the MES and in the ERP, and integration has to resolve that. In this episode of the iFAKT series, we examine how to align status, quantities and events across both systems. You will learn how to design integrations that avoid conflicting numbers. Who it is for: Integration architects, MES and ERP owners. Become a supporter of this podcast: https://www.spreaker.com/podcast/m365-fm-microsoft-365-azure-cloud-power-platform-ai--6704921/support. This episode includes AI-generated content.
Host: Mirko Peters
Transcript
Read the full transcript
iFAKT - MES and ERP Integration - Why Done Means Something Different in Each System 00:00:00 Speaker: Your order is complete in the manufacturing, execution system or Mess, but enterprise resource planning or ERP still shows it open so stock stays unavailable while your planner chases answers and the customer waits for a shipping promise. Both systems might be right. They may use the same word for different business milestones, and will trace that disagreement from factory events to shared rules that people can explain. Let's follow one hypothetical machining order. We'll move with it through production quality and stock posting to understand why finished work doesn't always mean you have parts you can ship. Imagine your plant receives an order for one hundred machined parts. For this example, those parts need machining and inspection before the warehouse can receive them as stock available for the next step or a customer shipment. The operator finishes machining the last part and records the work from that operator's point of view. The assigned job has reached its end because every part in the order has passed through the machining operation inspection hasn't finished yet. The parts sit in a marked holding area, and although someone walking past can count, the full order quality hasn't decided which parts meet the requirements. Now your planner checks whether the order can support tomorrow's shipment. The Mez reports that machining has finished, but ERP doesn't show the stock the planner needs, so the planner calls production to ask where the parts are. Production gives a reasonable answer. They're finished. The planner then calls the warehouse where the person responsible for dispatch checks received stock and finds nothing available against that order yet. Nobody needs to have entered anything wrong for this to happen. Each person answers from the work they own, while the planner needs an answer that spans the whole journey from machining to stock that can actually leave. Put yourself in the operator's position for a moment. Another job is waiting. The machining work on this order has ended, and keeping that operation open would suggest there's still machining left to do. Quality faces a different question. The inspector needs to check the parts and record the outcome. So a finished machining operation gives them work to assess rather than permission to release the order. The warehouse has its own task. Staff need to know which parts they can receive and where those parts belong, before they can pick them for a shipment or issue them to another order. So the planner asks a more useful question how many can I use? That question needs a quantity and a purpose, because parts ready for another production step may still need work before they're ready for a customer. Let's let inspection move forward in this hypothetical order. Most parts pass inspection. A small group needs rework and the accepted parts can continue to wordstock while the others return for more work. The order now has more than one path through the plant. Some parts wait for warehouse receipt while the rework group needs attention from production. Before quality can assess whether those parts also meet the requirements. Your planner doesn't necessarily need the whole order at once if the accepted quantity covers the next requirement. A partial receipt might support that requirement while the remaining parts continue through rework depending on the plant's rules. But the receipt hasn't been posted yet. The accepted parts may already be waiting in the warehouse area, while the person preparing the next shipment still can't find available stock against the production order. This creates another phone call. Quality confirms that the accepted parts can move forward. The warehouse checks the pending receipt, and the planner waits for the recorded stock position to catch up. Notice how the question changes during that conversation. At first, the planner wants to know whether production has finished, but now the immediate concern is whether the accepted quantity has reached stock in time for dispatch. Meanwhile, production works through the remaining parts. Once rework finishes, those parts return for inspection and their next destination depends on that inspection rather than the earlier report that machining had ended. Eventually, the remaining acceptable parts reached the warehouse and the receipts are recorded. The people responsible for the order can then finish their remaining work and close it under the rules that apply in this plant. I wouldn't begin by asking which system tells the truth. I'd ask the operator, inspector, warehouse team and planner to describe the exact point they're referring to when they call this particular order done. Their answers give us somewhere concrete to start. The operator may mean the assigned work has finished, while the planner may mean enough accepted stock is available to meet the next requirement without another call. That's the distinction we need to name before we ask me as an ERP to agree, we need to define which kind of completion each team expects the systems to record. I'd start with the status definition in your actual system. A label such as complete can refer to a single operation, an execution record, or a broader production milestone. Depending on how your Ms. has been configured. That scope deserves a plain sentence people can agree on. For example, Operation Complete might mean that no further work is expected on that operation under the current production instructions, without saying anything about later approval. Ask what the status actually permits. Can someone still record work against that operation, or does completion stop further entries until an authorized person changes the status back through an approved process? Those details belong in the definition to a status that simply records progress, means something different from a status that prevents further work, even if both carry the same label in different plants, ERP needs the same level of care. Depending on the product and process, its production statuses may refer to recorded output, delivery against the order, remaining transactions, or permission to close the order. A production receipt records output into inventory under the applicable rules. That receipt may place the output into a restricted stock category. So receipt alone doesn't necessarily establish that the quantity can be used or shipped. Delivery status answers another question. It can describe how much output has been received against the production order, rather than whether a customer delivery has left the warehouse or reached its destination. Even the word delivery needs a qualifier. When someone describes an order as fully delivered, ask whether they mean production output received or a customer shipment, because those statements refer to different business processes, then there's settlement where the ERP supports that process. Settlement generally deals with transferring costs or balances according to configured accounting rules. Rather than confirming that the physical work has just reached its end, financial closure needs its own definition. An order might still allow cost entries after production ends, while a later closing status restricts further postings according to the rules your finance team applies. I wouldn't impose one sequence on every manufacturer. Some processes receive output before quality release, others require release first, and the system settings need to represent the process that your plant has actually approved the milestone. Names can still stay clear. Physical work finish describes execution quality release describes the relevant approval. Stock posted, describes the inventory transaction and financial work. Closed. Describes the accounting boundary for that order. Each statement needs a scope. Quality released for one portion of an order doesn't describe the remaining portion. Just as physical work finished on one operation doesn't describe every operation attached to that order. Keep those full names in the shared language. If a meeting uses the word complete, the next question should establish which milestone people mean. Rather than letting each team silently supply its own definition, I'd write a short definition beside each milestone. It should state what has happened, what the milestone covers, and which further business conditions remain outside that statement. Using words the responsible team would actually use for physical work finished. The definition might name the relevant operation for stock posted. It should describe the inventory receipt that has been recorded, and make clear whether that stock still carries restrictions under the plant's process. Use those terms wherever people make decisions. A shift handover and a management report shouldn't shorten quality release to done. If that change invites someone to read it as permission for a different business action, you can still have an overall completion label, but give it an explicit purpose such as production, administration complete and define the milestones it requires. Instead of asking one word to carry every department's expectations. That also changes what agreement between systems means. Mez can correctly record an execution milestone while ERP records a later milestone is pending, provided both statements have clear definitions and remain consistent with the underlying work. The goal isn't identical labels everywhere. It's a shared understanding of what each label allows someone to conclude without silently adding quality, approval, inventory availability, or financial closure to a statement about execution. Collapsing those milestones changes the decisions people take. If one completion flag gets read as permission for every downstream action, someone can act before the condition they actually depend on has been met. The first cost can arrive before anyone notices a discrepancy. A customer service colleague sees a completion report, assumes the output can leave the plant, and confirms a dispatch date without knowing that release is still pending. That promise changes someone else's plans. The customer may book, transport or schedule assembly around that delivery, while your dispatch team starts preparing work based on stock that hasn't yet cleared the required checks. Call that false availability. The parts exist, but the business conclusion attached to their completion creates permission that nobody responsible for release has actually given, and the customer receives that conclusion as a promise. The mistake doesn't require anyone to bypass a rule deliberately. A person can follow the normal planning process and still commit too early. If the information they receive hides a condition that would have changed their decision, then the promised dispatch date gets closer. Someone discovers the restriction. Customer service has to revise the commitment, and production faces pressure to solve a delivery problem that started with an assumption about status. The opposite error creates a different kind of pressure. Accepted output can remain absent from posted inventory, leaving the planner to respond to an apparent shortage. Even though the plant already holds parts that could satisfy the requirement, a shortage can trigger more work. Depending on the planning process, someone might request another production run or move a different job aside, using scarce capacity to cover a gap that exists only in the recorded stock position. That decision affects orders beyond the disputed one, a job that genuinely needs machine time may lose its slot because another job appears short, and correcting the inventory view later won't automatically restore the time already spent. This is why planners keep their own working records. A spreadsheet can hold the accepted quantity confirmed by quality. A note about a pending receipt, and the name of the person who checked the parts this morning. That spreadsheet fills a real gap. It brings together details the planner needs for a decision. Often with enough local context to explain why the official stock position doesn't yet match what people know about the order. I wouldn't dismiss that as resistance to change. If your system can't answer whether a requirement can be covered, calling a trusted colleague and writing down the answer can be the most responsible option available. but that answer has a limited life. A note that supports one planning decision can become unsafe when someone copies it into tomorrow's plan, without checking whether the stock remains available for the same purpose. Local knowledge also depends on who's working, and experienced planner may know which pending receipts usually clear before dispatch. While a colleague covering the shift sees the same shortage and takes a different course, manual checks reduce uncertainty one order at a time. They also consume the attention people need for real exceptions, because every disputed status sends someone back through calls and records to establish facts that should already support the decision. Over time, checking becomes the normal process. People stop treating the system as an answer and start treating it as a prompt to ask someone else, even when the information happens to be correct. Trust then spreads beyond the original error. If a report repeatedly suggests stock that can't ship. Users may begin doubting its other figures too, including figures that accurately describe machine activity or completed work. Accurate signals can still support the wrong conclusion. A machine count may faithfully record output, yet a report that presents that count as available. Stock asks the signal to answer a business question it never measured. More frequent updates won't repair that interpretation. They can simply repeat the wrong conclusion sooner, giving users a more current reason to distrust the report, whose source measurements remain perfectly sound. So I traced the decision that went wrong, which information caused the early promise or unnecessary production request, and what would the planner have needed to know to choose a different action at that moment that gives the investigation a practical starting point? Did the information mean the wrong thing, or did the right information reach the decision too late, after people had already committed capacity or promised a shipment, I'd separate that investigation into distinct failure types before changing anything. The wrong meaning, a missing message, a late message, and a rejected transaction can leave similar symptoms, but each points to a different repair. Start by asking whether the expected information exists at its source. If nobody recorded the relevant business fact, searching the connection for a lost message won't explain the gap because there may have been nothing to send. Then check what actually left the source a recorded fact and an outgoing message aren't the same evidence, so you need to establish whether the integration captured that fact and passed it onward. A missing message needs a precise location. Perhaps the source recorded the work but produced no outgoing message. Or perhaps the receiving side has no record of a message that the sender believes it handed over. Late means something different. You can find the expected message. Its meaning matches the intended action, and its journey is still in progress rather than absent from the exchange altogether. Rejection means the receiving application reached a decision. The information arrived, but the application refused the requested transaction under its business rules, leaving the intended change unapplied even though communication worked as designed. Consider a connection that passes every technical health check. In our hypothetical plant, the integration reads Operation Complete from mess and translates it into a request to close the production order in ERP without checking that those milestones match. Every message could arrive promptly and without a technical error. The fault sits in the translation rule because the source describes the end of an operation while the receiver gets instructions for broader business action. That case needs a rule change, not a faster connection. I'd compare