Power Automate in 2026: What Actually Changed
Power Automate has shipped more meaningful capability in the first nine months of 2026 than in the previous two years combined. The driver is the platform's pivot from rule-based automation to AI-augmented and agent-driven workflows — a shift that is not just marketing language but is visible in the product's connectors, flow designer, and licensing structure. For enterprises running meaningful automation workloads, there are both significant new capabilities to adopt and important architectural patterns to reconsider.
This article covers the most important changes in Power Automate through September 2026 — the AI Builder actions that are genuinely useful in production, the Copilot Studio integration pattern for agent-triggered flows, the agentic loop flow type (still in preview but rapidly stabilising), and what process mining GA means for organisations serious about automation ROI. We include real flow configurations and code patterns throughout, not just feature descriptions.
AI Builder Actions: What to Use and What to Skip
AI Builder has grown its action library substantially in 2026. Not all of these are equally mature or useful in production. Here is a practical assessment of the most relevant AI Builder actions for enterprise Power Automate flows:
Summarise Text (GA) — Genuinely Useful
The Summarise Text action uses a Microsoft-hosted GPT-4 model to generate a configurable-length summary of any text input. In production this works well for: summarising incoming email bodies before routing to Planner, generating executive summaries of SharePoint document content, and producing incident summaries from concatenated Teams messages. The action accepts up to 32,000 tokens of input and allows you to specify a maximum word count for the output. Latency averages 2–4 seconds for a typical 2,000-word document, which is fast enough for synchronous approval flows but should use asynchronous patterns for batch document processing.
Extract Information from Documents (GA) — Production-Ready with Caveats
Document intelligence extraction has reached a level of reliability that makes it viable for production data capture — invoice processing, purchase order extraction, receipt digitisation. The prebuilt invoice model extracts vendor name, invoice number, line items, totals, and due dates with high accuracy for standard invoice formats. The accuracy degrades significantly on heavily customised or non-standard document templates — for those, training a custom extraction model in AI Builder is worthwhile if you have 50+ sample documents.
Important: the extraction action returns a JSON structure, not flat fields. You will need to use the Parse JSON action and a generated schema to access individual extracted fields downstream in your flow. Set up error handling for the case where extraction confidence is below your acceptable threshold — the action returns a confidence score per extracted field.
Create Text with GPT (GA) — Use with Caution in Automated Flows
The Create Text action allows you to send a prompt to the hosted GPT-4 model and use the completion in downstream flow steps — drafting email replies, generating document sections, creating report narratives. This works well for human-in-the-loop flows where a person reviews the output before it is sent. For fully automated flows that send AI-generated content directly to customers or external parties, establish a policy and approval step — hallucination risk is real and the organisation is accountable for content sent under its name.
AI Builder actions consume AI credits, not standard Power Automate API requests. Each tenant has an AI credit pool — Summarise Text costs approximately 1 credit per 250 words, Create Text is approximately 1 credit per 500 tokens. Monitor AI credit consumption in the Power Platform admin center. A flow that processes 1,000 documents daily can exhaust a standard AI credit allocation within a week.
Copilot Studio and Power Automate: The Agent-Flow Pattern
The most architecturally significant change in 2026 is the tight integration between Copilot Studio agents and Power Automate flows. Prior to this year, connecting a bot to a flow required a manual REST call or a dedicated flow trigger. Now, Copilot Studio agents can directly call Power Automate flows as "actions" — surfaced in the agent canvas as callable tools with typed inputs and outputs.
How Agent Actions Work
When you create a flow with the "Run from Copilot" trigger, it becomes available as an agent action in the Copilot Studio action library. The flow's input parameters become the action's input schema, and the flow's response body becomes the action's output. The agent's AI orchestration layer decides when to call the action — based on user intent and the action's description, which you write in plain English in the action configuration.
