iFAKT - Copilot on the Shop Floor - What Industrial AI Can and Cannot Do Today 00:00:00 Speaker: Imagine a night shift where machine four stops mid-order. A technician asks the plant's co-pilot for help, and it gives a clear answer from a service manual for an older version of that machine. The answer sounds useful, but the instructions don't apply. Reliable assistance depends on the exact asset approved sources, and enough current plant context to tell whether the advice fits the situation in front of you. So which tasks can you safely hand to an assistant today? Let's start with the decision facing that maintenance team. For this episode, we'll use a hypothetical plant with mixed age equipment, a maintenance crew shared across several production areas, and customer orders with little room for delay. Nothing about that setup requires an unusual factory. Some machines have been upgraded, while others still run with the controls they shipped with. And the people who understand those differences aren't always working the same shift. That's a normal operating constraint. Machine four stops while producing parts needed for an order due soon, and the operator calls maintenance because the usual approved response hasn't cleared the fault, the technician now faces a choice should they inspect the fault further, call a specialist, or tell production that the machine needs to stay out of service until someone can assess it. None of those choices comes free. An inspection takes time from another job, and calling a specialist might mean waiting, while a production supervisor asks whether the order can move to another machine. Meanwhile, parts aren't coming off machine for. I'd start the design discussion right there with the choice the team needs to make and the information that could change that choice rather than with a chatbot demonstration. What would actually help the technician? Maybe they need to find the relevant service information without searching several folders, or collect the operator's description and recent maintenance notes before calling someone with deeper knowledge. That's useful work for an assistant. The point isn't to make the conversation sound clever, it's to reduce the effort needed to gather evidence so a qualified person can decide what to inspect. Who to involve, and whether production should remain on hold. Now consider the production supervisor standing beside that technician asking for an estimate, because another team needs to know whether to keep material and people ready for the order, they need a different answer. The technician might have enough evidence to describe the fault, but still lack enough evidence to estimate a repair time that production can use with any confidence. Those are separate decisions even beside the same stop machine, and the assistant can help keep that distinction clear by collecting what the team knows, recording what still needs checking and preparing a handover that doesn't turn an early guess into a commitment. A neat sentence shouldn't become a delivery promise, and there's pressure to restart, and there's pressure to give an answer if the assistant fills gaps with plausible advice. The consequences can include an unsafe intervention, damaged parts or a restart that produces scrap before anyone notices. Holding production also carries a cost, because a decision to wait can affect later operations and leave the planning team with less time to protect the promised delivery date. Caution still needs evidence. That's why I define success around the work. Can the team find the relevant information, understand its limits, and reach the right person sooner without adding another source of confusion during a breakdown? You might decide that the assistant can gather records and draft a maintenance request, while the technician still decides whether that request accurately describes the condition and the work needed. That's a clear division of responsibility. Another plant might permit a narrowly defined administrative action. Once required, checks pass, but that permission needs an agreed boundary and a person who owns the workflow. Authority shouldn't grow just because the conversation feels natural. We'll examine assistance with maintenance, work instructions and quality investigations through that lens. Asking what evidence the assistant needs and what decisions remain with the people responsible for production. The scope matters more than the label. I also want to separate useful help from extra work. If a technician spends longer correcting an answer than finding the information directly, you've added a task during a breakdown rather than removed one. So keep the stopped machine in mind as we work through this, because every proposed capability needs to earn its place in that technician's actual sequence of decisions. A fluent answer is only one small part. Before we assign any of those jobs to industrial AI, we need to separate the different methods grouped under that name and understand which part of the problem each method can handle. They don't all do the same work. Explaining an alarm, estimating failure risk, and finding another production slot. Sound related. But each question asks for a different kind of result, so I'd separate the methods before deciding how an assistant should bring those results to the person asking. Start with generative AI, which produces text from the information and instructions supplied. A language model can turn a technical explanation into plain language, draft a fault description, or help someone ask a more precise question about an alarm. That's a language task. When you ask what an alarm means, the model can describe the documented condition and explain technical terms in words that fit the person who needs the answer. Its strength here is working with language. It can connect a short question with a longer explanation, adapt the wording and keep a conversation moving. When someone needs to clarify a term or ask a follow up prediction answers a different question. Instead of asking how to explain an alarm, you're asking whether measured condition suggests a particular event might occur within a defined period under specified operating conditions. A predictive model estimates that outcome from patterns in data for a maintenance task. Those patterns might involve vibration readings and operating load, with the model tested against the particular failure or condition you want to detect. The estimate needs a defined meaning. A prediction about failure during the next production run differs from a prediction about failure during the next month, even if both appear beside the same machine name. Testing also has to match the task. You'd want to understand how often the model misses the condition you're concerned about, and how often it raises an alert when that condition doesn't follow. Those errors create different kinds of work. A missed warning might leave a fault undetected, while repeated false alerts might consume maintenance time and eventually cause people to stop paying attention to the prediction. A fluent explanation doesn't establish predictive accuracy. The language model can explain an estimate from a tested predictive model, but the evidence supporting that estimate comes from the predictive models, data and evaluation. Optimization handles another kind of question. Given a set of choices, defined limits, and an objective. An optimization method searches for a choice that meets those limits and performs well against that objective. Finding another production slot belongs in that category. The task involves evaluating where an operation can fit within available capacity and the stated production requirements, rather than writing a convincing account of where it might fit the objective changes the answer you should expect. A search aimed at protecting the earliest delivery date can produce a different result from one aimed at reducing set up effort, even when both use the same available capacity. Sometimes the limits leave no workable choice. A scheduling method should be able to report that conflict, because an honest result that nothing fits gives the planner a reason to reconsider a requirement. Return to the alarm on machine for for a moment. A language model might explain the fault description. A predictive model might estimate a defined failure risk, and an optimization method might evaluate another slot for the affected operation. Each result needs its own kind of check. You check an explanation against its supporting information. Assess a prediction against task specific performance evidence, and assess and optimize choice against the limits and objectives supplied. These methods can work together. An assistant could receive a question in everyday language, pass a suitable request to a predictive model or optimization method, and explain the returned result without changing its meaning. That combination gives the conversation a useful role. Someone could ask whether a maintenance concern affects an order, then receive an explanation that keeps the estimated risk separate from the production choice being evaluated. I'd ask which method produced each claim. If the answer describes failure risk, I want the estimates basis. If it proposes a production slot, I want to know which limits the scheduling method actually checked. The assistant becomes the way people reach those methods. Its conversational ability can make them easier to use, but we still need to define what that assistant connects to and which tasks we expect it to handle. The word copilot needs a working definition. Before you choose a system in a project meeting, someone might use it to mean a workplace assistant, while someone from production assumes it means an application connected to equipment records and shift activity. Those expectations lead to different projects. If you leave the term undefined, you can approve a demonstration that answers office questions well and discover later that production expected access to information the application never received. I use copilot here to describe an assistant pattern. You give someone a conversational way to ask for help with a task, and the application provides that help through the sources and functions you've chosen to connect. Microsoft also uses copilot in specific product names. Those products have different intended uses and connection options, so a capability demonstrated in one product shouldn't become an assumption about another product carrying the same word. The product name doesn't settle the scope. You still need to check the particular service, its configuration, the available connections, and the licensing terms that apply to the application you intend to build or buy. Consider a general workplace assistant used by a production manager. Depending on its setup and permissions, it might help prepare a meeting brief from accessible workplace content without any connection to the system's recording. What's happening in production. That can be a perfectly reasonable use. The mismatch starts when someone expects the same assistant to answer a specific question that requires records outside its connected sources, simply because both tasks involve the factory. A document assistant has a more defined information boundary. You might configure it to answer questions from a selected collection of engineering documents, giving users a focused way to find and discuss information within that collection. Its answers stay within that designed scope. Access to documents doesn't itself provide access to current work orders, and a question about an unresolved maintenance job might require information that this application can't retrieve a plant specific application can reach further when you build the connections, it could combine a conversation with selected records and approved functions from plant systems, but those connections require design work rather than a broader interpretation of the word copilot. I describe that application by its actual job. An assistant for checking the status of maintenance requests tells your team much more than our industrial copilot, because people can test what checking status means in practice. Then you can ask what happens during a conversation? Does the application only return information, or can it also prepare something for another system? And where does the user's responsibility begin once that preparation finishes? Those choices shape the workflow people will actually use an assistant that prepares an entry for review differs from one that sends the entry onward, even if both respond with the same friendly sentence about helping you. Sources define what the assistant can draw from permissions. Define which parts a particular person may access, while connected tools define which functions the application can request when answering that person's question or carrying out the agreed task. Workflow design ties those parts to someone's work. You decide what information the assistant asks for, when it returns a result, and what the person must do next, rather than leaving those steps to conversational improvisation. So I'd write a task statement before comparing products, name the person using the assistant, describe the result they need, and state the further step the application may take without another person's involvement in that workflow. For our plant, that statement should sound quite ordinary. A maintenance coordinator might ask for the status of an existing request, with the assistant permitted to retrieve and explain that status but unable to change it. That gives you something specific to build and test. You can check whether the required information is reachable, whether the intended users can access it, and whether the application stops at the boundary you've assigned. You can expand that job later through a separate decision. Starting with a clear assignment lets you judge each proposed addition against a real task, rather than treating a longer conversation as evidence that the assistant deserves more authority. Let's give maintenance that same level of precision. Instead of asking for a co-pilot that helps with breakdowns, we need to choose the first piece of maintenance work that the application should actually take on. I choose preparing the fault hand over as the first maintenance task because it gives the crew something they already need without asking the assistant to decide how to repair the machine. Start with the information someone would otherwise gather by hand. Back at our hypothetical machine for the operator reports that the machine stopped during a cycle, and that the same alarm appeared again after the permitted response. That's the starting report, not a diagnosis. The assistant can help capture that report without polishing away the uncertainty, keeping the operator's own observations separate from any explanation someone adds during the conversation, stopped during the cycle describes an observation. The drive failed claims a cause. A useful follow up asks what the operator actually noticed before the stop, rather than suggesting a fault and encouraging them to agree with it under pressure. The wording of the question matters. For example, the assistant might ask whether the operator noticed anything unusual before the alarm, then record the reply as an operator report rather than an established equipment condition. An unanswered question should remain unanswered. Next, it can assemble the recorded alarm history around the reported stop so the technician doesn't have to reconstruct that sequence from several separate records while also taking a call. Keep the sequence intact. An earlier warning may help the technician decide what information to examine, but the assistant should present the recorded events without turning their order into an explanation of the failure. An alarm sequence gives clues, not a confirmed cause. The handover should distinguish the time of the recorded alarm from the time the operator described it, because those events can differ even when both refer to the same stop. That distinction helps the next person follow the account. Then the assistant can locate the relevant material in the approved service manual and return the procedure reference alongside the fault description and alarm sequence already collected. Finding the reference is a useful result on its own. I'd keep the manuals, procedure, title and revision attached to that result, together with the