the source statement with the action requested in ERP, then ask the responsible business teams whether the first statement actually authorizes the second action. Now take a different case involving a production receipt. The source information correctly supports that receipt, and the integration prepares the right request. But the request waits in a queue before ERP can process it. Here, the meaning can remain untouched. The investigation needs to establish why the request is waiting, whether that delay falls within the agreed operating expectation, and when the receiving application can process the pending work. Keep rejection separate from that waiting state. If ERP refuses the receipt because the order no longer permits that posting, the request needs attention to that business condition rather than more time in transit. An acknowledgment can also mislead the investigation. A response that confirms receipt of a message may only confirm that an interface accepted it for processing, while the actual business transaction still awaits a decision, read the acknowledgments definition before treating it as success. Ask whether it proves a rival acceptance for processing or a completed business posting, because those answers place the investigation at different points in the journey. I use one disputed transaction to establish the evidence. Begin with the source record and identify the exact fact recorded there, including its scope. Rather than starting with a general claim that the order hasn't synchronized. Next, inspect the message that represents that fact. Check whether its content preserves the source meaning or changes it into a stronger claim, particularly where a local status becomes an instruction to the receiving application. Then examine the rule that consumes the message. Find out which action that rule selects, what conditions it checks, and whether the receiving application accepts or refuses the action that the integration asks it to perform. Finish with the resulting business state. An interface log may report successful processing while the order remains open for a valid reason. So compare the actual business outcome with the specific outcome the request intended. That path narrows the repair without guessing at the cause. A translation error belongs with the mapping rule. Awaiting request belongs with delivery, processing and a refused transaction belongs with the receiving application. Stated reason. There's one assumption underneath that whole investigation. Before comparing the source fact with the receiving state, you need evidence that both records refer to the same piece of work, rather than records that merely look related. Which order did each system actually identify? Even a careful comparison can point to the wrong fault if the message refers to one order or operation and the investigator checks another at the receiving end, an order number needs a namespace. If two plants each issue order forty two thousand, sending that number alone leaves the receiver without enough information to choose which plants work. The message describes include the company and plant in that identity where your business separates legal companies from production sites. Keep both distinctions because a site name alone might not identify the company that owns the production order, you don't need one numbering scheme across the whole business. You need an agreed way to identify each order across the integration boundary, with its local number preserved, so people can still find it in the source. The operation needs its own scope to operation twenty on one order doesn't identify operation twenty on another, and a receiving system may use a different internal reference for the corresponding step on that same order. Then identify the particular execution you're talking about. If a team performs an operation in separate runs, the integration needs to distinguish those runs rather than treating every report against that operation as the same execution. Consider a morning run and an afternoon run. Both can belong to the same production order and operation, but each needs a stable reference. If people must connect reported work to the run that produced it. Resource names create another identity trap, a machine that operators call Myl7 might carry an equipment number in Mez and a different resource reference in ERP with neither name matching its physical asset. Tag names can change while the machine stays put. If maintenance renames that machine after a workshop move earlier, records still need to point to the same asset without silently inheriting the identity of its replacement. Keep the stable identity separate from the display name, maintain the relationship between system references, and retain enough history to explain which machine a local name referred to when someone recorded the work against it. Splitting an order adds another layer. Suppose the plant divides the machining work between two child lots, while ERP continues to track the requirement under one parent production order with its original reference, those child lots need explicit links to the parent. Each execution should identify the child lot it belongs to. So a receiver doesn't mistake a report from one branch for a report covering the whole order. Partial execution also needs a distinct identity. If a team pauses one run and starts another against the remaining work. Define whether those records continue the original execution or create separate executions linked to it. Merged work needs relationships in the other direction. When one execution handles work associated with several orders, retain those associations rather than choosing whichever order number fits into a field that only accepts one reference. Rework deserves the same care. If your plant creates a separate rework order, link it to the original order and affected lot while keeping the Rework Order's own identity intact for everyone who manages that work. 00:17:39 Speaker: Someone has to maintain these cross-references. I'd assign that responsibility to a named role with a clear process for adding relationships. When work splits, mergers, or moves into a separately identified rework order. A mapping table needs more than technical access. Whoever maintains it needs evidence for the relationship, approval where required, and a record of the change. So another person can explain why. Two references now point together. Avoid silent replacement of an existing mapping record. Who changed the relationship when the change took effect, and whether it applies only to future work, or also corrects references attached to earlier executions. Unknown references need a defined response. When an incoming report names a child lot the receiver can't identify, hold it for resolution, and give the responsible person enough context to establish the missing relationship. Don't guess from a similar looking number. A controlled exception keeps the identity gap visible, while an automatic match based on a familiar prefix can attach real production work to an unrelated order without anyone noticing. Once those references resolve, you can locate the work consistently. The next question concerns what each system believes that identified work requires. Because matching an order and operation doesn't establish agreement about the instructions behind them. Shared identity gives you a common reference point. Two systems can point to the same order and still carry different versions of the product or route. Which brings us to the master data behind that order. Master data defines what the plant is supposed to make. It connects the product to its required materials, production steps, permitted resources, units of measure, and approved revisions before anyone starts reporting progress against that work. Start with the product definition behind the released order. A matching product code won't settle whether both systems expect the same material grade, drawing, revision, or manufacturing method for the work currently on the floor. Materials need that same level of detail if an approved material substitute appears in one system, but the other still expects the original material, a valid execution record can fail the receiving system's checks. Resource definitions describe where the work is permitted to run. A route might require a particular machine capability or approved resource group, so a technically available machine doesn't automatically qualify for every version of that process. Units belong in this agreement as well. The products production unit, material consumption unit, and inventory unit can differ, but those differences need approved definitions rather than assumptions buried inside an interface mapping. The route connects those definitions to the sequence of work. For our machining example, suppose the released ERP route contains machining followed by inspection, while the Mas instructions also require a deburring step between those operations, that difference changes what finishing machining actually means under one route. Inspection comes next, while under the other production still owes a separate step before the parts should enter the inspection process at all. A status translation can't settle that disagreement. The teams first need to establish which route governs this order, then correct the conflicting definition through the approved process rather than hiding the missing step. Route comparisons need more than matching operation numbers. Check the purpose and sequence of each step, including whether a task belongs inside another operation or requires a separate completion record under the approved instructions. I would compare the orders released instructions, not just today's route. Many systems copy or derive production instructions when an order is created or released, though the exact behavior depends on the application and its configuration. Later engineering changes create a timing problem. A new route revision might become valid today, while an order released yesterday still carries an earlier approved revision that remains valid for that particular work. The newest revision isn't automatically the right revision. You need the rule that determines which revision applies, whether that's based on order, release, planned production date, lot assignment, or another approved condition in your process. An effective date needs a clear business, meaning it should explain which work receives the change and how. Work already released or underway must be assessed rather than merely recording when someone edited a master record. Suppose engineering adds a cleaning step after machining. If that change applies only to newly released orders, copying the new route into every active order could change instructions that production already started under an earlier approval. Other changes may require action on work already underway. In that case, the responsible teams need an approved decision about the affected orders with revised instructions where required, rather than letting an overnight data update decide. Keep evidence of the revision assigned to the order. That gives you something specific to compare with the execution instructions. Even after the Product Master and Standard Route have moved on to newer approved versions. Then validate against that assigned revision at the integration boundary. An incoming record should refer to a revision the receiver recognizes and can check against the order, rather than merely carrying a product code that looks familiar. Unknown revisions need a controlled outcome depending on the interface. Reject the request with a clear reason, or hold it in quarantine until the responsible Master Data team resolves the mismatch and approves further processing. Avoid falling back silently to the current revision. That shortcut replaces missing evidence with an assumption, and the receiver may then accept a record against instructions that never governed the work being reported. Once the definitions agree, quantities still need careful treatment, both systems might recognize the correct product and revision, yet attach different units to its output, leaving an apparently simple number dependent on a conversion nobody has checked. A quantity needs more than a unit attached to a number. You also need to know whether that number counts, individual pieces, complete packs, material weight, usable length, or the output of a batch under a defined measurement rule. Those measures answer different questions. A batch identifies a body of work, but saying one batch finished doesn't tell the receiver how many pieces, kilograms or meters that batch actually contains. Packaging adds a distinction that's easy to miss. A pack can represent a fixed number of pieces or a physical container whose contents vary. And those meanings require different rules when converting reported production into inventory. Take a separate hypothetical packing example inspection accepts ninety six pieces, and the integration sends a quantity of ninety six to an ERP interface that expects its stock quantity in packs of twelve pieces each. The intended receipt is eight packs. If the receiver interprets ninety six as packs instead, it records twelve times the intended stock. Even though nobody mistyped the number reported at the source, the conversion needs an approved source for this product and pack definition dividing accepted pieces by twelve produces the stock quantity, but the interface shouldn't apply that factor to other products just because they also use packs. Packaging changes can change the factor. If another approved pack contains a different number of pieces, the conversion must use the applicable pack definition rather than a single factor stored against the word pack. Weight conversions need their own checks. Converting grams to kilograms follows a fixed unit relationship, while converting kilograms to pieces depends on product properties and may require actual measured weight rather than a nominal piece. Weight length presents a similar choice. A reported roll length might include material that can't enter usable stock, so the agreed measurement basis must state whether the quantity describes total measured length or usable output. Precision belongs in that agreement. Decide how many decimal places each quantity can carry, which calculations retain extra precision, and where the receiver requires a final rounded value before accepting a transaction. Rounding needs a business rule, not a programming default for individually counted parts rounding a fractional result upward can create a piece that nobody produced, while rounding downward can leave a real piece outside the recorded quantity. Consider a receipt of ninety five pieces in our packing example that doesn't form eight full packs of twelve, so the plant needs an approved rule for loose pieces, fractional packs, or holding the quantity until packing meets the requirement. The interface shouldn't choose among those options. Its job is to apply the product's approved quantity rules and explain when the reported amount can't be represented in the receiving unit without breaking those rules. Product specific limits can help catch mistakes. A conversion may require whole piece results, permit fractional weight or restrict variation around a measured pack quantity, depending on what the product and receiving process allow. Keep the original measurement alongside the converted quantity that lets someone check the ninety six accepted pieces against the eight posted packs, without trying to reconstruct the source amount from a rounded result recorded later. Now put that meaning into the message contract. The quantity type should say what the amount represents, such as accepted output, while the unit states whether the amount arrives as pieces, packs weight or length. The operation and lot establish where that quantity belongs. Without those references. A receiver might combine measurements from different stages or different lots and treat them as one amount. Even though the numbers describe separate work, the measurement basis explains how the amount was established for weight. That could mean measured net product weight rather than container inclusive weight. For pieces, it could distinguish an individual count from an estimate derived through weighing. You don't need an essay inside every message. Use agreed terms that identify the measurement basis. Preserve the original value and unit, and make any applied conversion rule traceable to its approved definition. That gives the receiver a quantity it can interpret. But even a correctly converted amount still needs a quantity type, because a count of produced pieces and a count of accepted pieces describe different conditions of the output units. Settle how much we're