Power Automate AI Builder - Simply Explained

Welcome to the M365 FM podcast, where we explore the people, ideas and technologies shaping the future of work. Why everyone is suddenly talking about AI builder. It's a Monday morning and someone on your team is typing invoice totals into a spreadsheet by hand. That exact job is what AI builder was built to delete, and the tool is sitting inside Power Automate right now. Unused. This is Microsoft's no code way to add real AI to a workflow. Reading receipts. Pulling numbers off invoices. Sorting feedback, translating text. Building a model costs nothing but every real run, Burns credits. And that's the part nobody warns you about. Stick with me and you'll finish knowing what it is, what it costs, and where its limits really are. What AI builder actually is. Let's start with the plain version. AI builder is a feature inside Microsoft's Power platform, and it lets you create and use AI models to handle parts of your business that used to need a person. You don't train a model from scratch and you don't write code. You pick something that already understands a task, or you teach it with your own examples, and then you drop it into a workflow. It comes in two flavors, and the split between them is the thing to hold on to. Prebuilt models are ready the moment you open them, because they already know how to read a receipt or detect what language a block of text is written in. Custom models are the ones you build, train, and publish yourself using your own documents and your own data. The reason AI builder exists is easy enough to understand. Most businesses have the same handful of chores, and most of those chores come down to reading something and deciding what to do next. That work is usually too varied for a strict rule and too repetitive to keep paying a person for AI builder goes after that middle ground, and it's built for the person who understands the process but isn't a developer, which is a much larger group than the people who write code. You find all of this in the maker portal under a section called the AI hub. And then AI models, if you've ever opened Power Apps or Power Automate and clicked around the left hand menu, you've walked right past it. That last part is worth sitting with for a second. The barrier that used to separate the people who automate from the people who build AI has mostly come down. If you can build a flow, you're already close to using a model. There's one honest prerequisite worth checking before you get too excited to use AI builder, you need a power platform environment that includes Dataverse. Think of Dataverse as the storage layer underneath power Apps and Power Automate, the place where structured data lives. If your organization already runs Power Platform seriously, you probably have it. If you're working from a personal account, check first because this is the step where beginners get stuck most often. The easiest way to hold the whole idea is this. Azure is the workshop where developers build AI from raw parts, and AI builder is the vending machine down the hall, already stocked and labeled where you press a button and get a result. It's the same ingredients with far less assembly. The pre-built models you can use today, pre-built models are the fastest win, so let's put real names on them. There's a receipt reader, an invoice reader, an identity document reader for passports and driver's licenses, and a business card reader. You hand one of them a file and it comes back with the fields you'd otherwise type out yourself. Then there's text recognition, which most people just call OCR. It pulls every word off an image or a PDF and hands you the raw text, which makes it the workhorse of the set. Once you have that text, you can search it, filter it, or feed it into a rule that looks for a specific phrase The text help us do smaller jobs that add up fast. Sentiment analysis tells you whether a piece of feedback reads positive or negative, and language detection tells you what language you're looking at. Key phrase extraction surfaces the terms that matter, which is how a team can catch every social media post mentioning their brand. Text translation moves a support ticket into a language your team actually reads. There are standard versions of category classification and entity extraction as well. Classification sorts text into buckets. You define, and entity extraction pulls the specific things out of it, like a product name, a date or a place. Picture a shared inbox that collects receipts from a sales team. Every time a receipt lands a flow, scans it, the model extracts the merchant and the total, and the numbers get written into a spreadsheet or a Dataverse table. Nobody retypes anything, and one person built that in an afternoon. That's the moment where this stops being theory and turns into a saved hour every week. Now the honest limit because this is where people get burned. OCR isn't perfect and it never will be. It mixes up a zero and the letter O a one and a lowercase l a four and the letter A a five sitting on a line can come back as a dollar sign. Microsoft's own guidance is blunt about it, saying AI builder can't fix those misreads, so you improve the quality of the source document instead. If your scans are faint or crooked, your results will be to. That ceiling matters because the built in models only cover the common stuff. Plenty of paperwork doesn't look like everyone else's, and that's where the next part of the story begins. Building a model from your own documents. This is where AI builder stops feeling like a gadget and starts feeling like a tool. A custom document processing model lets you point AI at the documents your business actually runs on, not the generic ones Microsoft imagined. The process is simpler than the name suggests. You collect a handful of real documents, upload them, and tag the pieces you want pulled out things like the invoice