asset record used to select it and the event time under discussion. Those details let the technician check what the assistant actually used. Instead of returning follow the fault procedure. The handover could identify the approved manual name, the relevant procedure, and state that it concerns the reported alarm on machine for the technician now has a specific reference to verify. There's also a practical difference between locating a procedure and extracting an isolated line. That sounds like the next repair step, especially when the surrounding text contains conditions the technician needs. Give the technician access to the complete approved procedure. For the starting task, I'd have the assistant prepare a short handover that separates the operator's report, the recorded events and the service reference, rather than merging everything into one flowing account. Clear labels prevent accidental changes in meaning. The draft might explain that the operator reported a repeated stop, that the event record contains a particular alarm sequence, and that an approved service reference has been located for review. It shouldn't conclude that a repair will resolve the fault. Before anyone passes that hand over to the next crew. The technician checks whether it reflects the reported condition and whether any wording suggests more certainty than the available record support. Reviewing a short account should take less effort than rebuilding it. Suppose the technician hasn't inspected the machine yet, but the draft implies that the equipment has already been checked and that work can begin once the next person arrives. That wording needs correction before I hand over the plant's isolation. Procedures still govern any inspection or intervention, and the qualified technician remains responsible for checking conditions and following the approved requirements before work starts. A prepared handover doesn't establish a safe working state. This gives maintenance a modest but testable starting point, a coherent account of the fault, the relevant recorded history, and a service reference that the receiving technician can inspect. No repair verdict is needed. There's still a condition attached to that service reference, though, because finding a manual associated with machine for doesn't establish that every instruction in it applies to the equipment installed. Today, the technician needs to confirm that match. That takes us from finding information to checking its fit. Because the next question concerns the actual machine the technician faces, not just the name used in the handover. Which equipment does that procedure cover? Machine four might only be a local name. The manufacturer identifies the equipment by model and serial number. While your maintenance records may describe changes that no longer match the configuration listed when the machine arrived, the equipment family gives you a starting point to judge whether advice applies. You need the installed machine's identity and the changes that affect the procedure, including its controller version, when that changes how faults behave. Suppose our hypothetical plant owns two machines from the same series. One still uses its original drive system, while machine four received a replacement drive and controller update during a retrofit that changed some diagnostic behaviour. The outer housings might look almost identical. A technician could recognise the equipment family correctly and still receive the wrong guidance. If the assistant selects service information without checking which control system runs that particular machine. That difference can change the meaning of an alarm. In this hypothetical case, the same displayed fault code could point to different checks across controller versions, because the supplier changed the conditions that trigger the code, the assistant needs that distinction before interpreting the code. Its response should connect the reported alarm to the installed controller version, rather than assuming that matching words and matching equipment names establish the same fault condition. I'd ask which record supports that configuration. A serial number may establish the original build, but a later retrofit record may be needed to establish which components and software actually sit inside the machine today? That record can also have gaps if the maintenance system lists a controller replacement without its version. The assistant has part of the equipment history, but still lacks the detail needed to choose between competing instructions. A useful clarification asks for the missing detail. Instead of asking the technician to provide more context, the assistant could explain that the available procedures cover different controller versions and request the installed version through an approved identification method. Sometimes the technician can't confirm it immediately. The assistant can still present the known asset details and explain which configuration remains unresolved while withholding the version. Specific advice that depends on information nobody has established yet. Repair history adds another kind of evidence. If a previous job on this exact machine records the same alarm, the assistant can bring that job forward as a relevant record for the technician to assess. But a closed job rarely tells the whole story. A note reading replaced sensor running again, records work, and an observed outcome without proving that the sensor caused the stop or that another fault won't produce similar symptoms. 00:17:18 Speaker: Even the fault label may be uncertain. Someone might have selected a broad category to close the job. While the free text notes describe several checks and adjustments, with no clear account of which change restored operation, I'd preserve those limits when presenting the history. The assistant could report that a sensor replacement appears in an earlier job involving this alarm, while making clear that the record does not establish the cause of today's stop configuration. Changes also affect which history deserves attention. A repair from before the retrofit might concern hardware that machine for no longer contains, so its relevance depends on the installed equipment at the time of that earlier event. Conflicting records need a direct response. If one record lists the original controller and another lists a replacement, the assistant should identify that conflict and ask for confirmation rather than choosing whichever entry sounds more complete. An unresolved conflict needs an escalation route. The next step might involve the maintenance engineer responsible for the retrofit records, or a supplier specialist who can confirm applicability from the serial number and installed configuration. That escalation should carry the disputed details with it. The receiving person needs to know which records disagree and which proposed guidance depends on resolving them, so they don't have to reconstruct the question from scratch. Once that match is established, the search becomes more precise. We can look for evidence that applies to the confirmed configuration, rather than treating every document associated with the equipment family as equally suitable for the maintenance question. Now we need to examine how that evidence reaches the answer. The assistant has to find the relevant material and use it faithfully, which introduces a separate question about what happens between locating a passage and explaining it. There's a name for this approach retrieval augmented generation, usually shortened to rag. In plain terms, the application searches a selected collection of sources, then supplies material from that search to the language model so it can draft an answer. The search and the writing are separate operations. That gives you two things to examine when an answer goes wrong, because the application might retrieve unsuitable material or explain suitable material in a way the source doesn't support. Follow a question. Through that process, a technician asks what a documented alarm condition means, and the application turns that request into a search against the service information available for the confirmed machine configuration. Search can use exact terms, meaning or a combination. An exact search helps with a specific alarm code, while a meaning based search can help when the technician describes a symptom using different words from the manual, the application selects passages from the search results. It then supplies those passages alongside the question and its response instructions, giving the language model a limited set of material from which to compose its explanation. The model doesn't necessarily receive the whole manual. That means the quality of the answer depends partly on which passages reach it, including whether they contain the conditions needed to interpret the relevant text correctly. Imagine the search finds a paragraph explaining a pressure related alarm. A separate paragraph states that the explanation applies only during a particular operating mode, but that paragraph falls outside the material supplied to the model. The retrieved text can be accurate yet incomplete. If the answer leaves out that operating condition, the technician receives a broader claim than the manual supports, even though every quoted sentence genuinely appears in the source passage. Boundaries deserve engineering attention. When you split a manual into searchable pieces, each piece needs enough surrounding material to preserve its meaning without pulling in so much unrelated text that the relevant detail becomes hard to find. Warnings need to travel with the text they govern. If a warning sits above several steps, the preparation process needs to retain that relationship rather than treating the warning and each step as unrelated blocks of text readable text also needs checking before search begins. A scanned page might yield a plausible looking sentence while losing a negative word, confusing a unit, or separating a value from the label that explains what it measures. Those errors can survive into an otherwise polished answer. Test the extracted text against representative source pages, paying particular attention to fault codes, symbols, and any layout that carries meaning beyond the order of the words. Revision tags should stay attached to each searchable passage. When a passage reaches the answer, the application should still know which document revisions applied it, rather than trying to recover that identity from the generated wording afterward. Now consider the citation attached to the response. It gives the technician a route back to the source, but it doesn't establish that the source supports every claim beside that reference in the answer. A real citation can accompany an unsupported conclusion. The manual might describe a possible explanation for an alarm, while the generated response presents that explanation as the confirmed condition on the machine being discussed. Checking the reference means checking the relationship. Does the cited passage support the actual claim, including its conditions and degree of certainty, or does it merely discuss the same general subject using similar technical words that check needs to happen. Claim by claim. If an answer contains a documented definition followed by an inferred explanation, the application should keep those parts distinct instead of letting one citation appear to support both. For technical values, direct quotation can reduce unwanted rewriting. The assistant can identify the relevant passage and explain the surrounding terms while preserving the source wording. Wherever a small change could alter the instructions, meaning rack gives you a practical way to ground language assistance. Its usefulness comes from a connected process that preserves source meaning through preparation, search, selection and drafting with checks wherever information can change or disappear. There's still a question before any search runs, someone has to decide which documents belong in that collection, because careful retrieval can faithfully reproduce material that the plant should no longer use for the task. Document control now becomes part of the application's operating design. The team responsible for releasing a procedure needs a way to determine whether that procedure can enter the assistant's working collection and when it must leave. A shared folder rarely expresses all those decisions. It might contain an approved procedure beside an engineer's proposed revision, a supplier brochure, and an old copy someone kept because it includes notes from an earlier installation, those files have different purposes. The draft may support engineering review, and the brochure may describe equipment options, but neither should enter the approved procedure collection simply because someone stored it beside released service information. I'd separate those collections by their intended use if engineers need an assistant for discussing drafts, give that work a defined space that can't quietly feed unfinished content into the assistant used during production. Each released document needs a named owner. That owner takes responsibility for its approval, state the date it takes effect, and the conditions that require its replacement or withdrawal from use across the plant. An upload date won't answer those questions. A file copied yesterday might contain instructions withdrawn months ago, while a newly approved procedure might remain unavailable for use until a planned equipment change takes effect. The application needs explicit release information it can check. That means the connection from document control must carry approval, status and effective dates. Rather than asking the language model to infer them from file names or introductory text, consider a procedure withdrawn after engineering finds an incorrect inspection requirement. The document owner needs to stop further use even if a replacement isn't ready, which means withdrawal must work independently of publishing the next revision. That needs a defined technical path. The release process should identify where the document has traveled and trigger removal or exclusion wherever the assistant could still retrieve it as current guidance for the task. Deleting the original file might leave other copies active. A search index can retain extracted passages, and a cache can retain an earlier response, so both need an update process tied to the withdrawal event. I'd require confirmation that the change reached those locations. Document control should be able to distinguish withdrawal requested from withdrawal. Applied with a named person handling failures that leave the assistant's collection behind the approved release state. Saved conversations need attention to a technician. Might reopen a discussion from the previous shift and ask a follow up. While the conversation still contains text from a procedure that changed between shifts. Old conversation text should remain identifiable as historical content before using it. To support fresh advice, the application needs to check its source status and retrieve the currently permitted material for the task being discussed. That doesn't require silently rewriting the record. Keep the earlier exchange under the plant's record rules, but tell the user when a reference procedure has changed. Instead of allowing yesterday's answer to pass as current guidance. Now, suppose two approved sources disagree about a requirement. A local procedure might specify one inspection interval, while supplier service information specifies another, leaving the assistant with a conflict that document owners need to resolve. The response should expose that disagreement. In plain words, it can identify both documents, state the conflicting requirements, and direct the question to the responsible owner without blending the two into a new instruction. Nobody approved. Some plants have documented rules for resolving source precedents. Where those rules apply, the application can follow them explicitly and record the basis for selecting one source, rather than improvising a preference from how authoritative the wording sounds. Other conflicts require an engineering decision. The assistant should preserve the disputed text and its release details for that review, while limiting its response to information that doesn't depend on settling the disputed requirement. This work needs an owner beyond the initial