Enterprise customers use Autofill to extract metadata across thousands of files, but had no scalable way to know which results they could trust. Manually checking everything reduced the value of automation and introduced governance risk when critical metadata was wrong.
Instead of directly translating the proposed prototype into UI, I researched Microsoft AI patterns around trust, automation, and metered AI. I explored the end-to-end workflow and proposed an Autofill preview, explainable confidence states, and an in-context review experience. I also presented a lower cost alternative so Product and Engineering could weigh UX quality against implementation effort.
The team supported the direction of surfacing questionable AI outputs while keeping users in control of review and correction. Engineering preferred the full-canvas approach, while the alternative provided a practical path under tighter constraints. Although my contract ended before final handoff, the work helped move Autofill toward AI automates, explains uncertainty, and directs human attention where it matters.