number, the due date, the total, and the vendor. Then you train. The model training runs on its own, so you're not adjusting dials or writing formulas. You're showing it what matters by marking it up. The tagging is the part that decides how well it works. If you name a field for what it holds, like supplier or due date, the model learns faster and extracts better because the name itself carries meaning. Naming your fields the way the document names them is one of the few habits a beginner can adopt that measurably improves results. The number that surprises beginners most is how little it takes to start. For clean documents that share the same layout, Microsoft says five samples can be enough. Fifteen to twenty is the recommendation if you want better results, especially when your scans are rough. Five isn't a typo, and it's the best reason to try this yourself. Instead of assuming you need a data team on standby. That single number is what turns custom AI from a budget line into a Friday afternoon experiment. If your documents arrive in several different layouts, you don't need a separate model for each one. You group them into what AI builder calls collections, and a single model learns to handle all of them. Document processing is the big one, but it isn't the only custom model. Object detection learns to find things in images, which is how a warehouse could count stock from a photo prediction looks at your history and flags. Which records are likely to go a certain way, like which supplier might slip out of compliance? There's a real limit to plan around. A general documents model with a lot of collections can take hours to train, and in rare cases, it runs long enough to time out. Microsoft's own documentation warns that a heavy model can take up to a day to finish. You train a model once and use it many times, so this is usually fine, but don't plan to build one an hour before a demo. Once a model is trained, it just sits there waiting. The work starts when you put it inside a flow. Using AI inside a power automate flow. A Power Automate flow is a series of steps that run on their own. Something triggers the flow, the steps do their work. And the last step hands off a result. AI builder slots in as one of those steps. That's what makes it practical. Because a model on its own can't do anything until something hands it a file. Here's a concrete one. An email arrives with an invoice attached, and that arrival is the trigger. The next action reads the attachment and pulls out the vendor name and the total. Those values become normal pieces of data you can drop into any later step, exactly like a name or a date. So the flow posts a summary to a team's channel and then it starts an approval. So a human signs off before any money moves. That human step matters more than it sounds. Microsoft's own guidance suggests adding a review step after a prompt or a model. So a person checks the output before it goes anywhere that counts. You don't have to trust the machine blindly and you shouldn't. The result is reusable, which is the real advantage of doing this inside a flow. The same extracted values can update a spreadsheet, send a team's message, create a row in Dataverse, or trigger a second flow. You build the extraction once and every downstream step just reads it. The part that bites people is a setting, not a skill. Every AI builder action in a cloud flow has an asynchronous pattern option and it's on by default. Leave it on, turn it off, and the results don't come back properly, which means your flow breaks in a way that's genuinely hard to diagnose. There's one more limit worth knowing before you build anything. Document processing expects one form per file. If someone sends you a single PDF with three invoices stapled inside it, the model sees one document, not three. You either split the file first, or you use the page range option to loop through it one page at a time. Those two details, the async setting and one form per file, are the difference between a flow that works and a flow that mystifies you at eleven at night once you know them. AI builder inside a flow feels less like magic and more like a very good assistant who needs clear instructions. Models aren't the whole story, though. There's another side of this that works in a completely different way, and it's the part most people reach for first. The prompt side where you just ask everything so far has been about models which are trained to do one specific job. Well. The rest of AI builder is prompts, and prompts are about asking for what you want in plain language. A prompt is a saved instruction. You write yourself. You tell the AI what to do, and inside a flow you run it with an action called run a prompt. If you've seen older tutorials, this action used to be called Create Text with GPT using a prompt, and Microsoft renamed it in May of twenty twenty five. It's the same idea with a new label. The uses are mostly language jobs. You might summarize a long email thread into three lines before it reaches a busy manager. Draft a first reply to a customer question, or pull one detail out of a messy block of text where no fixed field exists. Anything where the job is language rather than extraction fits. Here. Prompts can take inputs too, and that's where they earn their keep. You can pass in a customer email, an order number, or a chunk of text from an earlier step, and the prompt works on whatever you hand it, that turns a one time instruction into something reusable across every record that flows through under the hood. These prompts run on GPT models powered by Azure OpenAI service, which is Microsoft's own hosted version of that technology. You never touch an API key or a developer portal because the whole thing is wrapped inside Power Automate. You can put a person in the loop to add an approval step after the prompt, and the generated text goes to a reviewer who can edit it or