Matching trust — designing the UX for goLance's AI cultural-fit engine

At marketplace scale, matching on skills alone kept producing mismatched, short-lived engagements — the real gap was cultural and working-style fit.

The company was betting on AI to close that gap. But AI matching only creates value if hiring users trust and act on what the model surfaces — my job was to make opaque, probabilistic outputs legible and credible inside real hiring decisions. I owned design end-to-end across the freelancer portal, client side, payments, and the platform's AI features; the models were built by the AI/data team.

The Cultural Fit Assessment: an LLM-supported, scenario-based evaluation where adaptive questions build a real cultural profile.

I designed it end-to-end — how AI-generated questions branch, and how answers resolve into a profile the matching model can actually use. The same structure types the company side too: the assessment returns a named culture type and plots it on a Competing Values quadrant, so a client reads their own organisation in the same terms a match is scored in.

Compatibility scores read as confidence, not verdicts — placed at the decision point, informing the hire without overclaiming certainty the model couldn't back.

Matching is a learning model rather than static scoring, and the interface had to say so honestly. "Best Match" indicators live inside the hiring flow, right where the decision happens.

A score that shows its work — the weights, the drivers, and which way each one moved, on the same screen as the number.

A composite score is only as trustworthy as its arithmetic is visible, so the interface states the formula instead of hiding it, and every driver carries both its own reading and its change since the last period. A client can see which part of the picture moved before deciding how much weight to give the number.

The company number is not the unit of trust — the rows behind it are.

The same score decomposes to the people it describes: each carries their own reading, the direction it moved, and a plain sentiment label rather than a colour alone. A client can see which parts of the picture drive the headline instead of taking one badge on faith — and designing a score that describes people means being deliberate about what it exposes, and to whom.

Designing for imperfect AI: on Mango AI, every summary, insight and flagged issue keeps its reasoning inspectable — so people trust it enough to act.

For the contractor-management assistant I prioritized what a client needs to verify before believing an automated flag, rather than presenting model output as fact.

The result: +25% project engagement, −22% checkout drop-off, −40% support load — and hiring users acting on model outputs they used to ignore.

Broader usability work cut task-completion time by 30%. Looking back, I'd push explainability further — surfacing why a match scored the way it did, not just how confident the model is.

UX research · information architecture · interaction design · AI/LLM UX · design systems · WCAG accessibility · cross-functional work with AI/data · Figma