Skip to content
artifx
← Case studies

UX for AI: designing chatbots and agents people actually use

How to design AI interfaces people trust: recognisable AI output, findings that lead to actions, confidence in plain language, and agents that ask before they act.

Zentra ticket detail on a laptop: activity timeline, details and AI risk assessment

Most AI features that go unused fail at the interface, not the model. These are the patterns I rely on, from Zentra and Vytrac.

Make AI output recognisable

In Zentra, AI cards carry a star marker, and strong colour is kept for status badges and that marker, so AI results never blend into one more chart.

Lead with a finding, then an action

Each Zentra AI card gives a plain-sentence finding and a suggested fix. One sentence on what the model found, one clear next step, and a way to dismiss it.

Show confidence in words

A score of 0.87 means little to a clinician. On Vytrac, the question was how AI insights, symptom trends and model confidence reach clinicians in a form they can act on:

  • Levels, not decimals: high, medium or low, with a short reason.
  • What it’s based on: the readings or period behind the result.
  • Lower confidence, lighter weight: a suggestion to check, never a verdict.

Put the evidence next to the claim

Vytrac’s timeline gives every vital its own lane and marks where readings crossed a limit, so the reason behind an alert is on the same screen.

Vytrac patient profile on a laptop: vitals timeline, team notes, outstanding alerts and patient information panel

Chatbots: design the hand-off first

Decide what the bot never answers, and when and how a person takes over, before designing the happy path. Urgent or sensitive requests should reach a person straight away instead of getting a best guess.

Agents: ask before acting

Zentra’s optimisation agent runs when a manager starts it, and its suggestions appear in the ticket list next to the tickets they concern. For any agent, decide what it may do alone, what needs approval, and how it reports what it changed and how to undo it.

Zentra all tickets list on a laptop with priority, AI risk and status, plus agent suggestions

Prototype AI features with realistic outputs, including wrong answers and empty states; that’s where trust is won or lost. See AI product design.

Questions about this work? Let’s talk it through.

Set a meeting

Let’s talk it through

30 minutes on Google Meet, at a time that suits you.

Set a meeting

Set a meeting

Book a call in an open slot, or just send a message.

What is it about?
Pick a time

Choose a day