AI in Power Automate: How Intelligent Automation Works

AI in Power Automate is the set of artificial intelligence capabilities you can run inside a flow, mainly AI Builder models, prompts that run on GPT models by default and text and document intelligence such as OCR and classification. It is not one feature. It is a toolkit you combine so a flow can read content, interpret it and act on it, rather than just move data from one place to another.

What does AI in Power Automate actually mean?

AI in Power Automate means running AI capabilities as steps inside a flow so automation can act on the meaning of content, not just its arrival. In practice that is AI Builder models, prompts running on GPT models by default and text and document intelligence like OCR, sentiment and classification, all called as actions alongside your normal connectors.

For years Power Automate handled rule-based work. A flow sent an email when a file landed or created a record when a form was submitted, without knowing what the file or the form said. In the Australian businesses we work with, the shift is towards flows that understand documents, classify requests and support decisions. That shift is one part of the wider picture in our AI for business guide.

The common misconception is that AI in Power Automate is a single switch. It is a collection of capabilities you assemble for the job in front of you. A flow might extract fields from an invoice with one model, classify the document with another, then draft a summary with a prompt. Each capability runs as an action whose output feeds the next step.

The capabilities you can call inside a flow include structured AI Builder models for extraction, classification and prediction, prompts for language tasks, optical character recognition for images and PDFs and text intelligence such as language detection, sentiment analysis and key phrase extraction. In its AI Builder in Power Automate overview, Microsoft maps each AI Builder action to the capability it uses.

What is AI Builder and how does it work in a flow?

AI Builder is the structured AI layer in Power Automate, built for predictable, repeatable business processes. It applies machine learning models without code and returns structured outputs a flow can reliably act on, which matters when AI is driving approvals, system updates or compliance steps. You build and test models for free and only consume credits when a flow runs them.

AI Builder suits steps where the output drives the next action, because its outputs are consistent and auditable. It is strongest where the outcome must be the same every time: extracting fields from invoices, forms or contracts, classifying incoming emails or documents, identifying entities such as names, dates or reference numbers and predicting outcomes from historical data. Because the output is structured, the flow can branch, validate and handle exceptions safely.

AI Builder models come in two forms. Prebuilt models such as invoice processing, receipt processing, text recognition and sentiment analysis run with minimal setup. Custom models, document processing among them, are trained on your own data for content that is unique to your business. We typically recommend a custom model when AI output directly affects customer communication, billing or operational decisions and when a prebuilt model is not accurate enough on your own documents.

How do you use GPT in Power Automate now?

GPT in Power Automate now runs through AI Builder prompts, added to a flow with the Run a prompt action. The old standalone Create text with GPT action has been deprecated, so flows built the old way should be migrated. The prompts run on GPT models through Azure OpenAI by default. The capability has not gone away. Only the way you add it has changed.

This matters if you are working from older guides or an older flow. Microsoft’s documentation marks the original text generation action as deprecated and points makers to its guide to using a prompt in Power Automate instead. The replacement action was called Create text with GPT using a prompt until May 2025, when Microsoft renamed it Run a prompt. You create a prompt once, give it input variables, then call it from any flow.

You also choose the model behind each prompt. Microsoft’s model version and settings page lists GPT-4.1 mini as the default model, with GPT-4.1 and GPT-5 chat among the other GPT models you can switch to. The same page lists Anthropic’s Claude Sonnet 4.6 and Claude Opus 4.6 and puts the Claude models at the experimental stage, meant for experimentation and not for production use.

Generative AI plays a different role to AI Builder. It is strong at understanding and producing natural language: summarising long emails or documents, drafting responses or internal notes, interpreting unstructured text and giving staff context for a decision. What it does not do well is structured, deterministic decision-making on its own. GPT output is probabilistic, so the same prompt can return different results. Microsoft’s overview of prompts is explicit that “human oversight is an important step” with GPT-generated content. Build a review step into any flow where the output is customer-facing or feeds a real decision.

AI Builder vs GPT, which should you use and when?

Choose AI Builder when you need structured, auditable output and GPT when you need to understand or generate language. The two are complementary, not competing. Mature flows usually use both: AI Builder extracts the structured data, GPT summarises or explains the context and Power Automate orchestrates the whole thing. The table below sets out how we guide that decision.

The practical rule we use with clients: if the next step in the flow depends on a value being exactly right and the same every time, that is AI Builder. If the next step is a human reading a draft or a summary, that is GPT. Most real solutions combine them.

How does Power Automate act as the orchestration layer?