counting. The next question is which parts belong in that count? Because finishing work on a part doesn't by itself establish that the part belongs in accepted output. Let's put actual counts against our machining order at this point in the hypothetical process. All one hundred parts have passed through. Machining with ninety two accepted, five awaiting rework and three classified as scrap. Those counts describe the parts at one defined moment. Each part belongs to exactly one of those outcome groups, so adding the groups accounts for the full quantity without counting any individual part twice. The five awaiting rework aren't accepted output yet. They remain work in progress because they need further work before anyone can decide whether they belong with the accepted parts or must move into scrap. That distinction needs a movement rule. When someone assigns a part to rework, the current quantity, view should remove it from its previous category and place it into rework rather than simply adding another count. Otherwise, the totals can grow without any new parts. If five parts appear in both accepted output and awaiting rework, the combined quantity suggests more physical output than the order actually contains. Hiding the unresolved work. Keep the categories tied to a specific stage. A part can finish machining and still await another required operation, so accepted output from machining needs a different scope from accepted output for the finished product. That also applies when you. Total quantities across operations. Adding the output of machining to the output of a later operation counts successive work on the same parts, rather than measuring how many distinct parts the plant holds. Rework makes that counting problem more obvious. Suppose the five parts return to machining for a corrective pass, and the operator records five completed pieces against that additional work on the same production order. The machine has now processed one hundred and five pieces across those passes, but the order still contains only one hundred distinct parts. Because the corrective paths worked on existing parts, rather than creating five new ones. Both measures can support a useful question. Processing activity tells you how much work the resource performed, while a distinct part count, tells you how many individual parts remain within the order's physical flow. Name the measure before adding it up. If a report combines first pass output and rework output, call that total processed quantity and retain the distinction between original processing and repeat work against existing parts. Now suppose all five reworked parts become acceptable. The current position changes to ninety seven accepted and three scrapped, with nothing remaining in rework, while the recorded processing activity still includes the extra machining paths. The earlier rework record still describes work that really occurred. Moving parts out of the current rework category changes their present condition. It doesn't erase the time or processing activity required to bring them into accepted output. Scrap follows a different physical path. For this example, the three scrapped parts leave the route to what finished output, so they remain accounted for as losses rather than disappearing from the order's quantity picture. This is where a material balance helps you compare what entered a defined process with what left as output, what remains inside as work in progress, and what left through scrap or another recorded loss. Choose the process boundary before doing that comparison. For our part count example, the boundary starts with the blanks entering machining and ends with accepted parts, scrapped parts and any parts still awaiting further work. Weight needs a different balance from part count. Machining removes material, so a weight balance must account for removed material as well as finished output. Rather than treating every difference between input and output weight as missing stock, keep the basis consistent within each balance. Comparing input weight with output piece count won't establish whether material is accounted for unless an approved relationship connects those measurements for the particular work under review. An unexplained difference deserves investigation, not automatic adjustment. Someone needs to establish whether parts remain between steps. Account includes repeat processing or a loss hasn't been recorded before. Changing the figures to force agreement. These categories describe a position that can change. An initial completion report might capture the end of processing before the final disposition of every part becomes known, leaving later decisions to change the accepted quantity. Acceptance, therefore needs its own place in the production record. When that decision changes after the first completion report, the systems need to represent the new condition without pretending the earlier processing never happened to those parts. Who can authorize those parts to move forward? I'd name the designated quality authority explicitly, because an execution record from Ms. establishes what production recorded. While release requires a decision under the plant's approved quality process, that authority may sit inside m.e.s.. Some plants manage release there, while others use a separate quality application or an ERP quality function. With the approved release decision recorded in that designated system. The distinction concerns responsibility, not the number of applications. Even when one application supports both tasks, reporting completed work and approving its output should remain separate actions, each performed by someone with the appropriate authority. A past inspection result can support a release decision, but where the process requires further review, that result remains evidence for the authorized decision, rather than permission for the integration to release stock on its own, I'd make the release statement specific enough to act on. It should identify the affected lot or portion of a lot, the approved quantity, and any conditions that limit where or how those parts can proceed. Partial release needs a physical boundary as well. If quality release is one portion while holding the rest, warehouse staff need a reliable way to distinguish those portions through lot subdivision container identification or another approved method. A quantity alone won't tell a picker which container to move. The release record needs to match the physical identification used in the warehouse, so staff can select release parts without also moving parts that remain blocked. The held portion keeps its restriction. Releasing another portion of the same order mustn't clear that restriction indirectly through an order level status that makes every associated container appear available for picking or use. Consider a fresh complication in our hypothetical plant. After an initial release, quality discovers that a measuring instrument needs investigation, and the review includes parts previously assessed using that instrument during the affected period. An earlier approval doesn't prevent a later hold. The quality authority needs to identify the affected stock and record the new restriction, with a traceable link to the release that people previously relied on. Preserve both decisions and their reasons. Someone investigating the order should be able to establish why the stock initially qualified for release. What new evidence changed that decision, and who authorized the resulting hold? The required action depends on where the parts have reached stock. Still awaiting picking may need a warehouse block while stock already picked needs intervention in dispatch, so the earlier release doesn't carry it onto a vehicle. Parts that have left require a different approved response. The plant needs to establish their destination and follow its applicable containment process. Because changing an internal stock status can't physically recover material already outside its control, a reversal also needs a precise meaning. Reversing a release decision changes quality permission while reversing an inventory transaction changes a stock record, and the approved response may require one or both actions, depending on the circumstances. Shipment permission has to reflect the current restriction. Dispatch needs access to the effective release position when it authorizes movement, including a later hold that changes what an earlier approval allowed the team to do. Release is one condition for shipping, not the whole decision. The required inventory record must also exist. The stock must occupy a usable warehouse position, and the shipment must satisfy the plant's allocation and dispatch rules. Those checks need to agree on the same stock. A quality record for one lot doesn't authorize shipment from another lot merely because both contain the same product and the requested quantity happens to match. I would keep the reason for a shipping block explicit if release is complete, but warehouse receipt remains pending. Dispatch should see that specific condition rather than asking quality to repeat approval for work it has already authorized. That detail. Directs the next action. Warehouse staff can resolve a receipt issue while quality handles an unresolved release without either team changing its own records just to make an unrelated shipping restriction disappear. Timing can still complicate an otherwise correct comparison. Quality in the warehouse might report identical quantities, yet one statement describes the position before a hold and the other describes the position after that hold took effect. Ask when each statement applies. Without that boundary matching numbers can conceal a changed permission, and different numbers can describe a valid change rather than a disagreement about the parts themselves. Which clock does that timestamp belong to? A production record can carry several valid times, and choosing the wrong one can move work into a different shift without changing anything that physically happened. The occurrence time describes when the work actually happened in our hypothetical plant. That might be when the operator finished the machining run. Based on the best evidence available from the execution process, the source recording time answers a separate question. It tells you when Ms. saved the confirmation, which can happen later if the operator enters the result after completing other tasks or resolving an uncertain count. Integration receipt time marks arrival at the integration boundary. ERP posting time then identifies when the receiving application records the business transaction rather than when production performed. The work described by that transaction keep those times distinct when diagnosing a delay, the interval between occurrence and source recording points toward a different part of the process, from the interval between integration receipt and the eventual ERP posting. Suppose a shift ends at six in the morning. An operator finishes the work shortly before the handover, but records the confirmation shortly after six. Once the next shift has started and the immediate handover tasks have finished, the work belongs to the earlier shift under this plant's reporting rule, a report based on recording time assigns it to the later shift instead. Even though the operator entered an accurate account of when production actually finished. 00:35:26 Speaker: Neither timestamp needs to be discarded, preserve when the work occurred and when someone recorded it. So a supervisor can distinguish late entry from output genuinely produced by the incoming shift occurrence. Time can carry uncertainty, too. If an operator estimates the finish time from a handwritten note, don't present that estimate with the same implied precision as a trusted, automatically captured production timestamp. Clock quality affects how much confidence those comparisons deserve. A source clock running ahead can make a message appear to arrive before its reported occurrence, creating an impossible journey without any actual reversal of events. Check clock synchronization before drawing conclusions from small timing gaps. Record the timestamps, origin and known limits where practical, especially when several devices and applications contribute times that people later compare as though they share one clock for exchange. I'd normally use coordinated Universal Time or UTC that gives systems a common time reference while keeping the relevant local offset and plant timezone available for interpreting shifts and local business dates. The offset and time zone answer different needs. An offset tells you the local difference from UTC at that moment, while a named time zone lets software apply the relevant rules for local clock changes. Daylight saving changes make local timestamps awkward. When clocks move backward, the same local time occurs twice, so a timestamp without an offset can leave two different moments indistinguishable in the production record. The spring change creates the opposite problem. Some local times never occur, so any process that builds shift boundaries from local clock values needs to handle that gap, rather than assuming every day contains identical hours. Daily totals. Need a business calendar as well? A plant running overnight might define its production day from one morning boundary to the next, meaning midnight doesn't necessarily start a new production reporting day. Accounting can draw a different boundary. The ERP transaction may carry a posting date governed by finance rules, so output assigned to yesterday's production day can legitimately enter today's accounting totals. After a late confirmation reporting refresh adds another time to the conversation. A report read after the posting might still contain an earlier extract, leaving its total behind the current ERP position until the next successful data refresh completes, make the bases clear beside the total. People need to know which production day or accounting date it covers and how current the underlying data is before they compare it with another figure for an investigation. Bring those boundaries together. Explicitly compare occurrence based production totals with the relevant postings, and refresh cut off, allowing for the specific late entries rather than assuming every difference represents missing output. A current status alone hides most of that sequence. Two records can now show the same position while concealing when work finished, when someone recorded it, and when the receiving system finally reflected that work. That limits what a snapshot can explain. To understand why an order appeared unfinished during a handover, we need its position at that moment, and the changes that brought it to its current position. To represent that position, give the order several related state dimensions. Each dimension tracks a particular responsibility, so a change in execution can affect what happens next without silently changing quality permission, inventory, receipt or commercial closure. Execution tracks the progress of the required production work. Quality tracks its disposition. Inventory receipt tracks where the output has entered the required stock records and commercial closure tracks whether the order permits further business activity under your process. These dimensions need rules connecting them. For example, your process might permit an inventory receipt while quality remains pending, but prevent commercial closure while required. Production confirmation still await approval from the responsible team. Define those dependencies explicitly rather than relying on a screen label for every permitted change. Establish the starting condition, the destination condition, the prerequisites, and the role or application authorized to approve that particular change. Take an operation moving from active to finished. The rule might require recorded output and resolution of outstanding work entries, with production responsible for confirming those conditions before the operation leaves its active state. Quality follows its own permitted changes. A transition from pending assessment to release needs the approval required by your process. Even if execution already qualifies as finished and nothing remains for the operator to perform, cross dimension rules prevent impossible combinations from passing without explanation. If commercial closure requires all expected receipts to be resolved, an order with an outstanding receipt should remain ineligible unless an authorized exception explicitly settles that requirement. An order level status then becomes a derived answer. It should reflect the relevant operations and their quantities under a defined rule, rather than copying whichever operation happened to report