What it does

When a UI or browser automation action fails at runtime with an Element not found error, self-healing captures a screenshot of the missing element, its parent window title and the full desktop image, works out which element was most likely intended, recaptures the selector and carries on. In preview as at July 2026.

Key facts

• It only fires on Element not found errors, and only for actions that interact with a single UI element — clicking a button, ticking a checkbox. Window handling, screen handling and anything targeting multiple elements are out of scope.
• Error handling runs in a fixed order per action: retry policy, then self-healing, then the Set variable / Run subflow rules, then continue or throw.
• Runtime only. It applies to console runs and to attended and unattended cloud runs, and has no effect at all during designer runs — so you can't see it working while you build.
• Two models power it in combination: GPT-4.1 mini and Claude Sonnet 4.5.
• Available to organisation premium accounts at no extra cost, subject to service limits. Not available in government clouds (GCC, GCC High, DoD) or sovereign clouds.
• You need Power Automate for desktop 2.66 or higher, and it's enabled per action in that action's error handling settings, not globally.
• Because Anthropic models are involved, admins must enable Anthropic as a subprocessor in both the Microsoft 365 admin centre and the Power Platform admin centre, tick Move data across regions on the environment outside US regions, and have Copilot features on for the environment. Microsoft enables Anthropic by default for most commercial cloud customers, excluding EU/EFTA and the UK.
• The portal doesn't surface self-healing detail in run action logs yet. A registry setting on the host machine keeps local logs, where a repairWithAIInfo property records whether a repair was attempted, whether it worked and which selector the AI produced.

When to use / skip

Turn it on for the handful of actions in a long-running RPA flow that break every time the target application gets patched — the ones you already know by name because you've fixed their selectors three times. Skip it for anything where clicking the wrong control is expensive. It's generative AI picking a control from a screenshot, and Microsoft say plainly it might occasionally pick the wrong one. On a "Submit payment" button that's not a risk worth taking for a bit less maintenance.

Configuration decisions

• Which specific actions get it enabled. Per-action is deliberate; treat it as an allowlist for low-consequence clicks.
• Whether the organisation will enable Anthropic as a subprocessor at all — for some regulated clients that's the whole conversation and the answer is no.
• Whether cross-region data movement is acceptable for the environment, outside US regions.
• Whether you turn on local action logging so you can actually audit what the AI repaired.
• What still happens after a failed repair, since normal error handling resumes and you still need a real branch there.

Gotchas

• You can't test it in the designer. Validation means a console or cloud run, which catches people out on day one.
• It's not a substitute for a retry policy — retries run first, and transient timing failures are still best handled deterministically.
• No portal visibility yet means a repaired run looks identical to a clean one unless you've set up local logging in advance. Do that before you need it, not after.
• The admin prerequisites are a chain of four settings across two admin centres. Miss one and the toggle is there but nothing happens at runtime.

Consultant notes

• Frame it as maintenance reduction on brittle automations, not as a reason to skip proper selector design. Well-built selectors still fail less often.
• Raise the Anthropic subprocessor point with the client's security or data protection people before you demo it. Finding out at go-live that the tenant won't allow external models is an avoidable problem.
• It's preview. Keep it off production critical paths, and say so in writing.
• Where a client's flows break constantly on UI changes, the real fix is usually an API or a supported connector. Self-healing buys time; it doesn't change the architecture.

---

Preview as at July 2026 — check whether it's gone GA, picked up window and multi-element support, or landed portal-side run logging