Power Automate does more than trigger AI actions. It orchestrates them across systems. It controls when the AI runs, validates the output, applies your business rules, handles exceptions and routes the result to the right person or system. That orchestration is what makes AI usable day to day.

A typical intelligent flow follows a clear sequence:

1. A trigger fires, for example a document lands in a SharePoint library or an email arrives in a shared mailbox.
2. An AI Builder model extracts or classifies the content into structured fields.
3. The flow validates the output against expected formats and a confidence threshold.
4. A GPT prompt drafts a summary or response where natural language helps.
5. Low-risk cases that pass validation proceed automatically. Higher-risk cases, failed checks and any drafted reply to a customer route to a person for review.
6. The result is written back to the system of record and the exception, if any, is logged.

We design flows so AI supports a decision without quietly taking it over. Confidence thresholds, human review points and fallback logic belong in the first version of the flow. This is the same discipline we apply across our Power Automate work, where the goal is reliable automation a business can stand behind. Where a flow uses personal information in a decision that could significantly affect someone, our guide to automated decision making under the Privacy Act shows how to check whether it must be described in your privacy policy.

What can businesses actually do with AI in Power Automate?

The highest-value uses cluster around document-heavy, classification-heavy and communication-heavy work. AI in Power Automate is most useful where staff currently read content, make a routine judgement and key the result into a system. That is exactly the loop a flow can take over, with a human reviewing only the edge cases.

These are the most common patterns we see in client environments:

• Process documents. Extract key fields from invoices, contracts and forms, validate them against expected formats and route each document by what it says. This pattern is strongest in finance, legal and operations, where manual review slows a digital process.
• Classify and route requests. Categorise incoming emails or requests by topic or urgency, route them to the right queue and escalate anything matching risk criteria. That cuts response times and keeps handling consistent.
• Make approvals smarter. Auto-approve low-risk requests, flag high-risk cases for review and give approvers a summary and the context they need to decide.
• Check data quality. Validate extracted data before it updates a core system, catch duplicates and anomalies and enforce consistent naming. This matters most when Power Automate is bridging Microsoft 365 and finance or CRM platforms.
• Communicate and monitor. Draft contextual replies and internal summaries and detect patterns that signal delays or unusual activity so teams get an early warning before a problem escalates.

In practice the value is not spread evenly across these patterns. The fastest payback we see is classification and routing. It removes a daily human bottleneck at low risk and is simple to validate. Document extraction pays off too. Standard invoices and receipts can often start on a prebuilt model, while other documents usually need a custom model. Either way the flow needs a proper validation and exception design before it can run unattended. We usually make extraction the second project, once a classification flow is running. The pattern that disappoints is pointing a prompt at a messy inbox and expecting a clean decision with no fixed categories, validation or fallback. A language model interprets what it reads. Its answer needs checking before the flow acts on it. Match the capability to the job and the returns are real. Try to make one capability do everything and the flow becomes the thing you cannot trust.

How is AI in Power Automate licensed?

AI in Power Automate is licensed on a consumption basis. Building and testing models and prompts is free and running them in a flow consumes credits. Microsoft overhauled this in late 2025. The licensing model new customers use is different to what older guides describe and it is worth understanding before you scale a flow that calls AI on every run.

The mechanics changed on 1 November 2025. According to Microsoft’s AI Builder licensing overview, new customers can no longer buy AI Builder capacity add-ons and must purchase Copilot Credits to run AI Builder features and AI Builder trials were discontinued. From 1 November 2026, new and renewed premium licences such as Power Automate Premium no longer come with seeded AI Builder credits. Microsoft’s End of AI Builder credits calendar sets out the dates. Microsoft’s licensing overview says eligible licences bought or renewed before 1 November 2026 keep their seeded credits until the contract term ends.

Microsoft’s own pages disagree on two points. Its licensing overview says any flow using an AI Builder action is a premium flow, while Microsoft’s Licensing and AI Builder credits page says adding an AI Builder action does not make a flow premium. On seeded credits, Microsoft’s licensing overview says they last until the contract term ends, while its credits page says they are removed on 1 November 2026. Any flow that runs on seeded credits should have Copilot Credits available to its environment from 1 November 2026. Confirm the licensing for the flow itself before you build. Our guide to what a Power Automate flow is explains what premium licensing covers.

We keep vendor pricing out of this guide because the figures change often and date quickly. The useful way to think about cost is consumption against a measured return on one workflow. If a flow runs an extraction model on a few hundred documents a month, that consumption is a known, plannable cost that is easy to weigh against the staff hours it removes. The trap is switching on AI across many flows before the first one has proven its return. McKinsey’s State of AI in 2026 survey found that one in five respondents says their organisation is limiting AI use because of operating costs. AI actions use credits each time they run, apart from a few free ones such as the business card reader. If the flow’s environment has no credits left, Microsoft says the model runs are blocked. The flow then fails at that step unless it is built to handle the error.