its latest change. Most recently, that rule depends on what the overall status supports. A production supervisor may need to know whether required execution remains, while a commercial owner needs to know whether the order still permits further approved business activity, a partial result needs to stay partial. If one operation finishes only part of its assigned quantity, its completion record shouldn't make the whole operation appear finished unless the remaining quantity receives an approved disposition. The aggregate also needs to handle work that won't continue when an authorized cancellation removes remaining work, the order can reach an appropriate terminal state without describing that cancelled quantity as output the plant successfully produced, I'd keep the reason visible in the derived result. Work finished and work cancelled can both leave nothing further to execute, but those outcomes support different conclusions about whether the original requirement has actually been fulfilled. Reopening needs just as much design as forward progress, a system that permits completion but provides no approved route back to actionable work forces people to invent workarounds when the physical process needs another step. Consider rework assigned after execution has finished. The rules should establish whether the original operation reopens or a rework operation becomes active and how that decision changes the order's remaining execution requirement. A correction can affect a different dimension if an approved quantity correction reveals that the recorded receipt no longer covers the required output, the receipt position and any dependent aggregate status need reassessment. Under the applicable rules, don't reopen everything automatically. A correction to one operation may leave the quality disposition of unaffected parts unchanged, so the transition should identify exactly which state and quantity require further action rather than resetting the entire order. A late quality hold creates another distinct path. It can change permission for affected output without changing the fact that production finished, while forcing any closure rule dependent on unrestricted output to be evaluated again. Commercial closure may restrict which paths remain available if the order no longer permits ordinary changes, the state model needs an authorized exception route or linked follow on work, depending on what the application and business process allow. People need to understand those paths before using them. Production, quality inventory and the commercial owner should be able to explain which change they control and which dependent decisions their change requires someone else to revisit. That gives each status change a specific meaning. We can now ask what occurrence justifies moving from one state to another, and what evidence the responsible system needs before it accepts that movement as valid. The state describes the position. An event describes a change. Following those changes lets us connect completed work to the next business decision without assuming that recording progress also proves every downstream action has already succeeded, an event states something that happened. A command asks another system to perform an action, and a result tells us whether that system accepted the requested business action or refused it for a stated reason. The wording helps keep those responsibilities separate. Operation completed describes an execution fact. Post-Production receipt requests an inventory action and production receipt posted. Confirms that the receiving application actually recorded that action. Those statements belong to different owners. Mez can report the execution fact it controls, but ERP decides whether the requested receipt satisfies its own business rules and can enter its inventory records. Sending a command doesn't transfer that decision to the sender. The integration can prepare the request using approved rules. While the receiving application retains responsibility for accepting or rejecting the transaction within its business boundary that affects how we describe success. Successful publication means the source fact entered the exchange. While successful receipt posting requires evidence from ERP that the requested business transaction actually reached its accepted outcome. Consider another complication for our machining order. Mes records a valid completion and the integration requests the corresponding production receipt. But ERP rejects that request because the destination inventory location no longer permits the requested transaction, the completion event remains valid. Production performed the work it reported, and the rejected receipt doesn't undo that work or justify changing the execution record. To suggest that machining still needs to happen, the receipt request now has a rejected result. That result belongs beside the request with the receiving application's reason, so the responsible person can address the destination restriction without asking production to report the work. Again, keep those outcomes separate in the operational language. Execution recorded receipt rejected tells the planner where the problem sits while integration failed leaves everyone guessing whether the work, the message, or the requested transaction needs attention, a rejection should identify the requested action precisely if the same operation supports several legitimate receipt requests, an error against one request mustn't imply that every receipt associated with that operation has also been refused. This is where correlation earns its place. Correlation means retaining the links between the original production fact, the request created from it, and the response that records the receiver's decision about that particular request. An order number alone can't provide that link. The same order can produce several completion reports and several receipt requests, so finding the right order doesn't establish which request a particular response actually answers. Give each request a stable identity. The response should carry that identity or provide a dependable way to recover it, so the integration can attach the result to the exact action it asked ERP to perform. Keep the originating fact linked as well. Someone should be able to start from the completion record and find the receipt request, then follow that request to its accepted or rejected result without matching records by guesswork. Acceptance needs the business transaction reference when ERP posts the receipt, retain the identifier of that posting alongside the request identity. Because the posting reference connects the integrations result to the actual inventory transaction, those references answer different questions. The request identity locates what the integration asked for, while the ERP transaction reference locates what the receiving system recorded and the relationship between them supplies evidence of the outcome. Don't manufacture a posting reference for a rejection. A rejected request may have a response or error reference, but that reference identifies the refusal and mustn't be presented as proof that inventory changed. Some interfaces return the final result later in that arrangement. Keep the request in a pending position until the later response identifies the business outcome, even if the initial exchange already returned without a technical error, that waiting position needs a clear name. A planner reading a waiting ERP result can distinguish an unanswered request from a rejected receipt. Rather than treating silence as either success or proof that nothing will happen, I want that chain to survive. An ordinary support question when someone asks why stock didn't increase, the records should connect the production fact to the requested receipt, and then to either posting evidence or a specific refusal. The relationship depends on information the messages actually carry. To preserve it across applications, we need an agreed minimum that identifies the fact and its meaning clearly enough for each participant to maintain those links. An event contract makes that minimum explicit. It's the agreement between the application describing a production fact and every application reading it, covering both the message structure and the meaning of its contents. I'd review it with production, not just the integration developers. If the people responsible for the work can't explain what an event asserts, a technically valid message still leaves too much room for interpretation at the receiving end, start with an identifier for the event itself. Define whether that identifier is unique across the exchange or unique within a named source, so consumers know exactly which combination identifies one recorded occurrence. The source needs a stable agreed name, identify the originating application and relevant instance, rather than using the name of a connector. That could change while the underlying production records remain in the same system, business Scope tells consumers where the contract applies an event type approved for one plants discrete assembly process shouldn't quietly acquire a different meaning when another plant starts publishing events under the same name, give the contract its own version as well. That version identifies the agreed message definition, allowing a consumer to determine which interpretation applies without inferring it from whichever fields happen to appear in the payload field. Definitions deserve more attention than their names for the order and operation references. The contract should state which identifier scheme each field uses and which surrounding business scope makes those references unambiguous to consumers. Quantity category should come from a controlled vocabulary. Each permitted category needs a written definition, so publishers can't introduce a new category through free text and expect every receiving application to interpret it consistently without review. The unit and event time need equally explicit rules. Specify the permitted unit codes and timestamp format, including what the event time represents, rather than leaving each developer to choose a reasonable interpretation for their own application. Also define what an absent value means. A missing quantity might be permitted for some event types, but it mustn't silently become zero simply because a receiving application initializes an empty numeric field. That way, trace context adds relationships without changing the production fact. A causation reference identifies the earlier event or request that directly prompted this event where that relationship exists and the source can actually provide evidence for it. Correlation groups records within a wider business interaction. Several events may share that grouping while retaining different immediate causes. So the contract should explain who assigns the correlation reference and how participating applications carry it forward. A source revision can identify a particular recorded change. Where supported, a sequence number can describe position within a named source stream, but the contract must define that stream before consumers draw conclusions from the number sequence doesn't automatically mean a plant wide order of events. If numbering applies only within one execution record, document that boundary and avoid implying that values from unrelated executions can be compared as one continuous sequence. Keep unsupported context honest if the source can't supply a revision or causation reference, define whether the field remains absent rather than inventing a value that appears useful but has no dependable relationship to the source, the contract will change as the process develops. Adding an optional field can preserve compatibility when existing consumers tolerate unfamiliar fields, and the established meaning of every existing field remains unchanged. Under the new version, even an extra category can break a consumer. An application that recognizes a fixed set of quantity categories may reject an unfamiliar one, so adding a permitted value deserves compatibility testing rather than an automatic minor version label. Changing meaning calls for an explicit breaking change. If a quantity field previously described an increment and will now describe a cumulative total, keep those definitions distinct and provide a planned migration for affected consumers. Consumer tests make that agreement executable. Use representative messages to check how each consumer handles permitted values, missing optional information and supported versions. Then run those tests when either the publisher or consumer changes. A signed document can't prove that applications follow its rules. Those tests provide evidence that the implemented exchange preserves the agreement while keeping the production meaning open to review by the people who own it. Now that agreement has to survive actual plant conditions. A clear contract tells us what a message means. The next design choice concerns how that message reaches the applications that need it when operating conditions change. Choose the exchange pattern around the person or process, waiting for it. Before selecting a connection method, I'd ask whether the next step needs an answer immediately or whether the work can continue while another application catches up. Request response fits when the caller can safely wait, one application sends a request. The receiving application checks it, and the caller uses the returned answer before allowing the person or process to continue with the next step. Consider an order release check before work starts. A supervisor requests validation of the proposed order details, and the application waits for the receiving system to confirm whether those details satisfy the checks assigned to that interface. Waiting needs an operational limit. Define how long the supervisor can reasonably wait and what the application should display. If no answer arrives, rather than leaving a person wondering whether to press the button again, that dependency belongs in the process design. If the remote application must answer before every routine confirmation, its availability becomes part of the operator's workflow, whether or not the integration team intended that consequence. Queued work fits a different need when production can record its confirmation without an immediate receiving system decision, the exchange can retain that confirmation as pending work until the receiver becomes available to process it. Imagine ERP undergoing planned maintenance while mess remains available. Operators continue recording permitted production activity, and the integration places the resulting confirmations in a durable queue. Instead of requiring each operator to wait for ERP, the queue changes. When processing must happen. Production can hand over the confirmation before the receiver handles it, provided the business accepts that interval and the queue retains the work under the agreed operating conditions. Buffering doesn't remove the receiving system's workload. It lets that workload wait. So the design still needs enough processing capacity to handle the accumulated confirmations within the time the business can tolerate after maintenance ends. Published events address another requirement. Several applications need the same fact. Rather than making the production application, contact each destination separately, it publishes the fact through a distribution mechanism that lets interested consumers receive their own copy. Different consumers can then use that fact independently. A production history application might retain it for analysis, while a notification application uses it to inform a supervisor that a particular execution milestone has been reached. That independence helps when consumers have different operating needs. A slow history import shouldn't force the notification consumer to wait, and adding another authorized consumer shouldn't require production to manage another direct conversation during execution. Publication and queued work can overlap in implementation. The distinction concerns intent. A work queue assigns pending work for processing, while publication distributes a fact to consumers that each have their own reason to receive it. Batch exchange remains sensible where the business accepts scheduled updates. A plant might send an agreed collection of confirmations at fixed intervals, because the receiving process operates on that cadence and doesn't need every record immediately, a batch makes its waiting period part of the design. If the next exchange runs after a planning deadline, increasing file transfer speed won't solve the mismatch between the exchange schedule and the time the planner needs the information. 