// "Run from Copilot" trigger — inputs become action parameters in Copilot Studio { "type": "object", "properties": { "employeeId": { "type": "string", "description": "The Azure AD object ID or employee ID of the user making the request" }, "leaveType": { "type": "string", "enum": ["Annual", "Sick", "Compassionate"], "description": "The type of leave requested" }, "startDate": { "type": "string", "format": "date", "description": "First day of leave in YYYY-MM-DD format" }, "endDate": { "type": "string", "format": "date", "description": "Last day of leave in YYYY-MM-DD format" } }, "required": ["employeeId", "leaveType", "startDate", "endDate"] }
The Copilot Studio agent handles the conversation — collecting the missing information through clarifying questions if the user's initial request was ambiguous — and only calls the flow once all required parameters are available. The flow then executes the business logic: checking leave balances, creating the SharePoint list item, sending the manager approval request, and returning a confirmation message that the agent relays back to the user in natural language.
Returning Structured Data to the Agent
The flow's Respond to Copilot action defines the output schema. Return both a plain-text message field (what the agent says to the user) and structured data fields (for cases where the agent needs to reason about the result). Returning only a message string works for simple confirmations; returning structured data allows the agent to make decisions — for example, if leave balance is insufficient, the agent can offer alternative dates or a different leave type rather than simply failing.
The plain-English description you write for the agent action determines when the agent decides to call it. Write it as a capability statement: "Use this action to submit a leave request on behalf of the user. Call this when the user wants to book time off, request annual leave, or apply for sick leave." Vague descriptions produce unreliable action invocation — the agent will call the wrong action or fail to call the right one.
Agentic Loop Flows (Preview)
The new "Agentic" flow type — currently in preview — represents the most significant addition to the Power Automate canvas in years. Unlike a standard flow that executes a fixed sequence of steps, an agentic flow contains an AI planning loop: the flow receives a goal, generates a plan, executes steps, observes the results, and replans if needed — all within a single flow run, without requiring separate flows or Copilot Studio orchestration.
When to Use Agentic Flows
Agentic flows are suited to tasks where the exact steps are not fully known in advance, where the outcome of one step determines what the next step should be, or where the flow needs to handle exceptions by trying alternative approaches rather than failing. Classic examples include:
- Data extraction and normalisation from varied source formats — the agent tries different extraction strategies and selects the best result
- Multi-system data reconciliation — the flow queries multiple systems, identifies discrepancies, and attempts resolution steps in sequence
- Document processing pipelines where document types are heterogeneous and routing logic is complex
- Support ticket triage — the flow analyses the ticket, queries knowledge bases, and attempts to find a resolution before escalating to a human
Configuring an Agentic Loop
An agentic flow uses a special Do Until (AI-Driven) loop that replaces the standard condition-based loop. Inside the loop, you place a set of available "tool" steps — the actions the AI planner can choose to execute. The planner reads the current state, selects the next tool to call, executes it, reads the output, and decides whether the goal is achieved or another tool call is needed. The loop exits when the planner determines the goal is complete, when a maximum iteration count is reached, or when the flow exceeds its 30-day run timeout.
Agentic loop flows are in public preview as of September 2026. The planning model, iteration limits, and billing model are subject to change before GA. Do not build production-critical workflows on agentic flows yet — use the preview for evaluation and prototyping. Monitor the Power Platform release notes for the GA announcement, expected in early Q1 2027.
Process Mining GA: Finding What to Automate
The most underutilised feature in the Power Automate licensing bundle is now generally available: Process Mining. Process mining answers the question most automation programmes struggle to answer objectively: which process should we automate next, and what will we actually gain from automating it?
The tool ingests event logs — a sequence of timestamped actions with case IDs — from your source systems and automatically maps the actual process as it happens, not as the process diagram says it should happen. In most organisations, the gap between the documented process and the actual process is substantial. Process mining surfaces: which process variants are most common, where bottlenecks occur and how long they last, which cases deviate from the standard path and why, and which steps are candidates for automation (high volume, low decision complexity, consistent inputs).
