If you run RPA in a government shop, you already know the failure mode. A vendor pushes an update to a court case management system, a permit portal, or a financial platform overnight. The button moved twelve pixels, the field got a new internal ID, and by 7 a.m. half your desktop flows are throwing “Element not found” and nobody downstream can file anything. The automation didn’t get smarter or dumber; the ground moved under it.
Power Automate’s 2026 Wave 1 introduces self-healing for desktop flows, which is Microsoft’s answer to exactly that problem. It is worth understanding now, because the version of this story that matters for government is not the demo you saw at a conference.
What Self-Healing Actually Does
Self-healing is a runtime fallback for one specific error: “Element not found.” When a supported UI or browser action can’t locate the element it expects, instead of failing the run, Power Automate captures context (a screenshot of the missing element, the parent window title, the full desktop image), uses AI to identify the most likely intended element, recaptures the selector, and continues. If it can’t, normal error handling resumes.
The order of operations matters, because self-healing does not replace your existing error handling; it slots into it. At runtime, an action evaluates: retry policy first, then self-healing, then your “set variable / run subflow” rules, then continue or throw. It is a safety net under your existing net, not a substitute for either.
Read the Limits Before You Celebrate
This is a preview feature, and the boundaries are narrow on purpose. Self-healing only handles “Element not found.” It only applies to actions that touch a single UI element, like clicking a button or selecting a checkbox; it does not cover window handling, screen handling, or any action that interacts with multiple elements. It runs only during console and cloud runs (attended and unattended), not during designer runs. And because it relies on generative AI to guess the right element, it can occasionally guess wrong.
That last point is the one compliance people should sit with. An automation that quietly repairs itself and keeps going is convenient until you need to explain, on the record, why a process did what it did. If the runtime made an AI-assisted decision about which control to click in a statutory workflow, that decision needs to be visible and reviewable. The capability is real; the governance around it is your job.
The Part Nobody Is Telling Government Teams
Here is the detail that changes the whole timeline for public sector shops: self-healing is not available in the government clouds yet. Microsoft’s own documentation states it is not currently available in GCC (Government Community Cloud), GCC High, or DoD. The reason is plumbing, not policy theater. Self-healing runs on external AI models operating as subprocessors, and those models are not enabled in the government clouds today. No subprocessor, no self-healing.
So the honest message for a government RPA lead is not “turn this on.” It is “this is coming, and the work that makes it useful is work you can do now, regardless of when it lands in your tenant.”
Self-healing rewards flows that were already engineered well. It does very little for flows that were held together with hope.
How to Prep Your Existing Flows Now
None of the prep work requires the feature, which is the point. First, stop relying on fragile selectors. Flows that anchor to stable attributes survive UI churn far better than ones pinned to coordinates or auto-generated IDs, and they give a future AI repair a fighting chance to find the right element. Second, put real error handling on every action that touches a legacy app; configure retry policies deliberately rather than leaving everything on defaults, because self-healing sits between your retries and your fallback rules. Third, instrument your runs so you can actually see what happened; if you can’t reconstruct a flow run today, you will not be able to audit an AI-assisted one tomorrow.
Then get the platform groundwork ready so you are not scrambling when availability arrives: confirm you are on a current Power Automate for desktop build, and understand which generative AI and external-model settings your environment will need turned on, who has to approve them, and where the data moves. In a government environment, that approval chain is the long pole, not the toggle.
The teams that benefit most from self-healing will be the ones whose flows were resilient before it existed. The feature buys you uptime when a vendor breaks their own UI; it does not buy you out of basic engineering discipline. If anything, it raises the bar, because now a flow can fail silently into a repair instead of failing loudly into a ticket.
Who’s Writing This
I’m Jacob, the engineer behind Puget Sound AI, a veteran-owned small business. I build Power Platform automation and AI agents inside production GCC environments, under the constraints that actually apply, not the commercial defaults. No account managers, no slide decks you get billed for; just the person who does the work.
If your desktop flows are one vendor update away from a bad morning, it is worth getting them resilient before Wave 1 logic reaches your tenant. Let’s talk.