Power Automate AI Builder: What It Does, and What the Credits Really Cost

Every guide to Power Automate AI Builder I read while researching this one is a tour of the features. A list of prebuilt models, a screenshot of the flow designer. None of them mentions that the currency AI Builder runs on is being switched off in about ten weeks.

Owners ask me about this tool because they already pay for Microsoft 365 and would quite like the invoices sitting in their inbox to type themselves into Xero. That is a fair ask, and AI Builder is genuinely one of the shortest paths from that wish to a working flow. I have no interest in talking anyone out of it.

What I do want is for you to see the meter before you build anything on top of it. So this is the version I would want before spending. What the tool does, what the credits in your licence actually buy, what changes on 1 November 2026, and what it will cost you afterwards. Every figure here comes from Microsoft’s own documentation, and where two Microsoft pages contradict each other, I say so rather than picking the tidier one.

What is Power Automate AI Builder?

AI Builder is the Power Platform feature that lets you drop a trained AI model into a Power Automate flow from a menu, without writing code. It reads documents, pulls named fields out of invoices, runs optical character recognition, scores sentiment and runs generative prompts, all as ordinary actions inside the flow designer.

It is not a separate product you go and buy. It appears as an AI models entry in the left hand navigation at make.powerautomate.com, and as a set of actions you can search for while building a flow. If you have ever added a “Send an email” action, you already know the interface.

The part that surprises people is the billing. Your flow runs are metered one way, under your Power Automate licence. The AI Builder action inside that flow is metered separately, in its own currency, at its own rate.

Two meters, one flow. Almost nothing written about this tool makes that clear on the first read.

Which AI Builder models can you actually use?

Two kinds. Prebuilt models that Microsoft has already trained and you can use immediately, and custom models that you train on your own sample documents.

The prebuilt list covers invoice processing, receipt processing, identity document reading, business card reading, contract processing (still in preview), text recognition, sentiment analysis, key phrase extraction, language detection, category classification and entity extraction. Alongside those sit AI prompts, plus a handful of ready made prompt actions Microsoft calls AI functions: AISummarize, AIExtract, AIReply, AIClassify and AISentiment.

Custom models are the interesting ones for most businesses. Document processing trained on your own layouts is the flagship, and it is the reason people bother with AI Builder rather than a generic chatbot. If every one of your six main suppliers sends a differently shaped delivery note, a custom document processing model learns those six layouts and returns the same clean fields every time. You can also train custom object detection, category classification, entity extraction and prediction models.

Two things are genuinely free, and nobody advertises them. Training a model costs nothing. Testing a model, in the AI models page or in the prompt builder, also costs nothing. So you can prove the thing works on your real documents before a single credit leaves your account. Do that.

How do you add AI Builder to a flow?

Five steps, and the first one is the only one that takes an afternoon.

1. Open AI models in Power Automate and either pick a prebuilt model or train a custom one on your sample documents.
2. In your flow, add an action and search for the model by name, for example “Extract information from invoices”.
3. Feed it the file content from wherever the document arrives, usually a SharePoint library or an email attachment trigger.
4. Map the fields it returns, such as invoice total or supplier name, into whatever comes next. Usually an Excel table or an approval request.
5. Test it. Testing is free, so run it against a dozen real documents before you trust it with anything.

The step that catches people is between four and five. A flow that works beautifully on ten invoices behaves very differently on four hundred, because that is the point where the credit meter starts to matter.

What do 5,000 AI Builder credits actually buy?

Less than you think. A Power Automate Premium licence seeds 5,000 AI Builder credits per user per month, which reads like a generous free allowance right up until you check the rate card.

Fifty pages a month. That is the honest answer for the model most owners actually want, the one trained on their own supplier paperwork. It is roughly two and a half invoices a working day.

Microsoft is not hiding this, to be fair. Its own licensing FAQ says the 5,000 credits let you “extract data from a few documents” or perform hundreds of basic text recognition extractions. That is a candid description of an evaluation allowance, and it is the opposite of how every partner blog frames the same number.

One more thing worth knowing before you plan around it. Unused credits do not carry over to the next month, so a quiet January buys you nothing in a busy February. Microsoft’s guidance is to size your purchase to your single busiest month, not your average one.

AI Builder credits are being switched off in November 2026

This is the part missing from every article currently ranking for this keyword, including the ones specifically about AI Builder pricing. In October 2025 Microsoft announced a progressive end to AI Builder credits as a currency, and the final date lands on 1 November 2026.

Read that middle row again if you are shopping for this today. As a new customer you have already lost the option to buy AI Builder capacity, and the free trial went with it. You are buying into Copilot Credits whether the tutorial you followed mentions them or not.

There is no conversion between the two currencies. Microsoft is explicit that seeded credits do not become Copilot Credits on 1 November 2026, they simply stop existing. What happens instead is a fallback: a flow uses AI Builder credits first, then Copilot Credits if those run out, and fails if neither is available.

One trap sits inside that fallback. Consumption is computed roughly daily, but Microsoft admits the calculation can lag five days or more, and the decision to fall back or block is based on the last computation. So a flow can quietly burn through far more capacity than you allocated before anything stops it. The good news is that overage is never billed. The bad news is that your flow stops dead, mid month, with an EntitlementNotAvailable error and no warning email.