reject it before it's used for anything customer facing. That's a sensible habit. Instead of a nice extra, two honest limits are worth carrying with you. This prompt capability is only available in some regions, and Microsoft warns that it can be subject to usage limits or throttling, which means heavy use might get slowed down. Some AI builder features are also still in preview, and Microsoft is explicit that preview features aren't built for production. Check the region and the release status before you promise your boss a launch date. That's the capability picture, and it leaves one obvious thing to settle, which is the cost, what it costs, and why the rules just changed. This is the part that surprises a lot of beginners, so let's be precise. Building and testing AI builder models and prompts is free. What costs money is running them. Every time a trained model processes a real document or a prompt generates real text. It consumes credits. Those credits work like a monthly allowance. Consumption resets on the first of each month, and unused credits don't roll over, so you can't bank them for a busy quarter. If you run out, the model runs get blocked until more credits are bought or assigned to your environment. That blocking is the detail that turns a working automation into a Monday morning problem. There's also a second cost layer that people miss. Any AI builder action is a premium connector, and a flow that uses one is a premium flow. That means it needs power automate premium licensing separate from the free tier. The AI itself isn't the only thing you're paying for. That isn't a reason to avoid it, and it's a reason to test before you scale a flow that runs ten times a day, and a flow that runs ten thousand times a day look identical in the designer and the bill at the end of the month doesn't. Now let's look at the part that changed. In October of twenty twenty five, Microsoft announced a progressive end of AI builder credits. Read that carefully, because it doesn't mean AI builder is going away. The features keep working exactly as before. What changes is the currency? Instead of AI builder credits, AI builder will be paid for with copilot credits. The calendar breaks down. In plain terms. New sales of the old AI builder credit add ons stopped on the first of November twenty twenty five. Existing customers keep using the credits they already have until their contracts end credits. The ones bundled into premium licenses wind down around the first of November twenty twenty six. If you're planning a project right now, that timing matters because you don't want to build a budget on a credit type that's being retired on relative cost without inventing numbers. The shape is this. Reading a page with custom document processing costs far more credits than a quick sentiment check. Simple text jobs are cheap, heavy document jobs aren't, and a couple of the pre-built readers business card and prediction come out free. You can also track all of it in the admin center, so this stops being guesswork once you're live. So the money question has a clear answer. It's cheap to explore, and it isn't free to run at scale. Which brings us to where AI builder stops being the right tool, where AI builder stops and Azure begins. AI builder isn't a separate island. Microsoft built it on top of Azure's document AI, which is the same engine developers use directly. So when you click through AI builder, you're using serious technology behind a friendly face. That's also the cleanest way to understand the boundary. AI builder is about speed and accessibility. Inside Power Platform Azure is about flexibility, scale, and reaching across systems. AI builder is aimed at makers the ones who build with clicks. Azure services are aimed at developers, the ones who build with code and APIs. The practical limits show you where that line sits. A single document can't exceed twenty megabytes. A model can process up to three hundred and sixty documents per environment every 60s, which is plenty for one team and a rounding error for a bank. Fields that split across a page boundary aren't supported, and several models are still in preview, so they aren't ready to run a business on those numbers. Sound abstract until you do the math on your own inbox, which is exactly the exercise worth running before you commit. So here's a simple rule of thumb stay with AI builder while your workflow lives inside Power Platform, your documents are reasonably structured, and your volume is a departments worth rather than a continents. Start looking at Azure when your paperwork arrives in wildly different shapes when your volume is enormous, or when the extracted data has to travel into systems far beyond Dataverse. None of that is a knock on AI builder. It's the opposite, because knowing the ceiling is what lets you pick it with confidence instead of hoping it holds. You don't have to choose one forever, either. Plenty of organizations run both using AI builder for the everyday flow and Azure for the heavy lifting, so they never have to trade speed for scale. And this space keeps moving. Microsoft renamed the prompt Action change how the credits work and keeps adding capabilities. So the ceiling you see today may look different in a year. What to remember and where to start. If you remember one thing, make it. This AI builder is the part of Power Automate that lets you add real AI without becoming a developer. It's free to build with. It costs credits to run, and it starts with a handful of your own documents rather than a big project. So here's my question for you. What's one task in your week that you'd happily hand to a machine if it only took five examples to teach it? Drop it in the comments, because the answers people share are usually the best ideas in the room. If this made the topic feel less intimidating, subscribe to M365 FM to master the tools and the mindset shaping the future of work.