Andrew Marks, UX & Product Designer

I was one of five on the provider evaluation. I sourced an option, facilitated the calls, and gave my recommendation to the PO.

• Sourced one of the government options under review and facilitated the evaluation calls.
• Recommended a commercial provider over the government option I had sourced, because its models were a better fit for our work.
• Organized communication channels across the org to surface what other AI projects were underway, which gave the AI team a basis for sizing the initial token request.
• Walked two other teams through the request process and got them set up on the tooling.
• Documented the value the tokens were delivering and where the existing cap was blocking work. Access and spend doubled, and held through a round of org-wide cuts.

The product team I support and the value stream I sit in had no shared process for coordinating. I was tasked with supporting the product team while sitting on the other side, which put me on both sides of the gap. After talking with PMs and POs across both, I recommended starting below the leadership level rather than at it.

• Proposed the sequence now in use: a leadership sync to approve a small group of PMs and POs, a scoped brainstorm, then a larger working group with report-backs to leadership when the work warrants it.
• Roughly five cross-team initiatives have started this way since.
• Early results from those groups gave leadership on both sides a reason to reengage on larger work.
• Built the data-lake workflows behind our user sourcing, and stay hands-on with it.
• Lead the research validating the five-year roadmap the product team owns and takes to leadership itself.
• Identified and now lead the effort connecting our research outcomes to a sibling team's learning and notifications work.

Using the model access and the data-lake integration, I built workflows that identify the specific users a campaign needs. I own the pipeline end to end.

• Studies field in roughly two weeks, down from four to five.
• Response rates on the two most recent research initiatives run 10 to 15 percent, against 1 to 3 percent at best before the pipeline.
• The gain comes from targeting: users who were invisible to us before, and the most active users at the organizations whose feedback matters most.
• On a recent merge of two front-end applications, interviews surfaced needs the design had missed (power-user filters and collapsible sections) and the feature changed before it shipped. That work is recent, so I'm not claiming a downstream metric yet.

I built the research tooling so people other than me could run it, which meant the operating knowledge had to live outside my head.

• Wrote the custom system prompts our researchers work from. Three researchers on my team use them, plus three from other teams on collaborations.
• Synthesis now runs in hours. Getting findings reviewed and presented takes a week or two, against the three to four months it took before. Measured across three collaborations and four campaigns.
• Trained the team on the prompts and on efficient use of the model access, and leveled up teammates with the org context their initiatives depend on.
• Documented the advanced data-lake workflows in our research repo. Sourcing still routes through me, since I hold the only researcher access.
• Working now to extend that access to other researchers and to PMs and POs on other teams, so they can pull sourcing for the feedback their own work needs.