An internal AI assistant employees had stopped trusting

Turning a mandatory category picker into a system that answers with confidence.

Industry
Enterprise software
Role
Product Designer
Team
1 Researcher, 1 Product Owner, 4 front-end and back-end developers
Duration
1 year
Year
2026
Platform
Web app (internal tool)

This case study was completed under NDA. Identifying details have been anonymized; performance metrics shown are accurate.

Context

An internal AI assistant that lets employees ask questions against their team’s document base, with access scoped by team, meant to save people from digging through internal documentation themselves. By the time I joined the product it had already been live for about a year.

I was brought onto the project with a narrow, specific brief: fix the onboarding experience. That was supposed to be the fix. It did not take long working alongside the researcher to realise the onboarding was only the surface of it. Adoption and retention were low for reasons that went well beyond a first-run screen, and the scope expanded from there into the deeper problems below.

Challenge

Before designing anything, we needed to understand why people were leaving. I worked closely with our researcher on the discovery phase by helping shape discussion guides, taking notes in sessions, and building the prototypes we used to test ideas with real users, on real work tasks.

What we found reframed the whole problem: users were not confused about the tool. They were dropping off at a specific, fixable point, and leaving for good after one bad answer.

Product setup

  • 1 Researcher
  • 1 Product Designer (me)
  • 1 Product Owner
  • 4 front-end and back-end developers

Key design decisions

A reworked onboarding flow

Before any question was answered, users had to select a tile from hundreds of categories organized around internal document codes (the largest single category covered fewer than 10% of the questions people actually asked). Most people read this step as a sign the tool was broken and left before typing a single question. I designed a short, re-accessible onboarding flow that replaced the category wall with a clearer first impression: what the assistant could and could not do, set in plain language, with a way to reopen it later. This shipped early as a quick win while the deeper research was still running.

From a mandatory category grid to task-based groups

Once the research confirmed the category step as the primary drop-off point, I redesigned the entry point itself, moving from a mandatory grid of hundreds of tiles to a small number of task-based groups people could actually recognize themselves in. Selection became optional rather than a gate.

A confidence-threshold answer experience

The old system routed every question through category matching, regardless of how well it understood the question. I designed the interface for the new model behind it: when the assistant is confident, it answers directly; when it is not, it asks one clarifying question instead of guessing. Designing this meant defining what “confident” and “uncertain” should each feel like, so an uncertain answer reads as a helpful next step, not a dead end.

Making sources feel trustworthy, not just present

Part of rebuilding that trust was how answers showed their work. I redesigned how source documents were surfaced alongside an answer, moving from a bare citation list to something more legible and engaging, so users could quickly judge whether an answer was grounded in the right document before acting on it.

Bringing structure to a team that had none

Design and research were not the only things missing structure. When I joined, there was no documentation of what needed to be done, by whom, or by when. In practice, that meant the product owner was going straight to developers with requests, bypassing the design process entirely.

I proposed a fix: a regular cadence of team discussions, weekly and sometimes more, where we aligned on the current iteration and planned the next one together. It was not a design deliverable in the traditional sense, but it is what made the rest of this work possible: without it, none of the research findings would have made it from a report into things actually being built.

Status

Onboarding is live, and most of the other updates were already built when I left. Usage went up after the trust-signal update, although I do not have the numbers. I was removed from the project because of the layoff and the company closing its operations, so I do not know what happened next.

Impact and lessons learned

This product reinforced something I already believed going in: the power of reframing conversations with product owners, and the power of being honest and open to collaboration. Getting a product owner who was used to bypassing design to instead sit down weekly and plan together was not a given. It took consistently showing up with openness rather than territory to defend, and reframing disagreements as shared problems to solve rather than positions to win. That is ultimately what let the research findings turn into real, prioritized changes instead of a report nobody acted on.