What governance and security considerations apply?

Governance is not optional with AI in a flow. The core controls are choosing which AI models makers can use and where those models process data, adding human review for low-confidence output and customer-facing drafts, logging AI decisions and naming a clear owner for prompts and models.

The risks AI introduces are mostly about data and trust. Generative output can be confidently wrong. Generated content that is customer-facing, financial or compliance-related needs a person to sign off. That is why the review step belongs in the flow by design. EY’s Global AI Sentiment Study, fielded in March 2026, found 68 per cent of Australians are worried about losing control over decisions made by AI on their behalf. We read that as one more reason to keep a named person on any AI-drafted reply and any higher-risk decision that reaches a customer.

Data residency and access also matter. AI actions should run only on content the flow is entitled to. Turn on Secure inputs and Secure outputs for actions that handle sensitive fields, which keeps the values out of the flow’s run history. Pass a prompt only the fields it needs. Model choice plays into data residency as well. Microsoft’s prompt model availability table lists GPT-4.1 mini and GPT-4.1 as generally available in Australia. It marks the GPT-5 chat and reasoning models and both Claude models as cross-geo for Australia, which Microsoft says means they might process data outside your region. In the Power Platform admin center an admin decides whether makers can pick models hosted outside Microsoft, such as Claude, using the setting described in Microsoft’s guide to allowing external models. That setting does not cover GPT-5, which Microsoft hosts on Azure OpenAI. The other lever is the environment’s Move data across regions setting. Microsoft’s page on moving data across regions says that with it selected, prompts and results might be processed outside your region, for example when the required model isn’t deployed locally. Before a prompt handles sensitive data, check which model it uses and whether the environment has Move data across regions selected.

A real example: extraction and classification for a Sydney client

A Sydney-based organisation came to OfficePro Consulting with a high volume of inbound documents that staff had to open, read and sort by hand before any real work could start. We built a Power Automate flow that uses AI Builder to extract the key details, classify each document and route it, with people stepping in only on the exceptions.

The manual step was the bottleneck. Every document waited for a person to categorise it and pull out the details that decided where it went next. That did not scale as the volume grew. It was inconsistent too. The call depended on who happened to be working the queue that day and how much of a backlog had built up behind it.

The real design question was whether the flow could trust what the AI returned enough to act on it without a person in the way. That is why we built it on AI Builder and not on a free-text prompt. AI Builder returns structured fields. The flow routes on those values. A low confidence score sends the document to a person. A validation step between the extraction and the routing catches a missing or malformed field where it happens, before it can reach the wrong queue.

The read-and-sort step that used to gate everything largely disappeared. Handling became consistent across the whole intake regardless of who was on. The hours that had gone into repetitive triage moved to the cases that needed a person’s judgement. The job shows where AI in Power Automate earns its place: a high volume, content-driven, repetitive task, structured output the flow can rely on and a human kept in the loop exactly where the risk sits. The client is kept anonymous at their preference and the outcome is described qualitatively.

When is AI in Power Automate the right choice?

AI in Power Automate is the right choice when AI output needs to trigger an action, when the process spans Microsoft 365 and business systems, when business users need visibility and control and when low-code delivery matters. It is the execution layer that makes AI usable in day-to-day operations, not a replacement for a data science platform.

There are limits worth naming. For heavy custom modelling or advanced analytics, direct Azure AI services or a custom application are a better fit. Knowing when not to build is half the skill. When an AI project stalls, in our experience the cause is usually not technical: a poorly defined problem, messy underlying data or treating AI as a software install when it is a change in how people work. We cover those failure modes in detail in our guide to why so many SMB AI pilots fail. If your interest is building AI agents on top of this same Microsoft stack, the companion guide on how Microsoft Copilot Studio works is the natural next read. Our AI and Intelligent Solutions services also cover agents, Copilot rollouts and Azure AI builds.

Frequently asked questions

Final thoughts and next steps

AI in Power Automate is worth adding where it lets a flow understand content, cut manual effort and support better decisions, with governance and a human in the loop where it counts. Applied to the right repetitive, content-heavy work, it turns Power Automate from a workflow tool into a practical operational layer.

If you are weighing up where AI in Power Automate would pay off in your business, book a free initial conversation with OfficePro Consulting. We will look at the workflow you have in mind, tell you plainly whether it is a good fit and show what a sensible first build looks like.