build. Someone must maintain the connection between document releases and the assistance collection, investigate missed updates and confirm that withdrawn content no longer appears as permitted guidance that gives you a controlled basis for the next task. Once the plant can govern which instructions count as released information, we can examine how an assistant helps prepare clearer instructions without becoming the person who approves them. Writing a clearer instruction starts with a specific editing task. You give the assistant approved engineering content and ask it to prepare wording for a defined audience, while keeping the technical requirements intact. For the people who review the draft, the useful change might be quite small. A dense paragraph can become a set of direct steps, with each action stated plainly and unfamiliar terms explained using language that operators already use in that plant. Take a hypothetical instruction for replacing a fixture during a changeover. The engineering source describes prerequisites, the removal sequence, installation requirements and checks that must pass before release, but packs much of that information into long sentences, and the assistant can help separate those actions into a readable draft. The person reviewing it can then compare each proposed step with the engineering source. Checking that clearer wording hasn't changed what someone must actually do. Step order needs deliberate protection during that rewrite. If the source requires a condition to be confirmed before a component is removed, the draft must preserve that dependency rather than treating both actions as interchangeable tasks. Even a small edit can change the required sequence replacing after, with, and might make two actions sound independent, so the reviewer needs to inspect the relationships between steps as well as the wording within them. I'd give technical values a stricter editing rule, preserve the stated unit's limits and conditions exactly in the draft, then have the responsible reviewer confirm them against the approved source before any instruction enters the release process. A limit isn't just another number in a sentence. An upper limit, a lower limit, and a target. Describe different requirements and a rewrite that substitutes one for another. Changes how someone judges whether the work passes warnings. Need the same care when wording becomes shorter, the draft should retain the hazard, the required precaution and the point where that precaution applies. Rather than trimming the warning into a general request to work carefully. Plain language still needs precise technical meaning. If a term names a particular component or approved method, replacing it with a familiar but broader word can leave the reader unsure which part or method the instruction requires. Translation creates another review task even when the wording sounds fluent. A translated instruction needs someone competent in the technical subject and the target language to check whether the required action remains the same for the people expected to follow it. Negation deserves a separate check. A sentence telling someone not to proceed until a condition is met must retain both the prohibition and the condition, because losing either changes the instructions, meaning. Local terms can also differ from dictionary terms. The word your supplier uses for a fixture may differ from the word operator's use, so an approved glossary can guide drafting without leaving the assistant to choose between competing names. Local practice still needs engineering review. If a translated draft changes a named isolation method into a phrase that sounds familiar locally, the reviewer must check that it still refers to the plant's approved requirement. Keep the draft status obvious throughout this work. The generated wording should enter the same review and release process as any other proposed instruction, with the responsible people checking technical accuracy and suitability before approving its use. I'd ask the assistant to prepare a comparison for reviewers. Alongside the proposed wording, it can identify the source passage and explain the intended edit, giving the reviewer a practical way to inspect changes without searching through unrelated material. That comparison also needs checking against the actual changes. The reviewer shouldn't rely on the assistant's claim that it only simplified wording, because an omitted condition can escape the model's own account of what had changed. Approval remains a recorded decision by the authorized owner. Once the draft passes the plant's required checks. The released instruction enters normal use through the existing process, rather than becoming approved because someone accepted the generated text. That leaves a separate question about using the instruction during production. A release document may read clearly in review, but the assistant still needs to help someone reach the right information while they're carrying out a specific task on shift, put the assistant beside an operator preparing a changeover. In our hypothetical plant, the next job uses a different product variant, and the operator needs to confirm which tooling and setup procedure apply before starting. The operator's question might sound simple. Which setup do I need? But the answer depends on which machine they're working at, which variant they're preparing, and which part of the changeover they're asking about at that moment. The conversation should establish those details without becoming an interview. If the workstation already supplies the machine identity, the assistant can state that identity and ask the operator to confirm the intended product variant before continuing. A shared terminal needs particular care here. An operator might walk over from another machine, so the application shouldn't silently assume that the equipment associated with the terminal is the equipment behind every question. Confirmation should name the equipment in plain language. Are you preparing machine for for this variant? Gives the operator a specific assumption to correct. Rather than asking whether an unexplained collection of background details looks right. Then ask about the immediate task, not the entire changeover. Someone checking the required fixture needs a different response from someone looking for the acceptance check after installation, even though both questions concern the same setup procedure. I'd keep the first response short enough to act on, identify the relevant procedure and tooling reference, then let the operator ask for the specific detail they need rather than reading out every part of the changeover. A spoken answer has a strict practical limit if the operator needs to remember several conditions while standing beside equipment, the assistant should pause and let them return to the approved procedure instead of extending the explanation. That pause shouldn't imply that the work has finished. A person saying next may only want more information, so the application shouldn't interpret that word as confirmation that a physical step has been completed correctly. Shop floor conditions also change how people can use the assistant. Gloves can make small controls awkward, and an operator carrying a tool may need to stop and put it down before entering a question or checking a reply. Voice input can help in some settings, but fans nearby machines, hearing protection and several people speaking at once can make a perfectly ordinary question difficult for speech recognition to capture without an error. Product names and tooling codes deserve extra caution. Two variants might sound almost the same, especially through a headset, so a plausible transcription shouldn't automatically become the confirmed basis for selecting the setup information. Repeat back the detail that could change the answer. The assistant might ask, did you mean variant B and wait for confirmation with another input method available? If repeated voice attempts still leave the intended variant unclear. Don't trap someone in a loop of failed recognition after a failed attempt, the operator needs a simple way to enter or select the identifier, or leave the conversation and reach the normal support route. Language differences add another practical constraint during an exchange. Someone may understand the procedure, but use a local term for the fixture, so the assistant should confirm the intended component instead of guessing from a near match. The reply also needs to fit the person asking. An experienced operator checking a tooling reference shouldn't have to hear a basic lesson, while a less familiar operator needs a clear route to someone qualified to help. That human route should remain easy to reach throughout the task. If the operator can't reconcile the answer with the equipment or the procedure. The assistant should help them contact the shift lead or designated support person without more questioning. Access to the approved procedure should be equally direct. The operator shouldn't need to persuade the assistant to provide the full instruction, and losing the conversation shouldn't prevent them from using the plant's normal method of accessing it. Try this interaction where the work actually happens. Have operators ask their real questions with the usual noise, gloves, interruptions and shared terminals. Then observe where the conversation forces them to stop work or repeat themselves. One more question sits behind this changeover interaction. Even when the operator identifies the machine and product correctly, the assistant still needs a defined basis for any statement it makes about the machines present. Operating condition is machine for still stopped? Sounds like a straightforward question. Answering it requires a machine signal tied to the correct asset, and a timestamp that tells us when the reported condition actually applied, rather than when someone opened the conversation. This is where the Internet of Things or IoT, enters the task. Equipment signals reach other applications through connected systems, carrying information about operating conditions that the assistant can request through a defined read only path. Follow one status update from the machine. A control system records a stopped condition, an approved connection collects that signal, and another service makes the recorded status available for the assistant to retrieve when asked. The signal needs more than a value. Its asset identifier must distinguish this machine for from equipment with the same local name elsewhere, and its event time must travel with it through that path. Arrival time answers a different question from event time. A stopped signal recorded during a network interruption might reach the receiving service after the connection recovers, even though the machine has already changed state again. 00:34:46 Speaker: The assistant needs to preserve that difference in its wording. The latest received record reports are stopped ten minutes ago. Describes the available evidence while the machine is stopped. Claims a present condition that might remain unconfirmed. Historical records still help, but they need their own boundary. A question about the previous shift can use stored events. While a question about the condition now needs evidence fresh enough for that particular task. Define fresh enough before anyone starts asking those questions. The acceptable age for a shift summary can differ from the acceptable age for a current status answer, because those answers support different kinds of work. I'd attach a freshness rule to each supported request. The application can compare the signals event time with the permitted age, and return a clear status, rather than letting the model decide whether an old reading sounds recent. A stale value should remain visible as stale. If the last update falls outside that limit, the assistant can report its time and explain that the current condition can't be confirmed from the available connection. A missing update doesn't describe a machine condition. Silence might mean a broken connection, a collection failure or equipment that sends updates only when something changes. So the interpretation depends on how that signal behaves. That behavior needs an explicit agreement with the signal owner. For a change only feed. The application needs a separate way to establish connection health before treating the last reported state as evidence of the present state. Duplicate events create a different kind of confusion. Some delivery paths resend a message after an interrupted exchange, and counting both copies could make the assistant describe repeated stops when the machine recorded only one. Handle those copies before generating the answer. Where the source provides an event identifier or sequence information, the receiving service can use it to recognize repeat deliveries and preserve the intended event history. Clock differences also deserve a deliberate check if equipment and receiving systems disagree about the time a record can appear newer or older than it really is, which weakens any freshness decision built on that timestamp. Keep the time basis explicit across the connection record which clock supplied the event time account for time zones, and flag timing problems that prevent the application from reliably judging the age of a reading out of order delivery needs similar care. During reconnection, an older stopped event arriving after a new or running event shouldn't replace the newer state simply because the receiving service encountered the older message last in its queue. Keep this first connection read only while testing those conditions, you can learn whether the assistant reports state accurately under delays and missing updates without giving that same connection any ability to change equipment settings or issue commands. Test the failure paths as deliberately as the normal path. Interrupt the connection in a controlled test. DeLay an update and check whether the assistant describes the evidence honestly instead of continuing to speak in the present tense. Knowing the latest recorded state is a useful starting point once its timing is trustworthy. The next question concerns what that machine's condition means for the work connected to it, which requires more than another status signal. Ask which customer order depends on machine four and the information requirements change. The assistant now needs a reliable connection between the equipment, record, the operation assigned to it, and the order that requires the parts from that operation start with how those records identify the same work. Your enterprise resource planning system or ERP might hold a production order number while your manufacturing execution system or mess tracks that order through separate operation and execution records. Those identifiers don't have to be identical. They need a maintained mapping that lets an application establish which execution record belongs to which production order, without guessing from descriptions that happen to contain similar words. That mapping also needs a defined level of detail. One order might split into several production lots, so linking every lot directly to the order won't tell you which lot currently needs the stopped operation. For our hypothetical plant, follow the affected lot it belongs to a production order, requires a particular operation, and has an assignment to machine for that connects the business requirement with the equipment doing the work. Now add the requirements of that operation. Producing this part might require a particular fixture and a worker qualified for the process. So a machine assignment alone doesn't describe everything needed to perform the operation. These are product, process and resource relationships. The product describes what you're making, the process describes the work required and the resources describe what can perform or support that work under the plant's defined conditions. Separate an actual assignment from an approved capability. Machine for might currently hold the assignment while another machine has approval for the same process, but only for certain product variants with specified tooling available. That distinction changes what the assistant can establish. It can report where the operation is assigned, and identify recorded capabilities, while leaving any further decision dependent on requirements that the available relationships actually