designing AI agents people trust: what we learned on zaki AI
people trust AI they can follow. here's how we design agents that show their work, ask before they act, and keep people in charge, from our work on zaki AI.
by joseph zakher · 3 min read
an AI agent that does things is a different design problem from a chatbot that talks.
the moment software can send a message, change a project, or join a call, trust becomes the whole product.
since 2025, i've been the founding designer of zaki AI: visible AI work for goals, projects, meetings and design, in english and arabic. it has been live at app.chatzaki.com since july.
here's what it taught us.
show every step
in zaki, a run shows every step as it happens. people see the plan, the tools, and the progress, in words anyone can follow.
it's the simplest trust pattern there is. people trust what they can follow.
victor yocco calls a version of this an intent preview in his smashing magazine piece on agentic UX: the agent shows its planned steps in plain language before it acts. it's one of six patterns he lists, and a good starting point for any team.
ask before anything leaves
in zaki, anything that sends waits for your yes. you can approve it once, or allow it for the session.
that small choice works like an autonomy dial. careful people approve each step. busy people allow the session. both stay in charge.
a good rule: the further an action travels from the person, the clearer the yes should be.
announce the bot
zaki's minutes bot joins a call in the open, and asks first.
it's always clear when people are talking to AI. we bring that rule to every AI product we make.
give agents their own visual language
people should always know who is speaking.
in zaki, actions are ink, and one red belongs to zaki alone. zee, the 8-bit mascot, gives each product its own character: agent, spaces the librarian, design with a beret and a brush, and minutes the note-taker.
on zunou AI, the workspace where people and AI agents work as one team, we went further: its visual language, a glass office, and a daruma ritual for agents.
when the agent has a face and a color, people can tell its work from their own at a glance.
design every state, with real answers
an AI answer has moods. it can be thinking, sure, or less sure.
we design every state the AI can be in, using real answers from the model, early in the design weeks. a screen designed with placeholder text looks calm until the first real answer arrives.
the less sure state matters most. when the agent is less certain, the screen should say so, so people know when to look closer.
make every answer easy to change
people make the final call, and every answer is easy to change.
an agent that's easy to correct is an agent people come back to. the edit button is a trust feature.
design for both directions from day one
zaki works in english and arabic. switch to arabic, and the whole shell mirrors, right to left.
the design system behind it has more than 120 components, in light and dark, right to left in arabic: question cards, plan steps, weekly usage, and tool rows. one system, shared by the web app, the websites, and the phone app.
building both directions in from the start keeps the system clean. and it opens the product to millions more people.
where to begin
you can start with one small thing this week. that's enough.
- list every action your agent can take that leaves the product. give each one a clear yes.
- design three states for every answer: thinking, sure, and less sure.
- give your agent a name, a color, and a voice that's clearly its own.
see the full story in the zaki AI case study, or book a call if you're building an agent of your own.