If you would rather find out where automation genuinely pays before you commit to anyone’s credit meter, the AI Opportunity Scorecard is a free two minute test that estimates the hours and money automation could save you.

What will it cost once the credits go?

More, for the document jobs most small businesses want, and slightly less for the simplest ones. Microsoft publishes both rate cards in US dollars, so here is the same work priced in each currency.

Invoice processing is the one to look at. Content processing under Copilot Credits is a flat 8 credits per page regardless of which document model you use, which is simpler to reason about and five times the old rate for invoices. Custom document processing rises more gently, from five cents a page to eight. Plain text recognition actually gets marginally cheaper.

So the change is not a straight price rise. It is a flattening, and flat rates always punish whoever was on the cheap end.

On Microsoft’s UK pricing page, Power Automate Premium is £11.50 per user per month paid yearly and Power Automate Process is £115.30 per bot per month, both before VAT. Copilot Studio appears at £153.80 for 25,000 Copilot Credits per month paid yearly. There is no AI Builder line at all on the UK page.

That £153.80 prepaid pack is the number to be careful with. At 8 credits per page it covers roughly 3,100 document pages a month. If you are processing two hundred invoices, you are paying for capacity you will never touch, and it does not roll over. Buy pay as you go Copilot Credits instead, at a cent each, and let the meter reflect what you actually did. The prepaid pack only starts making sense once you are reliably consuming most of it.

Do you need a premium licence? Microsoft answers this twice

Ask this question and you get two different answers from two current Microsoft Learn pages, so budget for the more expensive one.

The AI Builder credit management page says plainly that in Power Automate, adding an AI Builder action to a flow does not turn it into a premium flow. The AI Builder licensing overview says the opposite, that AI Builder actions are premium connectors and any flow using one is a premium flow requiring Power Automate Premium licensing. Both pages carried 2026 update stamps when I checked them in August 2026.

The two pages do agree on Power Apps. Put an AI Builder action in an app, or embed a flow that contains one, and the app becomes a premium app.

I am not going to pretend I can resolve a contradiction inside Microsoft’s own documentation. What I would do is assume the premium licence is required, price it in at £11.50 per user, and get your Microsoft partner to put the answer in writing before you build a department around it.

Where does AI Builder earn its keep?

When your documents already live in Microsoft 365 and the job is boringly repetitive. That is a real and common situation, and in it AI Builder beats almost anything you could assemble yourself on a Tuesday afternoon.

The strongest case is a custom document processing model trained on layouts that never change. Six suppliers, six templates, four hundred delivery notes a month, and a finance person who currently retypes them. The model learns the six shapes once and returns the same fields forever. Nothing about that requires a developer, and the SharePoint and Outlook plumbing is already there.

The weakest case is the one people usually try first. A single owner, forty invoices a month, all from different suppliers, hoping the prebuilt invoice model handles the lot. It will get most of them right and quietly mangle the rest, and you will spend more time checking than you saved. Volume plus consistency is what makes this worth it. One without the other is not enough.

Worth saying plainly: prebuilt models are trained on the world’s documents, not yours. If your paperwork is unusual, the accuracy you see in a demo is not the accuracy you will get. Test on your own worst documents, not your neatest ones, and remember that testing is free.

What I would build instead

For lower volumes, or for anything shaped oddly, I would write the job myself in Claude Code and pay nothing per page. This is the awkward part of the AI Builder pitch and I would rather say it than dance around it.

A flow that reads a folder of PDF invoices, pulls out the supplier, the date, the net and the VAT, then writes them into a spreadsheet is not a hard piece of software any more. You describe it in plain English, Claude Code writes and runs it, and when it works you own the thing outright. No credit meter to watch, and no November date on which the terms change underneath you.

The trade is real, though, and it cuts both ways. AI Builder hands you a supported product with tenant level governance and a support contract behind it. Building it yourself hands you control and a flat cost, and makes you the person fixing it at nine at night when a supplier redesigns their invoice template.

If you have a compliance team, take the supported product. If you are the compliance team, the second route is usually better value. I have written more about how I work through that choice on my about page.

The deciding question is not really technical. It is whether this job is worth having on a meter at all.

Where I would leave it

If you already pay for Power Automate Premium, go and use your 5,000 seeded credits this month, because in about ten weeks they are gone. Train a custom model on your ugliest supplier paperwork and see what comes back. Training and testing cost nothing, and by November you will know from your own documents whether this is worth paying for rather than from anybody’s blog post.

If you are starting from scratch today, I would not lead with AI Builder. Price the same job as pay as you go Copilot Credits and against an hour of your own time in Claude Code, and pick on the number rather than on the brand. My honest view is that AI Builder wins on volume and governance, and loses almost everywhere else.

If you want a second opinion on which of your processes is genuinely worth automating first, that is exactly what the AI Opportunity Audit is for. It is a £500 fixed fee 60 minute session where I go through your actual processes, find the highest impact place to apply automation, build one live in Claude Code so you can see it running, and leave you with a prioritised roadmap. There is a free 15 minute intro call first, so you can check I am the right person before spending anything.