describe. A knowledge graph can make those relationships explicit. It represents identifiable things such as orders and machines, and connects them through named relationships that applications can follow. When answering a question about the factory, the relationship names carry meaning assigned to describes a different connection from qualified to perform, so the model needs to preserve that difference. Rather than treating both as a general association between equipment and work. Follow those connections from the stopped asset. An application can find the assigned operation, identify its production order, and retrieve the product and resource requirements attached to that operation without asking the language model to invent missing links. You don't need a graph database for every version of this problem. A relational model with well-defined tables and maintain joins may handle the required questions perfectly well, especially when the relationships remain straightforward and the scope stays narrow. I choose the structure around the questions you need answered. Graph based modeling becomes worth considering when following varied relationships becomes difficult to maintain, rather than because a proposal sounds more industrial with knowledge graph in the title. A digital twin addresses a related modeling need for this discussion. It's a maintained digital representation of selected physical assets and their state, built with enough detail to support a defined operational purpose rather than reproduce the whole factory. Its scope needs a plain statement. A twin supporting order impact questions might represent machines, their operating state and assigned work without attempting to model component wear or the physical behavior of every machine part. The relationship model and the twin can work together. The relationships establish what the asset connects to, while the maintained representation supplies the selected asset details and state needed for the questions this particular application supports. Neither model maintains itself just because you've named it. When engineering changes a process route or production changes in assignment, someone must own how that change reaches the representation the assistant uses for its answers. That ownership makes the context usable beyond a demonstration. You can trace an answer about an affected order through maintained records and explicit relationships, rather than accepting a plausible connection assembled from unrelated descriptions during the conversation. We can now ask where Microsoft fits into that information journey. The next architecture decision concerns which services move and organize those records. While the plant retains responsibility for defining what the relationships mean in production. I'd place Microsoft Services by following the records they need to handle. that starts at the plant's existing integration points, where equipment, connections and business applications make selected information available without requiring the assistant to reach directly into every source for machine events. That could mean an approved plant gateway. The gateway passes selected records toward an Azure ingestion service, while the controls team determines which signals can leave the production environment and how that connection operates. Business records can take a different route into the application. A supported connector or an integration service can bring across selected order and maintenance records using the interfaces that the source systems permit, rather than an improvised extraction. Those routes don't all need the same destination or timing. Records prepared for investigation can enter an analytical store, while information needed during a conversation may remain in its source system until the application requests it. Microsoft Fabric can support the analytical part of that arrangement depending on the source and supported connection. Its data integration and storage capabilities can help bring selected production and business records into a shared environment for analysis. The engineering work sits around that movement of data. Your team still needs to map fields, reconcile definitions, and establish what each source contributes. Because loading two records into fabric doesn't resolve a disagreement between their meanings. For a quality investigation, fabric might hold prepared inspection and production history. An approved analytical model can then define how the application relates those records, with manufacturing specialists deciding which comparisons make sense for the process under investigation. Power BI fits where people need governance and reporting from those records. Its semantic models can hold agreed measures and relationships for reports, so the application team doesn't need to recreate a different definition of the same measure for every audience. That reporting role has a clear limit. A measure describing production performance doesn't establish whether a machine can perform a particular operation unless engineering has supplied and maintained the relevant capability information somewhere the application can use. The language application has a separate set of needs for an Azure based implementation, and Azure hosted language model can handle the conversation, while Azure AI search can provide search across the document collection prepared for that application. The application connects those services around the maintenance task. It sends the question through its retrieval process, supplies selected material to the model, and returns the response through whatever workplace application the crew will actually use. There's no requirement to copy every record into the search service. Searchable procedure text belongs there when appropriate. While transaction records can remain in their operational systems or analytical stores reach through the connections designed for those particular records. Copilot studio is another possible part of the implementation. Its agent building tools and connectors may suit a defined conversational workflow, provided the required system's interaction pattern and deployment conditions fit what the server supports in your environment. A custom application may fit other requirements better. If the maintenance conversation needs a specific interaction with a plant terminal or a detailed integration with existing software, your team can build that application around suitable Azure services instead. I'd compare those choices against the actual integration work. A quicker conversational build can still leave substantial work in connecting plant systems, and a custom build gives your team more responsibility for maintaining the application over time. Specialized manufacturing software retains its own job in either approach. The maintenance system records maintenance work, and production applications manage their established workflows, while the assistant provides another way for people to reach selected information and functions. Engineering supplies the manufacturing, meaning that those services need. Microsoft can provide data, search and language capabilities, but your plant must establish which processes are approved, which equipment is suitable, and which operating requirements apply. That distinction helps you assign responsibility when something doesn't work. A missing source record needs attention in its originating system, while a failed connection belongs with the integration owner, rather than every issue becoming a request to adjust the model. One part of this design still needs closer inspection. When the assistant needs a specific maintenance record during the conversation, the application needs a defined way to request it and return the result without asking the model to reconstruct the facts. Give that request a defined interface, an application programming interface, or API. Lets the assistant request a particular record or calculation through an agreed function with explicit inputs and a result the application can inspect. The tool should have a narrow job. For example, a maintenance lookup could retrieve an existing request by its reference number, while a separate production lookup could return the recorded status of a specified operation. The model proposes which function to call application code, then checks the proposed request before sending it, so a conversational interpretation doesn't pass straight through as an unchecked instruction to the maintenance or production system. Consider a question about outstanding work on machine for the application. Might need a validated asset identifier and a defined request status before a maintenance service can return the matching jobs from its records. Outstanding needs a precise meaning in that request. If the maintenance system distinguishes open jobs from jobs awaiting closure, the assistant may need to ask which group the person wants rather than silently treating them as equivalent. Typed inputs help enforce that contract. A tool can require an asset identifier in a specified field. A date in a defined format, and a status selected from values the receiving service accepts. The asset field needs a real validation check. Passing text that looks like an equipment identifier isn't enough. The application should confirm that the identifier resolves to an asset supported by that particular service. Return the facts in structured fields to a maintenance lookup might provide the request reference, recorded status, asset identifier, and last update time separately, allowing the application to check them before the model writes its explanation. That structure helps preserve distinctions during the reply. If the service returns awaiting inspection, the assistant should explain that recorded status without converting it into repair scheduled, which would claim a commitment absent from the returned record. Exact arithmetic belongs in a calculation tool. If someone asks how many requests meet the agreed conditions, a database query can count the matching records. Rather than asking the language model to count a long passage, check whether the response contains the complete result. Some interfaces return records in batches, so the first page of matching jobs can't support a claim about the total unless the service also supplies a reliable. Total limits deserve an equally explicit calculation. Suppose a process record contains a measured value and an approved acceptance range. A calculation service can compare them using the stated boundary rules and return the result. The boundary rule changes the answer. A value equal to an upper limit passes only if the defined rule includes that boundary. So the service needs the actual comparison requirement rather than a loose instruction to check acceptability. Units must remain attached throughout that operation. A comparison service should require compatible units or use an approved conversion, and it should reject a request when the available record doesn't establish what a number measures. Time conversions need the same discipline. When a person asks how long a job remained open, the calculation service should work from defined timestamps and time zones, including any daylight saving changes affecting the interval. Elapsed time also differs from working time. A request opened across a weekend might represent several calendar days, but fewer staffed hours. So the calculation needs the intended definition before returning an answer about maintenance duration. The returned result should expose that basis. The assistant can explain which records and calculation definition produce the answer, giving the person a way to distinguish a recorded fact from a result computed for this question. A failed call must stay a failed call. If the maintenance service times out, the assistant should report that it couldn't retrieve the requested status instead of supplying a likely status from the surrounding conversation. An empty result needs a different message. No matching records returned. Describes a completed search. While the search couldn't complete, describes unavailable evidence, and those outcomes shouldn't receive the same reassuring answer. The same rule applies to measured values. When a production tool returns. No reading, the assistant can't substitute a typical operating value and present it as though the source supplied an actual measurement. These checks become particularly useful during quality work. Once tools can retrieve records and perform traceable calculations, the assistant can help organize an investigation while keeping the measured results separate from proposed explanations of the defect. Suppose inspection flags more surface defects after a change over at our hypothetical plant. The quality engineer needs to establish which production lots are involved, what inspectors recorded, and what changed in the process before deciding where to focus the investigation. The immediate task is assembling an investigation brief. An assistant can help gather the affected lot, references, relevant inspection records and process history into a working account that the engineer can check and extend. Start by defining the reported problem in the inspector's terms. If the initial report describes scratches near a mounting face, keep that description intact rather than expanding it into a general claim that surface quality has deteriorated across the product. Then establish the period the engineer wants to examine that might include production before and after the changeover, with the boundaries stated explicitly so everyone understands which records belong to the comparison and which remain outside it. Affected lots need a careful description. A lot containing a recorded defect differs from a lot awaiting inspection, and the brief should keep those categories separate instead of describing every lot produced during that period as defective, inspection notes are often harder to organize than inspection results. One person might write fine scratches, another might record marks beside mounting face, and a third might use a local shorthand that means little outside the inspection team language assistants can help group those descriptions for review. The assistant can propose categories based on the wording while retaining each original note, so the quality engineer can check whether the proposed grouping actually describes the same defect. Some notes won't fit a category cleanly. A vague entry, such as surface issue, should remain unresolved until someone clarifies it, rather than being assigned to scratches. Because scratches dominate the surrounding reports in the collection, the engineer can correct those proposed groups before using them. That makes the assistant useful for reducing reading effort without quietly changing the defect classification that the plant uses to understand and report its quality problems. Next, bring the inspection results beside the production history, the brief can describe which product variant ran when the changeover finished, and which recorded settings or adjustments fall within the agreed investigation period for those lots. Keep recorded settings distinct from measured process behaviour. A setup record specifying a target doesn't establish that the process maintained that target, so the brief should describe exactly which kind of evidence the available history contains. The comparison needs an agreed measure to account of rejected parts means something different from a count of defects, because one inspected part can contain several recorded defects and therefore contribute differently to each measure. Ask the analytical tool for the comparison the engineer actually needs. It can calculate the selected measure for the defined periods and return the underlying counts, the inspection quantities and the records included in that calculation. The brief should state that calculation in ordinary language. For example, it can explain that the comparison concerns rejected parts among inspected parts. Not every part produced, and identify any lots whose inspection results aren't yet available. 