Raising Work Request Acceptance to 50%
How I built Anderson, an AI coach that helps product managers write better UX work requests.
GE Vernova | UX interaction designer | Spring-Summer 2026
The Problem
My team was turning into a reactive service desk. PMs sent requests through email, MS Teams, and random links – often with no clear problem, no defined user, and no explanation of why the work mattered to users. We'd start designing, then have to redo the work once the real requirements surfaced. That costly cycle kept repeating.
My Role
Staff UX Interaction Designer. Nobody assigned me this problem. I saw it hurting my team's credibility and took it on myself to do something about it. I designed the standard, set the rules it would enforce, and got a product organization I don't manage to adopt it. That meant convincing product leadership to change how requests get submitted, going well beyond fixing my own team's process. I led this from spring through summer 2026.
What I Did and Why
Building a coach that meets each request where it is, instead of forcing every request through the same checklist
I built Anderson, an AI coach in Aha! powered by Elle. PMs were moving fast, but weak requests were costing them more time in rework than a slower, better request would have. Anderson checks each request against 11 required fields before it can be accepted. It also sorts requests into four types first:
new feature
modernization/enhancement
discovery
a quick 30-minute UX consultation
Anderson coaches PMs based on the type, without showing them that logic. I wrote Anderson's prompt using the RICE framework (Role, Instruction, Context, Examples), created by Jeffrey Fowler and released under the MIT License, so its behavior stays consistent. I also used Amp, our company's AI agent, to review and refine what I'd built in Elle before Anderson reached real PMs. I continue to use Amp as a service bay for Anderson, running monthly diagnostics.
Making sure Anderson never does the PM's job for them
The last coach I can remember being a player/coach/manager was Pete Rose, back in 1986. The AI coach I created for product managers stays in its lane – it only coaches.
Anderson can't invent facts or fill in blanks for a PM. If a field is weak, it gets marked "Needs PM Input" instead of guessed at it. "Unknown" is only acceptable with a plan: who will clarify it, and by when. Even when a PM asks Anderson to just write the request for them, it must say no, and offer templates and examples instead. The PM owns the request. Anderson only helps shape it.
Getting people to explain the real problem, not just ask for a fix
PMs used to write requests like "just add a tab." When they use Anderson, it pushes back and asks for the real user problem instead. Using Anderson is optional, but PMs who use it get accepted more often. The same rule applies to the Clear Requirements field: it has to name the user, what they need, and the outcome. I can't be an engineering task dressed up as a user need.
Human in the loop – running reviews on a predictable schedule
I set up a weekly work request review loop for the UX team; this helps identify gaps in requests that might have been missed. Finding gaps/mistakes early helps me to respond to PMs faster so that we can work on closing those gaps without a week-long wait just to find out if edits helped or not.
Every review now follows the same format: accept or reject, top issues, and a field-by-field breakdown. A request that's already strong won't continue to receive coaching – Anderson encourages product managers at that point, telling them that acceptance is highly likely in the UX team work request review.
Early on, some "next steps" looked clickable but weren't – I fixed that, so every suggested next step is something Anderson can actually do. I also made sure Anderson holds its standard under pressure. Disagreement, urgency, and seniority don't count as new information, so a VP's priority gets the same scrutiny as anyone else's.
The Scope
Where It Stands
My agent is shaping how GE Vernoca requests UX work. Anderson launched to the entire org and is in its Beta, with a retrospective planned to gather feedback before it evolves further. Before Anderson, about 10% of weekly requests were accepted on first submission. With the PMs piloting it, that's grown to about 50% – early numbers, not yet company-wide, but a real signal it's working. Anderson has also gotten more disciplined through the pilot: it holds its standard under pushback, won't write a PM's request for them, and only suggests next steps it can actually carry out.