00:53:28 Speaker: Files still need explicit business, meaning an export containing an order's latest values needs to state whether those values replace a previous position or describe additional activity. Because the receiving application must apply the collection consistently. Change data capture starts from recorded database changes, often called CDC. It detects, inserts, updates, or deletions through supported mechanisms and can move those changes onward without requiring a separate application level publication for every relevant change. That can help when application integration options are limited, but one saved business action might change several database rows, and those row changes don't automatically tell a consumer which business milestone the application intended to record. Treat captured changes as evidence requiring interpretation. The design needs a supported understanding of the source data and its transaction boundaries before turning those changes into production facts that other applications can safely use. One order flow can use several exchange patterns. The choice follows who needs an answer, how long they can wait, and whether the exchange assigns work or distributes information, rather than a preference for one integration style everywhere. Now we can place Azure services against specific jobs. Each service should support the exchange behavior we've chosen, with its role clear enough that its product name doesn't substitute for an explanation of how the production process works. For a production receipt waiting to reach ERP, I'd consider Azure Service Bus. It supports durable business messaging with queues for shared work and topics with subscriptions. When separate, consumers each need their own delivery of a published message. Consumption stays under the receiving application's control. A consumer takes a message for processing and settles it through the messaging interface, allowing the application to coordinate its work with the broker rather than treating delivery as an untracked hand-off. A queue can support several competing processing instances. Each instance handles a share of the pending work, while the design controls concurrency, so the integration doesn't send more simultaneous requests than the receiving ERP interface can reasonably handle. Subscriptions give separate consumers their own pending messages. Subscription filters can select relevant messages using message properties, but those filters apply technical selection rules. They don't establish whether a production quantity qualifies for an inventory transaction. Event grid addresses notification and routing. If an application announces that a production document is available, Event Grid can route that notification to interested handlers without making the publishing application manage each destination itself. I distinguish notification from an assigned business posting event grid supports several delivery arrangements, so check the chosen configuration against the required behavior rather than assuming every event destination works like a service bus queue, event hubs addresses a different workload again. A continuous stream of machine measurements can flow through event hubs, where partition streams let consuming applications read the incoming records at the pace their processing supports. Consumers track their position in those streams. Separate consumer groups allow applications to read independently, and retained records can support rereading within the configured retention period, rather than disappearing simply because one reader processed them. That distinction affects the production integration design, a stream of equipment measurements, and a request to record an inventory transaction. impose different processing needs, even when both eventually contribute information about the same order, then we need somewhere to run. Application specific logic. Azure functions can host code that checks an incoming contract, resolves an approved reference, or transforms a recognized source representation into the structure expected by the receiving interface. Keep those transformations narrow enough to explain. A function. Can apply a defined mapping, but the business owner still needs to approve that mapping, especially where a technical conversion could change the meaning of reported production. Logic apps fits workflow and connector orchestration. It can coordinate calls between supported applications and services, with the available connectors and deployment options shaping how the workflow reaches the systems involved in your plant. Connector names aren't proof of process coverage. Before choosing a connector, check whether it supports the exact application version, operation, and authentication method you need, including whether it's returned response represents the business outcome your workflow expects. Functions, and logic apps can work together. A workflow might coordinate the application calls while invoking a function for a specific transformation that needs custom code without forcing every step into either code or a workflow definition. Neither choice supplies the manufacturing rules automatically. You still define which inputs qualify for which action, and the implementation needs to preserve that decision clearly enough for engineers and business owners to review it together. Access to the receiving interface needs its own control. Azure API management governs access to application programming interfaces, or APIs through policies that can check credentials, restrict traffic, and control how approved callers reach a back end service. A gateway policy doesn't own the production order. API management can reject an unauthorized call before it reaches ERP. While ERP remains responsible for deciding whether an authorized request satisfies the rules for the requested business transaction, I'd also check what sits behind an accepted gateway response. Passing through API management establishes that the call passed the configured gateway controls, not that the back end application has completed the requested inventory action successfully. These services don't all belong in every integration. Select the ones whose behavior matches the work. Then state their responsibilities plainly enough that removing a product name doesn't leave a gap in the explanation. There's still a boundary between service behavior and business guarantees. A broker can retain a message, and a workflow can complete its steps while proving that the intended production posting occurred requires evidence across the entire exchange. A timeout can leave the sender unsure about a completed posting. Suppose ERP accepts a production receipt and commits the inventory transaction, but the connection drops before the integration receives the acknowledgment confirming that transaction. The integration now faces an uncertain outcome. It can't conclude that the receipt failed merely because the response never arrived, and sending the request again could create another posting for output already recorded. This is an expected failure to design for even when both applications function correctly, a broken connection can separate the receiving application's completed transaction from the evidence. The sender needs to recognize that completion. I'd give that uncertainty its own processing state. Mark the request as awaiting outcome verification rather than failed. So recovery starts by establishing whether ERP committed the transaction. Before anyone authorizes another posting attempt. The required behavior has a name business Idempotency. It means that repeating the same intended receipt request produces one business posting, even when the integration delivers that request more than once during recovery. The identity must survive every attempt. If retry logic creates a fresh business request identifier each time, the receiver may reasonably interpret each attempt as separate work and accept several receipts for the same production fact. Keep delivery attempts separate from the intended action. You can record when each attempt started and how it ended, while retaining one stable request identity that ERP can recognize across all those attempts. ERP support determines how strongly you can enforce this. Where its supported interface accepts an external request reference and prevents duplicate use. Send the same reference again and recover the existing result through the documented behavior. Check exactly what that protection covers. A field that merely stores an external reference won't prevent duplicate posting unless the receiving application checks uniqueness within the correct scope as part of accepting the transaction. Without that support, the recovery design needs more care. After an uncertain result, use an authoritative lookup where available and hold the request for investigation when you can't establish whether retrying would create another receipt. A durable processing record supports that investigation. Retain the source event identity, the intended receipt request, its current processing state, and the ERP transaction reference. Once the receiving application confirms the posting, that record must survive a worker restart, keeping the relationship only in running application memory leaves recovery without dependable evidence when the process stops after ERP commits, but before the worker finishes recording the result, there's still a gap to account for. Recording success in an integration database and committing a receipt. In ERP are separate transactions, so a local success marker alone can't guarantee that both records change together. Recovery needs to close that gap with evidence. Use the receiving application's duplicate protection or supported transaction lookup to recover the outcome. Then update the processing record with the confirmed reference instead of assuming another attempt is harmless, broker A duplicate detection can reduce repeated message delivery. Its configured detection scope and time window don't replace a lasting business check, especially when recovery occurs after that window or through a separate submission path. Retries also need to distinguish failure types. A temporary connection interruption or a documented busy response may justify another attempt provided the previous outcome is understood or the receiving interface safely recognizes. Repeated requests give those attempts limits and increasing delays. Follow any receiving system guidance about when to retry, because repeatedly calling an unavailable endpoint consumes resources without improving the chance that the next attempt can actually complete. Business rule refusals need a different response if ERP rejects the receipt because a required business condition isn't satisfied, stop automatic attempts and retain the request with its reason for the responsible person to resolve. Repeating unchanged input won't repair that condition. Once an authorized correction addresses the refusal resume through a controlled path that preserves the original request history and establishes whether the intended action still remains appropriate. Don't disguise a duplicate as a correction. If the requested quantity or destination changes, record that change explicitly under the approved process, rather than quietly replacing the contents behind an identity previously used for another request. With those controls, repetition becomes manageable rather than dangerous. You can recover from a lost acknowledgement without creating another receipt, but duplicate protection alone doesn't establish which distinct production event reaches the receiver first. Separate requests can still arrive in an unexpected sequence. A design that recognizes the same request twice must also decide how to handle different requests whose arrival order doesn't match the order in which production recorded them. Picture a completion reaching the receiver before its start. Confirmation production recorded both correctly, but the start message paused in one processing path while the completion message traveled through another path and reached the receiving application first, a simple status updater can now move the operation backward. It processes completion, then applies the delayed start as the latest instruction, leaving finished work marked active because it mistakes a rival order for production sequence. The receiver needs rules for that situation. Those rules should distinguish the order in which messages arrive from the order of changes within the particular execution. Using evidence, the source actually provides and supports. First, decide which events require ordering relative to each other. Changes within one execution may need a dependable sequence, while a confirmation from an unrelated order usually has no reason to wait behind that. Executions messages, a plant wide sequence creates an unnecessary dependency. One unresolved event could delay unrelated production updates across the site, even though those updates concern different resources and work with no shared prerequisite for processing. Choose a narrower boundary that matches the business dependency that might be an execution stream or an order when changes across its operations genuinely require ordered handling, with the choice documented for producers and consumers, Azure Service bus sessions can help within that boundary. Messages sharing a session identifier can be handled in an ordered sequence with an exclusive session lock, preventing competing receivers from processing that session. At the same time, the producer still has to preserve the intended order. If it submits, completion before start, a session won't reconstruct the missing production sequence. It preserves the broker's ordered delivery rather than discovering what production intended. Parallel sending therefore, needs deliberate control. Where several producers contribute changes to one execution, establish how they coordinate sequencing before relying on a shared session identifier to deliver a coherent history to the consumer. The consumer needs discipline to if it receives messages in order but launches their business updates concurrently, a slower start update can finish after completion and overwrite the result that should remain current. Process dependent changes serially within the chosen boundary. Keep unrelated sessions free to progress, while ensuring that settling or advancing a message doesn't allow the next dependent update to overtake unfinished business processing for that execution. Now define what counts as a stale update. A lower source revision may describe an earlier state already superseded, provided the source guarantees that revisions increase within this execution and the consumer retains the last applied revision. Stale doesn't mean the historical fact has no use. The delayed start can fill a gap in the execution history without changing the current position from finished back to active, because those are separate effects of processing. A missing event needs a different rule. If completion references a predecessor the receiver hasn't seen. Decide whether that predecessor supplies a required business condition or merely adds detail that can arrive later without blocking progress where the missing fact is required. Hold the dependent change, record exactly what's missing, and set a bounded wait for investigation. Rather than allowing one execution to remain stalled indefinitely without anyone responsible for resolving it, other contracts may allow completion to establish the current execution state. In that case, retain the history gap explicitly, and accept the completed position under that approved rule without inventing a start time or claiming the missing confirmation arrived. Don't infer gaps from numbers unless the contract permits it. A sequence that can legitimately skip values won't prove that an event disappeared, so missing event detection must follow the sources documented numbering behavior rather than an assumed counting pattern. Late corrections require a separate route through these rules. An authorized correction may arrive later while referring to earlier work, so rejecting everything with an older occurrence time would discard a legitimate change to the recorded position. Recognize the correction by its defined business, meaning. Apply the permitted change to the affected execution while retaining its relationship to the earlier event, instead of treating it as either ordinary forward progress or a stale status update. These rules become especially relevant when connectivity disappears. Production can continue within its approved operating limits while the cloud connection fails, leaving the integration to encounter a delayed sequence once communication becomes available again. An outage needs a plant operating procedure, not just