00:52:25 Speaker: Now the assistant can prepare questions worth investigating. If the recorded increase follows the change over. It might propose checking the setup verification record and asking whether anyone documented an adjustment during the first production run. That remains a proposed explanation to examine. The brief should say which observation prompted the question, and identify the next piece of evidence needed, rather than allowing the timing alone to become a claim about the cause. A useful hypothesis comes with a specific next, check if the notes place scratches near a fixture contact area. The assistant can suggest asking the quality engineer whether an authorized examination of that contact area would help distinguish possible explanations. Avoid generating a long catalogue of generic causes. The engineer needs a few questions grounded in these records, with a reason for each question, rather than another page of possibilities that somebody must sort through during production. The resulting brief gives the investigation a workable starting point. It contains a defined defect description, a bounded set of lots, checked comparisons, and proposed questions whose supporting observations remain accessible to the person responsible for the investigation. The next step demands a different standard of evidence. Organizing those observations helps the engineer decide what to examine, but establishing a cause requires testing whether the proposed explanation actually accounts for the defect. For the scratch investigation, I'd first trace the affected parts through their actual production history, including the material used and the equipment involved in each operation. That's the purpose of batch genealogy. Genealogy connects a finished lot to its material lots, processing steps, and recorded equipment assignments so the engineer can compare exposure to possible causes across the parts under investigation, the links need to describe actual production. An approved route describes where paths should travel, but a rework operation or temporary routing change can alter where a particular lot actually traveled before inspection. Missing history leaves an unresolved gap. Suppose the affected lots passed through the same finishing operation while unaffected. Lots followed another route, even though both groups started with material from the same supplier batch. That gives the engineer a comparison to investigate before interpreting that comparison, check how the defect was measured because an apparent change in product quality can also come from a change in how people detect or classify defects. The measurement process needs examination to. Perhaps inspectors started using a different lighting arrangement after the changeover, allowing them to detect faint marks that the previous inspection conditions didn't reveal on otherwise similar parts. More detections wouldn't automatically mean more scratches. Ask whether inspectors reach consistent judgments when examining the same parts under the defined conditions, and whether the inspection method can reliably distinguish acceptable marks from unacceptable scratches. Those checks support the comparison itself. Sampling can distort the picture as well, especially if one period contains routine checks, while another contains extra inspections deliberately focused on parts that operators already suspected might be defective. Those samples answer different questions. Missing results also need a reason attached, because unrecorded inspections might come from a particular shift or product variant, rather than being scattered evenly across the production history. Treating them as ordinary passes would bias the investigation. Now consider two changes occurring together. A new material lot enters production during the same changeover that introduces replacement tooling, and the scratch reports increase afterward. Both explanations fit the timing. This is confounding. The available observations don't separate the effect of the material change from the effect of the tooling change, because both appear together in the records. The assistant should name that limitation directly. The engineer needs evidence that separates those possibilities, perhaps from comparable production. Using the new tooling with an earlier material lot provided the other relevant operating conditions are sufficiently comparable. Historical records might supply that comparison. If they don't, the quality and process engineers can define a controlled test, specifying which factor changes, which conditions remain fixed, and what result would support or weaken the proposed explanation. Production approval still governs the test. For example, an authorized trial could compare two link conditions while holding the material and defined process settings constant with inspection performed under the same agreed method for both conditions. The test needs a clear acceptance basis. I'd also ask what result would contradict the explanation before the trial begins, because otherwise almost any outcome can get folded into a persuasive account of why the suspected cause still fits. A useful test can challenge the hypothesis. The assistant can help prepare the test record and organize the resulting evidence, while the domain specialist judge whether the comparison controls the relevant differences and supports the proposed corrective action that judgment needs. Process knowledge machine vision adds another measurement system to examine rather than removing the need to validate how defects are detected across the actual products and operating conditions involved, its output needs separate performance evidence. A camera based inspection model needs testing against reference judgments on representative parts, including acceptable surface variation, relevant defect types, and the lighting and camera conditions expected during operation. Missed defects and false rejects need separate assessment. A language assistant explaining those inspection results can't establish that the vision model detects scratches reliably, and a well-written investigation report can't repair weaknesses in the inspection model's validation. Each model needs testing for its own task. Once the evidence supports a cause, the engineer can document the scope of that finding, including which products and conditions the tests actually covered, and where uncertainty remains unresolved. The conclusion shouldn't outrun the test. That same discipline applies when the question turns to what equipment failure, because estimating a future maintenance problem requires its own evidence, rather than confidence borrowed from a convincing explanation of past events, prediction brings another validation task. Consider an earlier point in machine force operating history before the stop, when condition monitoring flags a change in the vibration pattern around a bearing assembly. Maintenance now has a warning to assess predictive maintenance needs a defined condition to estimate from suitable operating data, supported by records that show how the equipment behaved as that condition developed. Something looks different is a starting observation, not a failure forecast. The maintenance team needs to know what the model actually estimates. It might detect a departure from expected behavior, estimate the likelihood of a defined fault or estimate remaining operating life under stated assumptions about future use. Those outputs support different responses. An unusual vibration pattern might justify a condition check, while a validated estimate of a particular failure risk could support a more focused maintenance review for that component. Rare failures make this difficult to validate. A plant might collect extensive normal operation, but record very few examples of the failure. The team wants to anticipate, leaving limited evidence about how reliably the model recognizes it. The failure records themselves can also contain uncertain labels. A bearing replaced during planned work doesn't necessarily represent a bearing approaching failure, especially when the technician's report contains no examination findings that confirm its condition. Training on those replacements can teach the wrong relationship. The model could learn when the maintenance team usually replaces a component, rather than which measured conditions reliably indicate that the component is developing a particular fault. Operating conditions can change the interpretation as well. A heavier product, a different speed range, or a changed mounting arrangement can affect the measurements, so the model's performance needs checking against the conditions where you intend to use it. The assistant can make those boundaries easier to inspect when the warning arrives, it can explain the predictive models stated output and retrieve the operating context needed for a maintenance engineer to judge whether that output applies. It should preserve the predictions actual meaning. If the model reports a departure from its normal reference pattern, the assistant can describe that departure without inventing a failure date or naming a damaged component that the model hasn't identified. Preparing the review is a useful language task. The assistant can assemble the alert, the relevant condition, trend, and the models documented limits into a short account with links to the evidence the engineer needs to examine. Then the practical maintenance questions begin. Suppose the engineer considers a bearing inspection justified, but the replacement assembly requires a supplier specific part that stores cannot yet confirm is available. For this installed machine configuration. A stock quantity alone won't settle that question. The maintenance coordinator needs confirmation that the available item matches the required specification, because a similar part on the shelf might not support the intended repair on machine for crew availability needs the same practical attention. Someone may be free during the proposed window without holding the required qualification or experience, and the work may depend on a specialist whose attendance needs arranging. The maintenance window must fit the actual work. That includes the required preparation and return to service checks, not just the estimated time to remove and replace a component. Once the equipment is ready for intervention, those details can change the response to the warning. A team might arrange an assessment first while confirming parts and specialist support, rather than treating the alert as an instruction to begin a replacement immediately. Any continued operation needs an engineering basis. The assistant can gather the relevant evidence for that decision, but a predictive estimate doesn't establish that running until a convenient window is acceptable for this equipment and condition, I'd want the review to produce a specific proposed response. That means identifying the work under consideration. The resource is still unconfirmed, and the person responsible for resolving those gaps before maintenance commits to an intervention. A useful warning creates time to prepare its operational worth. Depends on whether that time lets the team confirm the need for work, secure suitable resources and choose an intervention that the plant can actually carry out. Production still has to absorb that intervention. Once maintenance proposes taking machine four out of service, the planner needs to establish what can happen to its assigned work given the production requirements surrounding that proposed maintenance window. The planners first question is whether the displaced operation can fit elsewhere. Suppose machine four becomes unavailable for the rest of the shift, and two other machines appear to have room in their schedules for the affected work. An empty slot isn't yet a usable slot. One alternative might support the required operation, but need a lengthy setup, while the other might already have suitable tooling installed but remain committed to an earlier customer order. Start with the approved alternative routes for this product. The scheduling system needs to know which operation can move, which equipment can perform it, and whether moving. It changes the sequence of work that must follow. Then account for the setup between jobs. If the alternative machine currently runs a different product family, the required cleaning or fixture change consumes capacity before the displaced order can produce its first acceptable part. Setup time can depend on what runs immediately beforehand. Moving the operation into one gap might require a different preparation than moving it into another, so the schedule needs more than a fixed duration beside each job. Tooling introduces another limit on the available choices. A machine might be free, while its required fixture remains in use elsewhere, meaning the operation can't start until both the equipment and that fixture become available together. Material needs are timed availability check as well. An order may have sufficient material allocated overall, yet the required lot might still await release or remain at another operation when the proposed replacement slot begins. Worker availability can close an otherwise plausible gap if the qualified operator is assigned elsewhere during that period. The scheduling method must account for that commitment rather than placing the same person beside two machines. This is work for advanced planning and scheduling or apps. Such systems evaluate production schedules using modeled capacity and constraints. Although the particular constraints they support depend on the software and how your team configures it, I'd specify the required scheduling behavior before choosing the implementation. For this outage, the method needs to respect the maintenance block on machine for and place displaced work only where all modeled requirements can be satisfied. The solver also needs a stated objective. Protecting delivery commitments might be the priority with setup effort considered afterward, but the planner needs an explicit rule for handling orders that compete for the same limited capacity. Some requirements can carry penalties rather than absolute prohibitions. A delivery target might permit an assessed delay, while an equipment qualification remains mandatory, and those distinctions must appear in the scheduling model rather than only in conversation. Your ERP supplies much of the business requirement. Its orders, material records and planning transactions describe what production needs to deliver, but their timing and detail must be suitable for the scheduling task you're asking the solver to perform. The Mez contributes the execution position from which scheduling starts. It can provide recorded progress and work already underway, allowing the scheduling application to distinguish remaining operations from operations that no longer need a place in the revised plan. Reconcile that starting position before requesting a replacement schedule if an operation has already started, but the scheduling input treats it as untouched, the solver can produce a mathematically consistent answer to an inaccurate account of production. The copilot can help prepare the scheduling request. It might ask the planner which orders may move, collect the confirmed outage interval, and identify missing inputs that prevent the scheduling service from assessing the available alternatives. Feasibility belongs with the validated scheduling tool. The assistant should submit those inputs through the defined connection and explain the returned result without adding an apparently convenient production slot that the solver never evaluated against its constraints. A useful explanation identifies why the proposed slot works. It could state that the operation moves after another job because the fixture becomes available. then and explain which delivery commitment changes under that particular return schedule. An infeasible result needs an equally specific explanation. The assistant can surface the reported conflict, such as qualified labor being unavailable before the required completion time, so the planner knows which assumption or commitment needs reconsideration. A feasible schedule still leaves a practical question. The planner needs to examine how the proposed arrangement might behave when actual production differs from the durations and operating assumptions used to construct that schedule. A proposed schedule gives us something concrete to test. Simulation lets us estimate how production might unfold under that arrangement, using stated assumptions about durations, variation and the operating rules that govern the modeled process for machine four, we'd compare several responses to the maintenance concern. Those might include delaying the intervention until a later window, carrying