Getting Started: What Data You Need
Process mining requires an event log with three mandatory columns: a Case ID (the identifier for each process instance — an invoice number, a ticket ID, an order number), an Activity (the name of the step that happened), and a Timestamp (when it happened). Optional columns — resource (who performed the step), outcome, cost — enrich the analysis but are not required to begin.
CaseID,Activity,Timestamp INV-2026-0041,Invoice Received,2026-08-01T09:14:00Z INV-2026-0041,Validation Check,2026-08-01T09:22:00Z INV-2026-0041,Approved,2026-08-02T14:05:00Z INV-2026-0041,Payment Processed,2026-08-03T11:30:00Z INV-2026-0042,Invoice Received,2026-08-01T10:02:00Z INV-2026-0042,Validation Check,2026-08-01T14:18:00Z INV-2026-0042,Returned for Correction,2026-08-02T09:40:00Z INV-2026-0042,Validation Check,2026-08-04T11:15:00Z INV-2026-0042,Approved,2026-08-05T08:50:00Z INV-2026-0042,Payment Processed,2026-08-06T10:22:00Z
Export this data from your source systems (SharePoint lists, Dataverse, SQL, SAP) using Power Query in the Process Mining interface. The tool then generates an interactive process map and a set of conformance analytics automatically. The September GA release adds a Conformance Check that compares your actual process map against a reference BPMN model you upload — particularly valuable for regulated processes where adherence to a documented procedure must be demonstrated in audits.
Power Automate Desktop: AI-Assisted Recording
Desktop flows (RPA) now have an AI-assisted recorder that can generate flow steps from a natural language description without requiring a manual recording session. You describe the task — "open the HR system, find the employee with ID {employeeId}, copy their department and cost centre, then paste into the finance reconciliation spreadsheet at row {rowNumber}" — and the recorder generates a draft flow that can be refined in the desktop flow designer. The AI recorder works best for standard Windows application interactions; complex web applications with dynamic elements still require manual recording or script-level customisation.
Key Connector Updates in 2026
Three connector updates are worth calling out for enterprise developers:
SharePoint — Batch Request Support
The SharePoint connector now supports Graph batch requests natively. A new Send a batch of requests action allows you to bundle up to 20 SharePoint REST or Graph calls in a single action step, with the results returned as an array. This dramatically reduces the API call count for flows that need to read or update multiple list items — previously requiring a separate action per item, now handled in a single step with a foreach over the batch response.
Microsoft Teams — Adaptive Card with Response
The Teams connector's "Post adaptive card and wait for a response" action has been updated to support multi-select input controls and date picker elements, previously only available in Adaptive Cards Designer but not through the Power Automate connector. Approval cards can now capture structured form data — not just a yes/no response — directly from within Teams, without requiring a separate Power Apps form or SharePoint page.
HTTP with Azure AD — Enhanced Certificate Auth
The HTTP with Azure AD connector now supports certificate-based authentication for API calls to protected endpoints — eliminating the need to store client secrets in Environment Variables or Key Vault references. The certificate is uploaded to the connector configuration and rotated through the Power Platform admin center, which is a meaningful security improvement for flows that call APIs with sensitive scopes.
Key Takeaways for Power Automate 2026
AI Builder actions are production-ready for document intelligence, text summarisation, and human-in-the-loop content generation — but not for fully automated customer-facing content without review steps.
Copilot Studio + Power Automate is the right architecture for conversational automation — the agent handles the conversation and intent extraction, the flow handles the business logic and system integration.
Agentic loop flows are promising for variable-path processes but are still in preview — evaluate and prototype now, but hold production deployment until Q1 2027 GA.
Process Mining GA is the most underutilised capability in the Power Automate Premium licence — run it against your invoice processing, HR onboarding, or IT ticket workflows to find your highest-ROI automation target.
Monitor AI credit consumption proactively — AI Builder actions draw from a separate credit pool that is finite and not automatically replenished when exhausted mid-month.