a connection alarm. Specify which work can continue using locally available instructions. Which approvals remain accessible and which activities must stop when required. Information can no longer be checked. Keep execution authority within the approved plant boundary. Machine control and safety functions must retain their designed behavior independently of an integration connection with any loss of connectivity handled according to the equipment's approved operating requirements. That doesn't mean every job can continue indefinitely. A process requiring a fresh external authorisation may need to pause at that boundary, even while other operations continue, using instructions and permissions already valid for their work. Give supervisors those limits before the outage happens. They need a usable procedure that distinguishes permitted continued execution from activities requiring unavailable checks. Rather than asking them to interpret a technical connection error during a busy shift, local buffering also needs a defined point of acceptance. When the application tells an operator that a confirmation is saved. Establish whether it has reached durable local storage that survives a restart or loss of power. A cloud queue can't capture what never leaves the plant during a broken connection. The local application or approved edge component needs somewhere dependable to retain confirmations until it can hand them onward through the exchange. Size that storage against the outage the plant intends to tolerate. Allow for the actual confirmation rate and record size, including supporting data while reserving enough space for the local application to keep functioning under sustained load. Storage capacity and retention need separate decisions. A buffer may have room available, while an age based clean up rule deletes unsent confirmations, so pending work needs explicit protection from routine housekeeping and automatic file removal. Someone must own the buffers health. Assign responsibility for capacity warnings, damaged records, and recovery after a local failure, with an escalation path that remains usable when the usual cloud based support tools are unreachable. Define what happens as that buffer fills Backpressure means slowing or refusing additional intake before storage is exhausted, with the application clearly telling production whether another confirmation can still be retained safely for later delivery. Silent data loss isn't an acceptable overflow policy if the remaining capacity can't support continued recording, supervisors need enough warning to invoke the approved operating limit before operators perform work they can no longer document reliably. Recovery starts with a deliberate release of pending work. After connectivity returns, check that the receiving path is ready and feed confirmations at an agreed rate, rather than emptying every local buffer into ERP at once. Several lines may reconnect together after a side outage. Their combined backlog can compete with current business activity, so recovery needs a shared traffic limit rather than separate workers, each assuming the receiver has spare capacity. Watch the receiver while that backlog drains. Adjust the replay rate when response times worsen or capacity restrictions appear, and give support staff a controlled way to pause recovery without deleting the work awaiting delivery. Operators need to see their own pending confirmations, show whether the local record is safely retained and awaiting transfer, and make that distinction clear enough that a delayed stock update doesn't prompt another entry of the same work. The handover should include unresolved local work. An incoming supervisor needs to know what remains pending, how much recording capacity remains and which operating restrictions apply before deciding whether the next scheduled activity can proceed safely. Manual workarounds need explicit limits, too. If an approved fallback permits, paper records or direct entry elsewhere, define who authorizes it and how staff marked that works, so the pending automatic transfer doesn't create a second submission, avoid leaving that choice to individual operators. A handwritten note may support continued work under one approved procedure, while another process requires the activity to stop. Because the missing system supplies a permission the plant can't replace locally. Reconnection doesn't mean the business has stood still. A confirmation retained during the outage can reach ERP after someone changes the order or restricts its destination, leaving a valid account of production facing conditions that no longer match those present when the work occurred. Sometimes, the delayed confirmation isn't the record that needs changing while the connection recovers. A recount might reveal that the accepted quantity includes parts counted twice, requiring an approved correction to a fact already recorded. Take accounting error discovered after a receipt has posted, a supervisor finds that one container appeared twice on the inspection tally, so the accepted quantity exceeds the number of accepted parts physically present in that shipment group. The extra quantity never existed. This isn't scrap or newly rejected output. It's an error in the recorded count, and describing it as a physical loss would create a false account of production. 01:11:48 Speaker: Record the correction as a distinct authorized action. It should reference the original acceptance record and identify the affected quantity so another person can establish precisely which earlier statement the correction changes without searching unrelated entries. State whether the correction replaces a total or adjusts it. A replacement quantity and a negative adjustment can express the same intended result, but applying one as the other would introduce another error into the order. The reason needs enough detail to explain the change. Duplicate container count confirmed during recount gives the reviewer evidence to examine, while quantity fixed leaves the next shift unable to distinguish accounting mistake from an unexplained adjustment. Identify the person responsible for that decision. If an integration account submits the correction, retain the human actor and approval where required, rather than leaving the service identity as the only explanation of who authorized it. Keep the effective time separate from the corrections. Recording time. The effective time identifies when the corrected statement applies, while the recording time preserves. When the organization discovered and entered the change against its existing records. That distinction prevents a misleading historical account. Someone reviewing an earlier shipping decision needs to know what the system showed. Then, even if later evidence establishes that the quantity shown at that moment was wrong. ERP determines how it's posted. Receipt can change. Check the supported reversal and reposting procedure, including whether the application permits a partial reversal or requires the original transaction to be reversed before a corrected receipt can be posted. Use that approved transaction path rather than editing stored balances directly. A database change might produce the desired stock number while bypassing the accounting entries, and document relationships that explain how the inventory position reached its current amount. Existing stock activity can restrict the available correction if some receive parts have already been consumed or dispatched. Reversing the whole receipt may be blocked or require additional approved actions before ERP can accept the change. Finance also controls the accounting boundary. A correction discovered after a period closes may need an authorized posting in an open period according to the application's rules and the organization's approved accounting policy for that situation. Preserve the original production timing. Even then, an open period adjustment can carry today's accounting date while retaining references to the earlier work and counting error without pretending that production physically occurred on the adjustment date. The correction now spans separate applications and separate transactions. Updating the accepted count, reversing a receipt, and posting a corrected receipt won't ordinarily behave like one database transaction that either commits everywhere or disappears everywhere. Plan the recovery as individually tracked actions. Record which correction step has completed and retain its transaction reference, so a stopped workflow can resume from evidence rather than repeating every action from the beginning. Suppose the reversal succeeds, but the corrected receipt is refused. Inventory now reflects the reversal, and the correction case must remain visibly unfinished, while the responsible person resolves the refusal through the approved business process. A compensating action addresses a completed action through another authorized transaction. It doesn't erase history, and it might require a different operation from simply undoing the last step, especially when other business activity has occurred. Meanwhile, don't let recovery invent that decision. Define which compensation is permitted for each interruption, who approves it, and when. The safest response is to hold the case for review rather than automatically attempting another stock change, close the case only after its required outcomes are confirmed. The corrected acceptance record and resulting ERP documents should remain linked with any approved exception recorded, so support can explain both the intended correction and its actual outcome. Even that discipline won't catch every disagreement. Some differences never generate a failed request or a correction case, so we still need a way to discover where the recorded business positions have drifted apart. Reconciliation starts with a business question, not a delivery total. For each quantity that should have reached a particular business milestone, can we find the corresponding recorded outcome or a documented reason why that outcome remains pending? Matching message counts won't answer that question. The sender and receiver might each report the same number of processed messages, while one receipt contains the wrong quantity and another expected receipt never reaches inventory. Choose a comparison scope that people can investigate. I'd compare a specific order operation and lot using the relevant quantity category and an agreed time boundary. Rather than subtracting two plant wide totals and investigating whatever remains. The categories don't always match directly between applications. Compare accepted output with the receipts expected under your approved process. Accounting for restrictions and exclusions, rather than assuming all recorded production should already appear as unrestricted inventory, the cutoff needs to apply consistently. If the comparison covers facts recorded before a shift boundary, later receipts should be recognized as subsequent resolutions not quietly included on one side, while the other side remains frozen. Allow for the delay. The process actually permits a confirmation still moving through its normal processing window belongs in a pending category, while an unexplained difference beyond that window needs someone to establish why the expected outcome hasn't appeared. Age alone doesn't explain the difference. A recent request with an explicit refusal already needs attention, while an older request retained during approved maintenance may have a known recovery path and an accountable owner. Give each difference a reason where evidence supports one separate known pending work from unexplained discrepancies, and avoid clearing an exception merely because somebody expects the next scheduled exchange to resolve it without checking a missing receipt is one reconciliation case. The comparison finds output eligible for receipt under the agreed rules, but neither a corresponding ERP posting nor a documented pending request accounts for that output within the selected boundary. That finding doesn't authorize another posting immediately. The owner first needs to establish whether the receipt exists under a different reference, remains unapplied or never entered the receiving process using the records associated with that particular quantity. Duplicate quantity produces the opposite discrepancy. Posted inventory exceeds the quantity supported by the relevant production and release records, so the investigation must identify which transactions account for the excess before anyone chooses a corrective action. An unresolved hold requires a different comparison. Where the approved process requires affected stock to remain restricted. Reconciliation should check that the effective quality restriction has the corresponding inventory or warehouse consequence for that same stock. A hold can exist without taking effect downstream. The quality record may show the restriction while the receiving stock record still permits use, leaving a discrepancy that deserves attention even when every quantity total agrees. Stale closure belongs in the comparison as well. An order may remain closed despite a recorded condition that should trigger review or remain open after every prerequisite for its designated closing milestone has been satisfied. Neither case calls for blindly copying another status, check the closure rule against the supporting facts, then establish whether the status needs an authorized change, or whether an unresolved condition legitimately prevents that change from taking effect. Every unresolved case needs a named owner. Assign it according to the business issue with supporting evidence and a clear next action. Rather than leaving a shared exception list that everybody can read but nobody must resolve, keep the evidence with the case. Include the comparison scope, cut off relevant source records, expected outcome and observed difference so the owner can investigate the specific finding without reconstructing why it appeared on the list. Resolution needs an approved correction, not just a disappearing alert. The owner should record what action the responsible team authorized, why that action addresses the discrepancy, and which resulting records will prove that the intended business position now exists. Verify the result against a fresh compatible comparison. A completed task or accepted request isn't enough to close the case. If the affected quantity, restriction, or closure condition still differs from the position the correction intended, some differences will remain valid and documented. Record the accepted reason and its applicable boundary, so the next comparison distinguishes that approved exception from a new discrepancy. Rather than repeatedly asking someone to resolve the same condition, support needs a traceable root to investigate those findings. Once a case identifies the disputed business outcome, the team needs to follow its evidence through the integration and locate where the recorded journey stopped matching the expected result. Give support a starting point they can actually search. A planner should be able to supply a production reference and find the relevant exchange without knowing which Azure Resource processed it or which developer built the connection. The first useful answer describes where progress stopped for a disputed receipt. Support needs to distinguish a source event awaiting transformation from a transformed request, waiting in a queue or a request awaiting the receiver's business result. Record evidence at those boundaries as work progresses. Capture the source event, reference the transformation outcome, the messaging reference, and the receiver interaction, preserving their relationships. So an investigation doesn't depend on matching similar timestamps across separate logs. Transformation evidence should explain the applied rule retain its version and the validation result. So support can establish which mapping process this particular event without assuming that today's deployed logic matches the logic used when the request first arrived. Queue evidence needs to describe the current waiting condition. A message ready for consumption needs different attention from one held after repeated processing failures. Even though neither has yet produced the business outcome the planner expects keep the receiver's outcome attached to that journey. A completed call should lead support to the recorded business result or an explicitly unresolved result, rather than ending the investigation at a technical entry that simply reports success. Azure Monitor can bring the operational signals together. Application insights can capture application traces, dependencies, and failures from instrumented components, giving engineers evidence