out maintenance immediately, or routing the affected production through an approved alternative while the machine remains unavailable. These are scenarios, not instructions to proceed. Any delayed maintenance scenario belongs in the comparison only if the responsible engineering assessment permits continued operation under the stated conditions. Simulation can't establish whether that continued operation is safe. Each scenario needs a consistent starting point. Use the same recorded production position and customer requirements, then change the response being tested so the planner can understand which consequences follow from that particular choice. Consider the immediate maintenance option first. Taking machine four out now might interrupt the current sequence, but completing the work earlier could leave more room for later orders. If the intervention finishes within the assumed duration, a delayed intervention shifts that exposure elsewhere. The current order might finish sooner, while maintenance occupies a later period when more work depends on the machine, creating a different pattern of waiting and completion times. Routing work elsewhere changes more than its location. The simulation can estimate when parts reach the next operation under that route, including whether a different release pattern leaves downstream equipment waiting or builds a queue before it. Those consequences depend on the model's operating rules. If the plant normally releases a transfer batch before the entire order finishes. A simulation that waits for the full order will describe a different flow. I'd have the planner inspect those rules with production specialists. Otherwise, a comparison can look precise while assuming that material moves, work starts, or completed parts become available in ways the plant doesn't actually permit. Repair duration deserves more than one assumed value. Maintenance might provide a likely duration and a wider plausible range, allowing the comparison to examine what happens if the intervention takes longer than the central estimate. Yield introduces uncertainty into the production response. Two if a scenario assumes every completed part passes inspection, it can underestimate the time needed to supply the required quantity when the process normally produces some rejected parts, change over duration can move the result in another direction. An alternative route that looks attractive with a short preparation might lose that advantage when the simulation uses a longer, still plausible changeover for that equipment and product combination. Keep those uncertain inputs visible in the comparison. Rather than announcing one completion time, the result can describe how completion changes across the tested assumptions, and identify which input most strongly affects the delivery outcome. Some uncertainties also move together if a difficult restart can lengthen the return to stable production and reduce initial yield modeling. Those effects as unrelated might understate the disruption in the immediate maintenance scenario where suitable data exists. Repeated runs can explore that variation. The spread of results still depends on the distributions and relationships applied, so repeated calculation doesn't turn an unsupported assumption into measured evidence about the factory. The co-pilot can explain the comparison in ordinary language. It could report that alternate routing protects the affected order across the tested repair durations, but delays another order when change over exceeds the assumed range in those runs. That explanation needs to preserve the trade off. Calling one option best hides the planners decision. If the scenarios exchange earlier delivery for one customer against additional delay for another without an agreed basis for that choice. The planner also needs to know what remains outside the model. If the simulation doesn't represent a relevant downstream operation or a restart effect, its predicted completion times shouldn't be treated as covering those omitted parts of production. Review the result against actual operating experience before release. A recommendation that surprises the planner may reveal a useful alternative, but it may also expose a missing rule that requires correction and another comparison before anyone changes the schedule. That leaves a concrete boundary between testing and acting. Even when the comparison supports an option, someone still needs authority to approve the production change. Which brings us to the permissions and assistance should receive. Before connecting any action, I'd name the exact change it permits. Reading a record, drafting an entry, submitting that entry, approving the work, and executing a change. Each. Create a different responsibility for the application and the people using it. Those permissions need separate decisions. Even within one conversation. Someone asking the assistant to prepare a maintenance request hasn't necessarily agreed to send it, and sending a request doesn't authorize the maintenance activity described inside it. 01:10:08 Speaker: Start with a maintenance request draft for our stopped machine. The assistant can populate the required fields from the evidence already collected, while leaving any field unresolved when its value requires a judgment. The responsible person hasn't supplied the draft should carry its supporting evidence into review. That includes references for the reported condition and the relevant records, so the reviewer can check the entry without relying on the assistant's wording as the only account available. Required fields need more care than simply preventing empty boxes if the request requires a priority. For example, the assistant should use an established classification rule or ask the authorised person rather than inventing urgency from the operator's tone. Drafting is a relatively low risk starting point. Keep that draft outside the active work queue until someone checks it, because an unfinished entry can still trigger confusion if another crew mistakes it for an accepted assignment, submission creates a new operational consequence. Once the request enters the maintenance system, it may notify a coordinator or become visible for work allocation. So the user needs to understand that effect before confirming submission. Make that confirmation specific to the pending action. Submit the This maintenance request for review tells the person what will happen, while continue leaves them guessing whether they're saving a draft or creating a record. Other people will act upon approval. Answers a separate question about the proposed work. The maintenance owner may accept the request for assessment without authorizing a repair, and the workflow needs to preserve that distinction wherever the assistant reports progress. Back to the requester. An approval also needs a defined scope agreement to investigate. A fault shouldn't become permission to replace components, expand the job, or change the production arrangement when those decisions require further review. Under the plant's normal process, execution needs its own boundary even after approval. Creating an approved record and carrying out the associated physical work are different events, so the assistant shouldn't describe the intervention as completed merely because the request passed its approval stage for production changes. Retain the established release workflow. If a planner accepts a proposed scheduling option, the assistant can prepare the change for that workflow. But acceptance in conversation shouldn't silently update assignments that production teams already rely upon. The same care applies when an action changes during discussion. If the user revises the requested work after reviewing a draft, any earlier confirmation should cover only the previous version, not a different instruction assembled later in the exchange. I'd give each permitted action a short written contract. State what it changes, which conditions must hold, who can authorize it, and where. The application must stop when the request falls outside that agreed boundary. Keep safety controls independent of the assistance generated response. Emergency stops, protective functions and machine interlocks must retain their established control paths, with no assistant authority to bypass them because a generated explanation suggests the interruption is unnecessary. An instruction from a user doesn't remove that restriction. If someone asks the assistant to bypass an interlock to finish an urgent order, the application must leave that action unavailable and direct the issue through the plant's authorized safety process. Review these boundaries with the people receiving the resulting work. A maintenance coordinator may interpret a submitted request differently from a supervisor, so agree on what each status permits before the assistance starts creating entries during a busy shift. That's how you add automation without quietly expanding its authority. You can permit a useful administrative step while keeping approval and execution with the established owners. Instead of granting every action that happens to follow naturally in conversation, these agreements now need controls that hold outside the conversation. The next design question is how the application enforces each boundary when the user requests something beyond it, rather than depending on the model to remember where it should stop. The maintenance service should reject an unauthorized request on its own, even if the model requests the wrong function or supplies an incorrect explanation, the receiving application must check whether this person can perform this action on this record. Begin with the identity behind the conversation. Microsoft Entra ID can provide sign in and application identities where appropriate, but your application still needs rules connecting those identities to the permissions defined by the maintenance or production system. Signing in doesn't grant access to every plant. A technician assigned to one site might have permission to retrieve its maintenance records without being permitted to read equivalent records from another site within the same organization. Check that site boundary in the receiving service. The service should derive the permitted sites from trusted identity and policy information, rather than accepting a plant name supplied by the model as evidence that access is allowed. Separate the user's identity from the application's identity. A connector might authenticate through a service identity, but that identity shouldn't automatically give every person using the assistant everything the connector can reach in the source system. That distinction affects how you design each connection. Where supported, delegated access can preserve the user's permissions. Where an application identity is necessary, the application needs its own enforced checks before returning information or requesting a change. Give service identities only the access their functions require a component. Retrieving maintenance status shouldn't also hold permission to modify equipment records, simply because granting a broad account proved easier during the initial connection test. Search needs the same discipline before passages reach the model. If a user can't access a restricted engineering document in its source system, the assistant shouldn't reveal its contents through an answer assembled from the search index. Copying content doesn't automatically copy its access controls. Your indexing process must retain the permission information needed for filtering, and the retrieval service must enforce that filtering using the authenticated users permitted access before supplying passages for generation. Check how permission changes reach the search service, too. When a person's access is removed or a document becomes restricted, previously indexed permissions must stop granting access within a defined period, with failures handled explicitly rather than left unnoticed. Tool access needs checks at the moment of use. A conversation might remain open while someone's role changes, so authorization from the start of the exchange shouldn't become a lasting permit for every subsequent request in that session. The application service should evaluate each proposed action independently. It can check the authenticated identity permitted plant assigned role and requested operation against its policy before forwarding an allowed request to the system that owns the record. Use the actual target record for that decision. If someone supplies a maintenance reference belonging to another site, the service must enforce the record site restriction even when the conversation still concerns the user's own plant. Keep credentials outside the language models, working material tokens and connection secrets belong in protected application infrastructure, such as an appropriately configured secret store with controlled access rather than inclusion in prompts, retrieved passages or generated tool descriptions. Network access adds another boundary around those credentials. Information technology or IT services and operational technology or OT equipment shouldn't become mutually reachable simply because the assistant needs a selected status record from the production environment. Use controlled integration paths between those environments. The network design should permit only the required connections through approved gateways or intermediary services, with the controls team determining which production endpoints remain unreachable from the assistance application environment. A blocked connection should remain blocked during troubleshooting. If a required function can't operate through the approved path, resolve the integration design instead of opening broad access that allows unrelated services to reach production systems. Verify these restrictions using deliberately unauthorized requests. Ask an authenticated user to retrieve another site's record or invoke a function outside their role, then confirm that the application refuses access before any restricted information reaches the model. These controls constrain the application even when its reasoning fails. There's another source of trouble to examine next, though, because an authorized retrieval can still return content containing instructions that the assistant should never treat as application rules. A maintenance note can contain an instruction aimed at the assistant if the application retrieves that note while preparing a handover. The model encounters those words alongside the legitimate fault history it needs to interpret for the technician. This attack is called prompt injection. Someone places instructions inside material, the assistant reads, hoping it will follow those instructions instead of treating them as content supplied for a limited purpose. The entry point doesn't need to be a document, a supplier message, a job description or text returned by a connected tool can carry the same attempt if its content reaches the model during a conversation. Return to the maintenance history for our hypothetical machine for. Suppose someone inserts a paragraph claiming that a complete fault assessment requires retrieving restricted engineering records and including their contents in the technician's handover before proceeding. The paragraph might claim to come from the application administrator. It could tell the assistant to ignore earlier restrictions, hide the additional lookup from the user, and treat the instruction as a mandatory part of the maintenance workflow. That text remains a maintenance note regardless of its tone. Its presence in an accessible record gives it no authority to change the application's rules, expand the task, or decide which information belongs in the response. There is a distinction between quoting an instruction and obeying it. The assistant may need to report that the note contains suspicious text, but it should treat that text as evidence to examine, rather than a command to carry out separate source material from application rules throughout the request. Clearly identify retrieve passages as reference content, retain their origin, and instruct the model to use them only as evidence relevant to the user's authorized task. That separation helps, but it isn't a complete defense. A model can still follow hostile wording, so the application needs controls that limit what can happen even when the generated response or