What it does

AI Builder puts a set of AI models behind Power Automate actions — text classification, entity extraction, sentiment, OCR, translation, document and receipt reading, prediction. Some are prebuilt and ready to call; others you train on your own data first.

Key facts

• Prebuilt models you can call straight away include business card reader, category classification, entity extraction, key phrase extraction, language detection, sentiment analysis, text recognition and text translation. Contract processing, ID reader, receipt processing and image description were still marked preview as at January 2026.
• Custom models you train yourself: category classification, entity extraction, document processing, object detection and prediction.
• Each capability burns credits at its own rate. The rate card lives in the Microsoft Power Platform Licensing Guide PDF, not in the docs — don't quote a per-call number from memory.
• Actions consume AI Builder credits first; when those run out the system falls back to Copilot Credits. With neither available, AI Builder actions fail with EntitlementNotAvailable, NoCapacity or QuotaExceeded.
• Consumption resets on the first of each month and unused capacity does not carry over. Overage is a grace period and is never charged.
• Seeded AI Builder credits that come with licences (Power Automate Premium and Process both include 5,000, Power Apps Premium 500) are being removed on 1 November 2026. After that date only AI Builder add-on credits remain. New customers can no longer buy the AI Builder capacity add-on at all and must buy Copilot Credits.
• Adding an AI Builder action to a flow doesn't make the flow premium. Adding one to an app — or embedding a flow that contains one into an app — does make the app premium.
• Availability varies by region. Models are deployed wherever your environment lives. Text translation, for example, was only GA in Europe and the United States as at January 2026.
• Training models, testing models and testing prompts are all free. Prompts are the exception in the other direction — they consume credits even in preview.

When to use / skip

Use it when the AI work is a step inside a process you're already automating and the volume is predictable — reading a few thousand invoices a month, classifying inbound email, pulling entities out of free text. It's the fastest route from "we have unstructured content" to "we have a Dataverse row". Skip it if you need model control, custom architectures or per-call cost transparency; you'll be happier calling Azure AI services through a custom connector, and the client's data science team will thank you.

Configuration decisions

• Whether credits are allocated per environment or left as an unallocated tenant pool. Environments with allocated credits only consume allocated credits — there's no automatic fallback to the tenant pool.
• Whether to leave the tenant setting "Allow users to consume unassigned AI Builder credits" on. Turning it off means only environments with an explicit allocation can use AI Builder at all, which is a decent governance lever.
• Which regions the environments sit in, checked against the model availability table before you promise a capability.
• The path off seeded credits before 1 November 2026 — add-on renewal for existing customers, or Copilot Credits.
• Whether outputs are trusted straight through or routed to a human step.

Gotchas

• Leave the Asynchronous Pattern setting on for AI Builder actions in cloud flows. Turning it off is a known way to stop results coming back properly.
• Overage only blocks the environment that's in overage. Other environments carry on, which makes the global consumption graph misleading — it goes red when any one environment is over.
• Credit consumption for prompts depends on input tokens, output tokens and the model chosen, so a prompt that grows a system message quietly gets more expensive.
• If an admin turns off the AI Builder preview toggle, any prompt using a preview model fails everywhere until someone edits it onto a GA model.

Consultant notes

• The 1 November 2026 seeded credit removal is a real budget conversation, not a footnote. Raise it in the licensing session and get the client's procurement route agreed early.
• Estimate consumption from the rate card and expected volume before committing to a design. "It's included in the licence" has been the wrong answer for a while and will be entirely wrong after November.
• Push back on demos built in the default environment with tenant-pool credits. They tell you nothing about what production will cost or whether it'll even be allowed to run.

---

Recheck after 1 November 2026 when seeded credits drop off, and whenever the licensing guide rate card is reissued