about execution while they investigate the production request associated with those components. That evidence needs deliberate correlation across the exchange. Carry the agreed identifiers through message properties and application telemetry, because a trace won't reliably continue across an asynchronous handoff unless the sending and receiving components preserve the relevant context. Technical trace identifiers and business references have different jobs. Link them explicitly so support can move from a planner's order reference to the relevant processing attempts, including attempts that ran separately and therefore don't share one continuous application. Trace telemetry coverage also needs checking. Sampling, retention settings or missing instrumentation can leave gaps, so an absent trace entry shouldn't become proof that a production request never existed, or that it's business action never occurred. Operations needs measures beyond application response time. Track the oldest pending receipt using a defined starting milestone, and let staff open that measure to identify the actual request, its current waiting condition, and the last confirmed progress. The quantity gap needs its own business basis. Report unresolved differences in the agreed unit and category, keeping products with incompatible units separate, rather than presenting a plant total that adds pieces to kilograms and calls it production exposure. Reconciliation backlog adds another operational signal. Show how many cases await action and how long they have remained unresolved, with enough detail to distinguish new findings from cases that repeatedly return without reaching a verified resolution. Together, those measures help support choose where to intervene. A technically quiet integration may still contain an old receipt blocking dispatch, while a burst of harmless connection retries may have no unresolved effect on the production position. Error wording should direct investigation without inviting guesswork instead of displaying an internal exception alone, present the refused action, the receiving application's usable reason, and the supporting reference, while retaining technical detail for staff authorized to inspect it. Replay controls need to show exactly what will happen before an operator resubmits anything. Display the selected request, its known outcome and whether the action retrieves unchanged content or invokes a separately approved recovery procedure for that specific case. Make uncertainty visible before the replay can proceed. If the previous outcome remains unknown, the control should direct support to what verification, rather than presenting another submission as a routine remedy for an unanswered request or missing acknowledgment. Protect production data while making failures understandable. Support views can expose the references and reasons needed for diagnosis without routinely copying complete payloads, product specifications, or sensitive operational details into broadly accessible application logs and exported incident reports. Diagnostic access shouldn't automatically include permission to act. Someone may need to inspect an error and explain its consequence while remaining unable to alter a transformation rule or send a production request back through processing. That distinction leads to a practical question. Once support can identify the fault and propose a recovery, who can approve a rule change and who can authorize replay when that action could change the plant's recorded stock. I'd separate authority to approve from authority to execute. A support engineer might carry out an approved recovery, but the business owner should determine whether that recovery remains appropriate for the affected order and its current conditions. Assign ownership to decisions, not simply to applications. Production owns the criteria for declaring its assigned work finished quality owns release permission, and the inventory owner approves the rules under which output becomes a posted stock transaction. Order closure needs an accountable owner as well. Where closure depends on several departments, name the role that accepts the remaining obligations as resolved, rather than giving every department unrestricted permission to close the same order independently. Delegation should preserve those boundaries during nights and weekends. Define who can act when the usual owner isn't available, which decisions that delegate can approve and which conditions require the work to wait for further authorization. The connection itself needs a jointly approved operating boundary, information technology and operational technology, or IT and OT should agree on permitted access paths with plant staff, establishing the constraints that protect ongoing production and equipment availability. A reachable endpoint isn't permission to use it. Agree. Which component initiates communication, which destination it can reach, and whether access requires an approved gateway or another controlled route through the site's existing network boundaries. Support access needs the same discipline as application access. A temporary troubleshooting connection should have an owner and an expiry, with production agreeing when work can occur. If that access could affect a running system or shared equipment. Plan. Changes around the plant's operating requirements, certificate replacement firewall updates and identity changes can interrupt valid connections, so the approved maintenance procedure needs a way to verify access without experimenting against active production transactions. Give each integration component a recognizable workload identity. Microsoft Entra can authenticate supported applications and services, letting you distinguish the component, submitting a request from the person who approved the business action behind that request. Use managed identities where the participating services support them. They let eligible Azure workloads obtain access tokens without storing application credentials in code, though the destination still needs to support that authentication and grant the required permissions. Authentication only establishes who's calling scope, authorization to the operations and resources that component actually needs, so a service permitted to submit production receipts doesn't automatically receive permission to alter master data or administer the receiving application. Older interfaces may still require a stored credential. Keep those secrets in a protected store, such as Azure Key Vault, where suitable. Restrict retrieval and assign responsibility for rotation before an expired credential stops an otherwise permitted exchange. Human approval shouldn't disappear behind the service account. Preserve the identity of the approving person alongside the workload identity, while ensuring that a caller can't claim approval merely by inserting someone's name into an unverified message field. Contract changes need an approval path before deployment. The owner of the affected production, meaning should approve its interpretation, while the application owners confirm that the proposed change remains within their supported interfaces and authorized business responsibilities. Replay rights should be narrower than general support access, grant them to designated roles and require approval when the selected recovery could create a stock movement or change a result that another department has already relied on. Set that threshold before an incident creates pressure. Routine recovery may follow a pre-approved procedure, while exceptional replay needs a named approver who can assess the affected business action rather than simply authorize another technical attempt. Manual overrides need boundaries that people can enforce. Specify the permitted change, its business scope, and the duration of any temporary permission with additional approval. Where the override bypasses a control that normally requires another department's authorization. Keep an audit record of the decision and execution. Record the reason approver acting identity affected. references and observed outcome and restrict alteration of that record. So the person performing the override can't quietly rewrite its history. These controls belong in the working design as we follow the production information into ERP. Each handoff should identify whose authority it uses and what evidence remains when that authorized action finishes or needs further attention. Start at the machine's output signal. A sensor detects a part leaving the fixture, but the execution application still needs the active job operation and recording conditions before that signal can support a production. Fact. The signal doesn't carry all that context itself. Mess or approved logic close to the equipment associates the observation with the work actually running and determines whether it qualifies for recording under the configured execution rules. For example, a setup cycle might also trigger the sensor. The interpreting application needs to distinguish that cycle from production output using the approved operating context, rather than assuming every detected movement adds another part to the order. That interpretation belongs before the business handoff. By the time the source publishes an execution fact, it should have resolved which work the observation describes and committed the corresponding production record through its supported transaction process. We then reach a less obvious failure boundary, saving the execution record and separately sending a message creates an interval where the source can stop after saving production, but before handing its event to the integration. A transactional outbox can close that particular gap where the source application supports it. The same local database transaction saves both the production change and an outgoing event record, so either both records commit or neither does a separate publisher reads those committed outgoing records. It hands them to the messaging service and tracks delivery progress, allowing the source transaction to finish without requiring the remote messaging service to participate in that database commit. That protection depends on actual source support. Adding an unrelated table beside a packaged mess won't create atomic behavior unless the application writes that table within the same transaction that records the production change itself. Some systems won't provide that option. Use a supported capture mechanism instead, then reconcile the captured records against the source so a production transaction omitted during an interruption can be discovered and handed onward. 01:29:58 Speaker: The capture position needs durable tracking to whether the mechanism reads an application journal or supported change. Feed recovery must establish which committed source records have already entered the exchange and which still need capture. Now the event reaches the Azure processing boundary, the receiving component checks the caller's identity and validates the event against its supported contract, resolving the business references needed for the intended action before releasing valid work for processing. Unresolved input stays outside the executable work path. A structurally readable event with an unrecognized production reference should produce an explicit validation outcome. Rather than becoming a partly interpreted receipt. Request that another component must guess how to complete validated work, then enters its assigned queue for this receipt flow. A service bus consumer retrieves the pending item and prepares the receiving application supported request using the approved interpretation of the original execution fact. Keep that preparation distinct from invoking ERP. The worker should retain which source event produced the request. Then call the supported receiving interface without treating a direct right into Erp's database as an equivalent shortcut, ERP evaluates the request within its transaction boundary. If it accepts the posting, the integration records the returned business transaction reference against the request. An original event establishing a navigable relationship from execution evidence to inventory evidence. A refusal takes a different recorded path. The receiving application's explicit rejection remains linked to that same event, with the refused action and reason preserved. So the journey ends in an explainable outcome rather than an unexplained missing stock movement. An unanswered call leaves the journey unfinished. The processing record must distinguish that uncertainty from either confirmed outcome, so the event's history never claims a posting solely because the worker attempted the receiving call. We now have more than a connection between two applications. We have a source fact, a durable handoff, a validated request, and receiving evidence that together explain what crossed the boundary and what ERP actually accepted. That recorded journey can also feed reporting. Analytical. Consumers can receive execution facts and posting outcomes through an approved distribution path without becoming another required participant in the inventory transaction or delaying its completion while they refresh. The reporting copy follows the transaction. It doesn't authorize it. That separation lets us combine production and inventory evidence for analysis, while leaving each business decision with the application and owner responsible for recording it for reporting. I'd begin with the questions people need answered. Microsoft Fabric can bring analytical data from production quality and inventory together, but the combined data needs definitions that explain what each source contributes to those answers. Keep source lineage attached. As the data moves through fabric for a reported amount, people should be able to establish which source records contributed, which transformations applied, and which version of the reporting rule produced the result. Built in lineage helps describe supported connections between analytical items. Detailed traceability back to individual production records still depends on retaining the relevant references through ingestion and transformation, rather than assuming the platform can reconstruct relationships you've discarded. Freshness belongs in that design from the start. Agree how current each source needs to be for the intended decision. Then distinguish the expected update interval from the latest update that actually completed successfully. A successful refresh can still contain old source data. Record how far each source has progressed, so a combined report doesn't present a recent processing time as proof that every contributing application supplied equally recent information that matters when measures appear together. A production feed, updated this minute and an inventory extract from earlier can both be accurate. Yet their difference may describe unequal freshness rather than unresolved business work. Give each shared measure a written business definition. Produced should state the execution boundary it measures and how it treats repeat processing, with the calculation matching the particular question that the report claims to answer. Accept it needs an equally precise reporting rule. Establish which recorded disposition qualifies, how the measure handles a later change, and whether the result represents the current accepted position or acceptance recorded during a selected period. Those are different measures even when their names sound similar. A period total describes activity within that period, while a current position describes what remains in a condition. So name the distinction before people compare them. Posted should derive from confirmed inventory transactions. Its reporting definition needs to account for the approved transaction types and subsequent adjustments, so the displayed quantity represents the intended inventory position rather than a count of successful interface calls. Available requires agreement about the decision it supports. Define whether it means stock eligible for a particular requirement or a broader inventory position, and identify the authoritative availability calculation that the analytical measure represents. I wouldn't quietly recreate that calculation in a report where ERP supplies the relevant availability result, preserve its meaning and freshness. Where an analytical calculation is necessary, document its assumptions and name it so nobody mistakes it for transactional permission. Power BI can present these measures through separate reporting Perspectives. Execution progress should answer what work remains against the defined production requirement without expecting the listener reviewing that information to infer warehouse readiness from an execution