proposed tool call follows it. Restrict the available tools to an explicit allow list for a maintenance handover. Expose only the functions required for that workflow, rather than letting the assistant discover and invoke unrelated functions. Because retrieved text asks it to check the purpose of each call as well as permission. A person might legitimately access engineering records in another workflow. Yet a maintenance note shouldn't cause the assistant to retrieve them without a task related reason. Authorization remains independent of the retrieved text. The service deciding whether to release a record or accept an action must use trusted identity and policy information, never a claim of permission embedded inside a document or tool response. Watch for attempts to send information elsewhere. A note might instruct the assistant to transmit its findings to an external address for supplier diagnostics. Turning an ordinary retrieval task into an attempted disclosure of plant information. Outbound connections need a defined destination policy? Allow only the services required by the application, and prevent arbitrary web requests or message destinations from becoming available, simply because the model produces a plausible explanation for using them. Sensitive data filtering adds another check before anything leaves. Inspect generated responses and outbound tool payloads for protected information using the organization's data rules, while recognizing that filters can miss content or block legitimate technical material. Test those checks against realistic engineering content. A filter that catches passwords but ignores restricted drawings, process recipes, or copied maintenance records won't address every disclosure that this particular application could attempt during an attacked conversation. Adversarial testing means deliberately trying to break these boundaries, place attack instructions in controlled copies of maintenance notes, supplier documents and tool results. Then observe whether the application follows them, rejects them, or exposes information through its reply. Include attacks that sound helpful rather than openly hostile. Claims about urgent safety updates or required diagnostic steps can be persuasive, so test whether those claims change tool selection or cause the assistant to conceal an action. Inspect the actual calls, not just the final answer. A response might look harmless after an unauthorized lookup has already occurred, which means testing needs access to the application's retrieval and tool activity, as well as its wording. The useful outcome is a contained failure. Even if hostile text influences the model. The surrounding application should prevent unrelated retrieval, blocked actions, and data transfer to destinations that the maintenance workflow never needed. Ordinary missing evidence creates a less obvious problem than a hostile instruction. A technician can ask a reasonable question, receive an incomplete answer, and still need to decide what that answer permits them to conclude about. The job uncertainty needs to help them choose their next step. Saying this answer may be inaccurate leaves the person with the entire checking task without identifying what remains unknown or which part of the response they can actually use. I'd make the evidence state explicit in the reply. The application should distinguish an answer supported within a stated scope from one with partial evidence, conflicting evidence, or a source it couldn't reach when the question arrived. A supported answer still has a boundary. If the technician asks which procedure covers a confirmed alarm and configuration, the response can identify that procedure without suggesting that it establishes the machine's physical condition or the required repair. Partial evidence calls for a more selective response. Suppose the assistant can retrieve the approved inspection requirements, but can't establish whether the previous crew completed that inspection. It can explain the requirements while leaving completion unresolved. That distinction should shape the wording immediately. The procedure requires this inspection. I don't have a completion record gives the technician usable information without implying either that the inspection happened or that nobody performed it. Conflicting evidence needs a different kind of restraint. The assistant can state which question the disagreement, prevents it from answering, so the person understands whether the conflict affects a document reference and equipment detail, or the proposed next step. An unavailable source should be named, not hidden behind general doubt. The maintenance history service isn't available, so I can't check the previous inspection. Explains the gap and lets the technician decide whether another approved route can supply that evidence. Keep these states close to the claims they affect. A warning at the end of a long response forces the listener to reconsider everything, even when the missing evidence concerns only one specific part of the answer. The response also needs to distinguish different kinds of statements. A retrieved fact reports what a source records. A calculated result, follows a defined operation, and a proposed explanation describes something the team still needs to investigate. Those distinctions can sound natural rather than bureaucratic. The assistant might say the job record lists an inspection yesterday. The interval calculation places it before this stop. Whether the condition changed afterward. Still needs checking. Each statement carries a different evidential burden. The record supports the listed inspection. The calculation supports the timing relationship, and neither alone establishes what developed inside the machine between the inspection and the stop. Clarification should target the gap that changes the answer. If the question concerns which branch of a procedure applies. The assistant might need the confirmed operating mode rather than another broad description of the breakdown. Explain why that detail matters before asking for it. The procedure separates automatic operation from setup mode, which mode applied when the alarm appeared, gives the technician a reason for the question and a clear way to respond. Missing identity needs the same focused treatment. Ask for the particular identifier needed to distinguish the candidate assets, rather than asking the user to repeat the entire report or supply details the application already holds. Sometimes nobody can supply the missing detail during the conversation. The assistant should then stop at the supported portion of the answer and offer the designated escalation route without repeatedly rephrasing the question until the user guesses Escalation should preserve the work already completed. Send the responsible person the original question, the evidence collected, and the precise unresolved point so they can address the gap instead of starting another round of information gathering. A model confidence score won't resolve that gap. A percentage attached to generated wording doesn't establish that an asset matches a source, supplies or an unobserved condition exists, so it shouldn't become proof for the person making the decision. I'd rather hear exactly what remains unconfirmed. I can identify the applicable procedure, but I can't verify. The equipment's present mode tells the technician where assistance stops and which fact still needs an authorized check. That response policy needs to hold under pressure. We now need to examine whether the assistant actually maintains those distinctions across the plant. Questions it will encounter, including questions that arrive with mistaken assumptions already built into them. Testing starts with actual questions from the work you're assigning. I'd build an expert reviewed set of cases from approved procedures and representative maintenance scenarios, with an expected outcome established before the assistant attempts an answer. A case needs more than a question and model answer. Record the equipment configuration, the available source material, the user's permitted access, and the conditions that determine whether the assistant should answer, clarify or stop. Have the relevant specialists define what counts as correct. A maintenance engineer can identify which procedure applies, while the document owner confirms its release status, and the application owner checks the permitted behavior for that workflow. The expected response doesn't need one exact wording, it needs required facts, acceptable limits on the explanation, and specific claims that would make the answer unacceptable, even if the rest sounds sensible to a technician. Choose a straightforward case first to establish the basic behavior. Give the assistant a confirmed asset and a question with an explicit answer in the approved material. Then check whether it returns the right information without adding unsupported advice. Then change one condition while keeping the question unchanged. Replace the applicable manual with an obsolete revision in the test collection and establish whether the application excludes it or correctly reports that suitable evidence isn't available. Use cases where the wrong answer looks attractive. An older manual might contain a passage that matches the question almost word for word. While the current procedure describes the same subject using different technical language and additional conditions. Similar equipment names deserve a paired test. Ask about machine four at one site while providing another asset with the same local name, then inspect whether the answer stays within the confirmed equipment scope throughout the exchange. Missing records need a defined expected result to remove a required configuration detail and check whether the assistant requests that detail, rather than selecting whichever procedure happens to rank first among the remaining search results. Some questions should contain a mistaken assumption. A technician might ask which replacement part to order for a confirmed drive. Failure. When the supplied evidence establishes only an alarm, and the expected response should challenge that unsupported diagnosis. Score the claims, not just the overall impression for each technical statement, check whether the available evidence supports its meaning and conditions, then record omissions separately from statements that contradict the source or exceed its scope. 01:27:50 Speaker: Track abstention in both directions. Refusing an unsupported request is correct, but refusing a supported lookup creates unnecessary work, so the test set needs cases where answering is expected, as well as cases where restraint is required. Permission testing should use known identities, and known records, run equivalent questions under different access rights and inspect the retrieved material. Because an answer that conceals restricted content doesn't excuse an unauthorized retrieval earlier in the workflow, don't collapse every result into one average score. An assistant can perform well on routine lookups while repeatedly selecting the wrong equipment revision, and those failures need their own category rather than disappearing among successful answers. The conversation itself needs testing beyond the first question. Let a user correct an asset identifier, ask a follow up using that machine or contradict an earlier detail. Then verify which confirmed context the application actually uses. Test the complete workflow, not an isolated model response. For a maintenance request task. Follow the case through retrieval tool selection, input validation, draft preparation and the required approval boundary checking the recorded outcome at each stage, a friendly completion message proves very little. Inspect whether the intended record exists, whether its fields match the reviewed request and whether the application attempted any change before the required approval was actually supplied in the test. Failure handling belongs in that same examination. Interrupt a tool response or return an invalid result, then check whether the workflow stops cleanly and accurately reports what completed, what failed, and what remains unresolved. Keep some reviewed cases outside the development team's tuning set. Those cases help reveal whether changes improved performance on unfamiliar tasks, rather than merely teaching the application to produce acceptable answers to examples its builders repeatedly inspected. Passing these tests establishes a basis for further assessment. The next question is whether the person reviewing an answer during a breakdown can detect its remaining limits before approving anything that affects the work. Who actually has time to review the answer during a breakdown? Assigning a technician that responsibility only works if the shift arrangement gives them enough time. Access to evidence and authority to challenge. The proposed result without being pressured to accept it. I'd name the decision owner for each workflow and action class. That person needs to understand what they're accepting, what remains outside their responsibility, and who takes over when the question exceeds their competence or assigned authority. A generic approval queue can obscure that ownership. If several people receive the same item, each might assume someone else checked the technical detail, leaving an accepted entry with no clear account of who judged its suitability. Make the assignment explicit. When the review reaches someone, the reviewer should know whether they're checking the accuracy of a fault description, authorizing resource allocation, or accepting a production consequence. Because those reviews demand different knowledge and evidence, the review effort should match the consequence of an error. Correcting the wording of an internal note generally demands less scrutiny than accepting an entry that other people will use to allocate work or commit scarce maintenance resources. Reversibility changes that burden as well. An incorrect description can often be corrected before anyone uses it. While an accepted commitment may already have displaced another job by the time someone notices the underlying mistake. Available evidence determines whether a review is practical at all. Asking someone to approve a proposed priority without the relevant operating consequences forces them either to investigate separately or accept a judgment they can't properly assess from the material supplied. The review package should make that gap apparent. Put the proposed decision beside the specific evidence needed to judge it and identify any assumption requiring the reviewer's agreement, rather than burying that assumption inside a polished explanation. Reviewers also need an alternative to approving or rejecting everything. Let them return a draft for clarification, correct a specific field, or refer one unresolved judgment to another owner without losing the portions they've already checked. Downtime pressure changes how people use those options. Imagine the maintenance coordinator reviewing a prepared request while the supervisor waits for an answer, and another technician calls about a fault elsewhere. A quick acceptance becomes very tempting. That's a setting for automation bias. People can give a systems suggestion more weight than the evidence warrants, especially when accepting the suggestion removes an immediate interruption and challenging. It creates another task during an already crowded shift. A citation doesn't remove that pressure if checking it requires opening a long manual and reconstructing the relevant conditions, the reviewer may accept the response without performing the examination. The workflow assumes they completed. Warnings can meet the same fate under repeated use. A standard caution attached to every response can become background text, while the particular unresolved assumption that matters for today's decision receives no separate attention from the reviewer. Give the reviewer a concrete checking step instead for a proposed maintenance priority. That might mean comparing the stated production impact with the supervisor's confirmed account, then explicitly accepting or correcting the impact before the priority can pass. Review. Avoid turning every review into a lengthy checklist, though. Observe how qualified people actually examine the proposed result, then retain the checks that helped them catch consequential errors rather than