percentage, stock readiness answers a different operational question. It should express the quantity eligible for the stated purpose under the Agreed Availability rule, with its applicable scope and freshness clear enough to support a responsible decision. Unresolved mismatches deserve their own reporting perspective. Bring the conflicting measures together with their comparison, boundary and known explanation, so users can distinguish and identify timing difference from a difference that still requires investigation by its owner. Separate perspectives can still share the same governed measures. The point is to organize answers around decisions while keeping a consistent definition wherever the same measure appears. Instead of letting each department maintain a slightly different calculation, there's a limit to what that consistency establishes. A report can explain the recorded position and help someone choose the next action, but its analytical copy doesn't acquire authority to approve or commit that action. Treat reporting availability as evidence with the stated age before an irreversible business action, the responsible workflow still needs the current checks required by its transactional owner, especially when stock conditions can change between analytical updates. This leaves planning with a better starting point. The figures now have defined meaning and freshness, but knowing the current position still doesn't establish whether the plant can complete the next requirement within the time available. That's the next question I'd put to the architecture. Once reporting explains what the plant has recorded, planning needs to determine what the plant can actually do next. Rather than treating a clean production report as a feasible plan, a scheduler needs to place work into time that actually exists. Advanced planning and scheduling or apps uses a model of capacity and production constraints to work out when operations can run and which resources can perform them. Feasibility depends on what that model includes. If it checks machine capacity but ignores a required operator qualification, the resulting schedule may fit the equipment while assigning work that nobody on that shift can perform. I'd ask which constraints the scheduler actually enforces change over time. Resource calendars and permitted operation sequences can all affect the result, but their treatment depends on the selected system and the rules your organization supplies. ERP still has a central role here. It supplies demand, material requirements, inventory positions, and business transactions, and it may also provide substantial planning and scheduling functions depending on the application and its configuration. Start by checking those existing planning functions. A separate apps application makes sense when the production problem needs capabilities beyond what the configured ERP planning process can represent, rather than because another planning product sounds more specialized. Consider what remains on our machining order. The scheduler needs the outstanding operations and their expected durations, including whether the remaining work requires another setup or can continue under conditions already established at the resource. An order level completion percentage won't supply those details. Two orders showing similar progress can require very different amounts of remaining capacity, because one needs a short final operation, while the other still faces a lengthy, constrained process released quantities influence which work can start downstream if only part of the required material has been released. The scheduler needs to know whether the next operation permits a partial start or requires the full quantity. Together, that choice changes the schedule itself. A partial start may bring some work forward, but it can also require another setup when the balance becomes available. So the model needs the approved production rule. Resource state adds another kind of limitation. A machine can appear free in the schedule while remaining unavailable for the required work. Because maintenance has restricted its use or a required tool isn't currently available, a stopped resource needs more than a stopped flag, planning needs an assessed return time, or an explicit uncertainty about that return before it can responsibly place another operation into the capacity that appears open. Rework creates demand for capacity that the original route may omit. In our example, the corrective work needs an assigned operation, an estimated duration, and an eligible resource, rather than simply keeping the affected quantity out of available output. Otherwise, that capacity remains falsely free, the scheduler can assign another job to the same period, leaving the supervisor to resolve a conflict between the published schedule and corrective work already waiting beside the machine. The next design decision concerns when to reconsider the schedule. Updating every input continuously doesn't mean issuing a revised dispatch sequence after every small change, especially when operators have already prepared tools or staged material. Schedule changes have a physical cost. Moving a job can waste preparation and interrupt coordination between departments, so an update policy should distinguish information worth recording from a change large enough to justify another scheduling decision. Choose meaningful triggers with the production team, a revised resource return time, newly released material or newly authorized rework might prompt review when it affects a committed operation, while a minor timing variation may wait until the next planning cycle. Review doesn't have to mean automatic rescheduling. The system can assess the affected commitments and propose a change, while the planner decides whether its benefit justifies disturbing work already prepared or underway on the floor. Protect the near-term schedule where the process requires stability. Define which operations remain fixed, who can authorize an exception and which later work. The scheduler may move freely when incorporating the changed production facts into its next calculation. This gives context a practical test if the scheduler repeatedly needs relationships between operations, qualified resources, release material and corrective work, ask whether the current planning interfaces can supply those relationships clearly enough. A richer model should answer a specific planning question would a digital twin or knowledge graph help maintain that context, or would it duplicate relationships that your existing production and scheduling systems already manage adequately? I judge that richer model by the questions it can resolve. If you need to ask how a restriction on one lot affects a particular order, the model must represent that connection and its current business, meaning. A digital twin models, production entities and their relevant state. Here, that could mean a software representation of an order and its operations, with recorded conditions attached to the specific entities those conditions actually describe. A knowledge graph makes the relationships between those entities explicit. It can express that an operation belongs to an order, ran on a particular machine, and produced a lot affected by a recorded quality decision. The two approaches can work together. A twin can use a graph to represent relationships, but neither label guarantees that the model contains the evidence needed to answer a production question correctly. Model only the context the intended decision requires. For this integration problem, that means connecting the order to its operation, the operation to its execution resource, and the resulting lot to its applicable restrictions and posting evidence. Those links need business meaning, not just technical references. A relationship saying that a hold affects a lot should identify the affected portion and applicable period, rather than implying that every part ever associated with that lot remains blocked. Evidence should travel with the represented state. If the model describes output as received, it should point to the receiving application's posting record, allowing someone to verify the assertion rather than trusting another unexplained status. That gives industrial AI a more useful starting point. A generative AI assistant can retrieve related records and explain their apparent relationship, provided its tools expose the necessary evidence through access paths the organization has approved. Suppose your planner asks why a particular order remains unavailable. The assistant could retrieve the operation record, follow its relationship to the relevant lot, and check whether the applicable hold and inventory posting. Explain the recorded restriction. The response should separate evidence from interpretation. It might identify a current hold as the documented restriction, cite its source and state whether the available posting evidence supports or contradicts that explanation for the specified stock. Missing evidence needs an equally direct answer. If the assistant can't retrieve the posting record, it should report that limit rather than infer that no receipt exists merely because its search returned nothing. Permissions must apply when the assistant retrieves records. If the planner lacks access to detailed inspection findings, the governed retrieval layer should return only permitted information rather than relying on the language model to conceal restricted details. I'd also require the answer to expose its references, the planner needs a route back to the underlying records with their applicable times, so a fluent explanation can be checked before anyone acts on it. Language models can still connect the wrong facts. Similar lot descriptions, incomplete relationships, or stale records can produce a plausible explanation, so governed retrieval reduces avoidable errors without making the generated answer authoritative by itself. Keep business authorization outside the generated explanation. Deterministic checks should evaluate defined conditions, such as whether the requested action concerns the affected lot, and whether the current restriction permits that action. Under the approved rules, an authorized workflow then controls the actual change. If an assistant proposes releasing stock, that proposal should enter the same approval process as another request, with the designated Quality Authority retaining its decision and the application enforcing it. The assistant shouldn't possess broader rights than its task requires. A tool that retrieves receipt evidence doesn't need permission to create receipts, and permission to draft a proposed action shouldn't automatically include permission to execute that action. Even an approved request needs current checks at execution conditions may change after the assistant prepares its explanation, so the receiving workflow must validate the intended action against the authoritative position before committing a posting or release. I'd keep generated advice and committed outcomes separately traceable. That lets a reviewer establish what the assistant recommended, what a person authorized, and what the receiving application actually changed without treating those as one decision. A convincing answer is only part of the proof. We still need to establish whether this architecture preserves its boundaries when records are incomplete, applications refuse requests, or are seemingly reasonable recommendation conflicts with the actual production position. I'd prove that with a deliberately narrow production trial, pick one product family and one approved route with enough variation to exercise the difficult cases without bringing every exception in the plant into the first release. Put names against the acceptance decisions before testing begins. The production, quality and inventory owners should agree which completion milestones the trial must demonstrate, and who can accept the evidence for each milestone. Use actual order examples to write the expected outcomes in advance for each test state. What should change, what should remain unchanged, and which difference between S and P would be legitimate at that point? Begin with partial output rather than a perfectly completed order. Confirm that the trial represents the completed portion correctly, while leaving the remaining requirement actionable, without allowing a partial result to imply that the entire order has finished. Then send some of that output through rework. The test should establish whether the additional work reaches the correct execution record, and whether its completion changes the affected quantity without increasing the number of distinct parts. Deliberately submit the same receipt request again, check the actual inventory documents afterward, including after restarting the processing component. Because a duplicate warning alone doesn't prove that the receiving application recorded only the intended quantity. Reverse the arrival order of dependent messages in another test, verify the resulting business position and retained history rather than accepting the test simply because every message eventually leaves the queue without an error. Disconnect the exchange during a controlled outage, test production staff should demonstrate the approved recording procedure, then observe recovery through to confirmed outcomes, including a pending confirmation that survives a restart before the connection returns. A refusal belongs in the trial. To arrange a permitted test condition that causes ERP to reject a receipt and require the assigned owner to resolve it through the proposed support procedure without entering the production work. Again, follow that with a correction discovered after the initial posting, check that the approved adjustment reaches its intended outcome and remains connected to the original records, including when recovery pauses between the required correction steps. Acceptance needs evidence from each of those cases. The quantities should balance within the chosen process boundary, and every intended posting should lead to a receiving document or an explicitly unresolved outcome with an assigned owner. Require someone outside the build team to explain the differences. Give a plan of the trial records and ask why the two systems show different positions, then check whether the explanation agrees with the expected business outcome. Recovery must also work without the developer directing every step. Have the designated support person follow the procedure, locate the affected request, and complete an authorized recovery while recording enough evidence for the business owner to verify it. Only then move into shadow comparison. Let the proposed exchange calculate its intended actions alongside the existing process, without committing competing transactions, and compare its predicted outcomes with the authoritative records for the selected scope. Investigate differences before advancing to cutover. An unexplained mismatch shouldn't become an accepted exception merely because the overall totals look close, or because the project has already booked its production deployment window. At cutover, identify the exact boundary between old and new processing. Establish which outstanding confirmations remain with the previous path which enter the new path, and how staff will prevent either path from claiming the same work. The rollback path needs the same precision. Reverting an application deployment doesn't reverse inventory transactions already committed, so the plan must account for accepted postings and pending work before the previous process resumes. Responsibility. Set the stop conditions before anyone switches paths. Decide which observed failures require pausing the rollout. Who can authorize rollback and how production will handle the affected orders while the team establishes their recorded position. Finish the handover with monitored operation, not a document transfer. Support should demonstrate that it can detect a pending case, identify its owner and verify resolution while the project team remains available for agreed escalation. Expansion should depend on that observed performance. Call the first flow finished when its owners can account for the quantities. Explain the statuses and recover the tested failures without relying on undocumented intervention from its original developers. Shared production reality needs owners and evidence. Your shipment promise should depend on stock that's released and available for that shipment, with agreed milestones and traceable facts showing why the quantity qualifies when dispatch commits it. Which event makes an order done in your plant? I'd like to compare that definition with the event you're shipping team actually relies on.