collecting confirmations that nobody can meaningfully verify. I'd measure the effort required to complete that examination. If reviewing a draft takes longer than preparing the entry directly. Simplify the assignment or reduce the assistant's scope instead of treating slow approval as a failure of user adoption, corrections should enter a controlled improvement process. A technician can report that a response emitted a condition, while the application owner and relevant specialist determine whether the fault arose in the source retrieval instructions or connected workflow. Don't treat a correction as automatic training data. The feedback may concern one configuration or contain another mistaken assumption, so someone needs to verify its meaning before deciding which control change, if any, the application requires. Keep the person reporting the problem informed about its disposition. Knowing whether a correction was accepted, rejected, or held for investigation helps technicians distinguish an unresolved defect from a problem that someone has actually addressed in the working application. Human review consumes real shift capacity. Once responsibility and checking effort are explicit, we can ask whether that review arrangement remains workable when the assistant becomes part of everyday support across successive crews. The support arrangement needs to survive losing the assistant altogether. If a network fault cuts the plant off from the cloud during a shift, maintenance still needs a working root to approved procedures, and the people who support the equipment plan that fall back before the service enters routine use where approved local copies form part of the continuity plan. Document control needs to maintain them through a release process that works independently of the assistance cloud connection. Practice the switch with the crews who will use it during a controlled outage exercise. Confirm that people can reach the local procedures and normal support contacts without needing the unavailable assistant to explain where they've been stored. Service ownership also needs to extend beyond office hours. Assign a support route that tells each shift who handles an unavailable assistant, who can disable a faulty function and where unresolved incidents pass. When the responsible team changes, a slow service creates a different operational problem. An answer arriving after the technician has called someone else can interrupt work already underway, especially when the application gives no indication that an earlier request is still processing. Define response targets around the work being supported. Measure the whole interval from asking a question to receiving a usable result, with enough detail to distinguish model processing from time spent waiting for search or a connected service. Averages alone won't describe that experience. Adequately track the slower responses and abandoned requests to. Because a service that usually responds promptly can still become unreliable precisely when several crews needed during the same disruption. Keep operating measures separate enough to diagnose problems response delay usage, cost index freshness, and tool availability describe different conditions so an overall service status shouldn't conceal which part currently prevents the assistant from completing its assigned task. Cost needs an operating limit, not just a monthly invoice. Repeated searches, long conversations, and automatic retries can increase consumption, so the service owner needs alerts and a defined response when usage moves beyond its expected range. Index freshness deserves a direct operational check. Monitor whether scheduled updates actually complete and whether the searchable collection reaches the expected release date, rather than counting a running update process as evidence that its contents are current. Tool availability should reflect a working dependency, not an open connection. A service can accept network traffic while failing to return usable results, so monitoring needs controlled checks that exercise the functions this assistant actually depends upon. Changes to the application need an identifiable release record, the model version, prompt version retrieval settings, and connector versions together. Because a response can change when any one of those components changes. Even without a visible application update, that release record makes investigation much less speculative. When a crew reports different behavior after an update, the support team can establish which configuration handled the request instead of trying to reconstruct it from whatever settings remain deployed today, run regression tests before releasing the changed configuration. Reuse the approved workflow cases to check whether previously acceptable behavior remains acceptable, including interactions affected by a connector update or a different passage selection setting in retrieval. Keep a known working configuration available for rollback. The deployment process should specify who can restore it and how they'll verify restoration with attention to connection or data changes that an application rollback can't automatically reverse. ordered records need enough detail to reconstruct an incident. Capture the configuration used relevant source references. Tool activity and workflow outcome under defined logging rules, without assuming that every conversation must be retained indefinitely in full. Those records need access limits and retention rules of their own. Maintenance discussions can contain sensitive operating information, so restrict who can inspect the logs and define when records expire or require preservation for an investigation. Give incident response a concrete operating procedure if the assistant starts returning withdrawn material or misreporting completed actions, the service owner needs a route to contain the affected function, preserve evidence, notify users, and verify recovery before restoring access. Reliability brings continuing work and expense into the project. Once that burden is visible, you can judge whether the assistance delivered during an ordinary shift justifies maintaining it, rather than judging the application only by the quality of its demonstration, a pilot should answer a funding question, not just a technical question. Does this particular assistance improve a defined task enough to justify its full cost compared with how the crew completes that same task. Without it, I'd start with approved maintenance document retrieval for one equipment family, choose a family with enough recurring questions to assess the work and set a boundary around finding applicable source material rather than diagnosing faults or recommending repairs, write down what completing that task means before the trial begins. For example, the technician must locate the applicable released procedure, confirm its equipment scope and identify the requested information, or recognize that the available material doesn't answer the question. Measure the existing method first, including searches that don't succeed. Ask participating technicians to complete representative lookups through their usual channels, recording the time spent and whether they finish with the correct source or need help from someone else. Keep the comparison fair to the existing method. If technicians normally use bookmarks, local indexes, or a colleague's document reference, let them use those aids rather than forcing an artificial folder search that makes the assistant look better. Source selection needs its own measure alongside search time, a fast lookup that returns the wrong revision shouldn't count as an improvement. So have the responsible specialist judge applicability using criteria agreed before anyone reviews the pilot results. Task completion tells you something different from finding a document record. Whether the technician actually obtains the information needed for the lookup. Because reaching a relevant manual doesn't necessarily resolve the question that started the search. Also examine what happens when the lookup can't be completed. Assess whether an escalation reaches the appropriate person with a clear, unresolved question, and whether that person can respond without asking the technician to repeat the entire search. Choose participants across the shifts that will use the service. Include people who know the document collection well and people who don't. Then examine their results separately so improvements for one group don't conceal additional work for another. Avoid testing the same question twice with the same person. Someone who has just found a procedure will remember its location. So use comparable task sets and vary their order to reduce the effect of practice on the comparison. Agree on the decision rules before collecting results. Decide what improvement would justify continuing which errors require correction before another trial, and which outcomes mean the proposed task isn't a sensible investment in its current form. The cost record should begin before the first successful answer count document cleanup and the engineering time needed to establish applicability alongside the integration work required to make this limited collection available to the application include the work that continues throughout the trial as well. Review, effort, licensing and support belong in the assessment, including time spent resolving failed lookups or helping technicians recover from a confusing exchange during their normal duties. Keep setup costs separate from recurring costs that lets you judge whether an expensive preparation effort supports a repeatable task, or whether each equipment family would require so much additional work that expansion changes the economics. Roll out the pilot in stages with a decision between each stage. Begin with read only trials where participants compare results against the approved source, then allow routine retrieval only after the agreed evidence supports that step for the selected equipment. Family bounded workflows can follow through a separate trial if you later add a maintenance request draft, assess that addition against its own task and cost, rather than treating successful document retrieval as permission to expand the assistant's job stop criteria should be specific enough for a shift lead to apply, returning inapplicable instructions as current guidance, exposing restricted material or repeatedly claiming successful retrieval when the source is unavailable should trigger a pause in the affected function. Specify who can call that pause without waiting for a project meeting. Resume only after the responsible owner verifies the correction against the failed cases, and confirms that the pilot can continue within its agreed scope, the outcome might be a narrower service than you first expected. You could retain procedure finding while dropping conversational explanations. If checking those explanations consumes the time the search function saves, that's a useful pilot result, not an incomplete rollout. Then test the supported scope against a complete maintenance event we can return to machine for and follow where this limited assistance contributes, where another system takes over, and where an authorized person must make the decision. This time, follow the handoffs through to the recorded outcome. We're back at our hypothetical plant where machine four has stopped mid order, and the technician uses the assistant to prepare for an assessment rather than request a repair verdict, the technician confirms the installed configuration against the asset record. The assistant retrieves the current fault event and applicable approved procedure, then brings forward the recent maintenance history associated with that confirmed equipment identity for the technician to examine one recent entry records work near the affected assembly. The technician reads the underlying note and decides it's relevant enough to mention in the request, but not enough to establish that today's fault has the same cause. The assistant prepares the maintenance request with that distinction preserved, it describes the recorded stop, attaches the procedure reference, and includes the earlier work as background while leaving the cause and expected repair duration unresolved. For the assessment, the technician corrects one detail before accepting the draft. The operator reported the alarm after the stop, but the initial wording suggests the operator witnessed it appear, so the technician changes the account to reflect what was actually reported. That correction becomes part of the submitted request. The maintenance coordinator now receives a usable record with an identified asset and a clear assessment requirement, rather than a request that already commits the crew to a particular replacement. The technician then checks the conditions at the equipment. Following the plant's approved requirements, they establish whether the proposed assessment can proceed and authorize only the next step supported by those checks, without treating the assistance request as permission to intervene. For this scenario, the assessment requires a longer production hold. The technician records that conclusion and tells the coordinator that a repair duration remains unconfirmed, giving production a definite availability restriction without supplying an unsupported estimate of when the machine will return. That creates a specific decision for the planner. They need to decide what to do with the assigned operation during the confirmed hold, while keeping the unresolved maintenance duration visible in any production option they consider. The assistant prepares the inputs for the approved scheduling service. The planner checks the affected order and outage assumption, then requests alternatives through the validated tool. Rather than asking the conversation itself to invent a replacement production sequence. The returned options expose a delivery conflict in this hypothetical result, moving the operation to an approved alternative protects the affected order, but shifts another completion, so the planner must decide whether that trade off can be accepted. The assistant explains that consequence without selecting it for the planner. The planner consults the responsible production owner, checks the commitments affected by the move, and approves the chosen option only after the required review has addressed that displaced work release happens through the established production systems. Once the revised schedule is accepted there, the assistant can report its recorded status while maintenance continues under its own authorization, with no implication that rescheduling has resolved the equipment fault. Notice where the conversation stops determining the sequence. The technician owns the assessment decision. The scheduling tool evaluates the modelled production options, and the planner owns release of the revised schedule after the relevant production consequences have been reviewed. The event record should let someone reconstruct those handoffs afterward. It links the evidence used to the people responsible and the actions actually taken, including the maintenance requests submitted and the production revision released through the existing workflow. It also retains what nobody could yet establish. The cause remains under investigation at that point, and the return to service time remains unresolved, even though the team has already taken legitimate action to assess the fault and protect production. I wouldn't judge this event by how confidently the assistant spoke. I'd examine whether it's prepared material helped each owner reach their decision with less reconstruction, and whether the final records accurately distinguish the work authorized from the work still pending. That gives the shift handover a concrete endpoint. The next crew receives a record showing what maintenance has authorized, where production has reassigned the work, and which unanswered questions still belong to named people, rather than disappearing into the conversation. Every proposed action needs evidence, permission, and an accountable owner. If any of those is missing, a persuasive answer can't authorize the production change. However urgent the order or convincing the explanation sounds to the team. Which shop floor task would you give an assistant first, and which action would you explicitly keep outside its authority when production is waiting?