AI agent UX: best practices for agents people can trust

When a chatbot gets something wrong, the person reads a bad answer and decides what to do with it. When an agent gets something wrong, it may already have acted on it: booked the wrong date, emailed the wrong person, deleted the wrong folder.

So the question people ask changes. With a chatbot it is “is this answer right?”. With an agent it is “what did it just do in my name?”. Every pattern in this guide exists to make that question easy to answer, before, during and after the task.

Agents also work over time. A task can take ten seconds or ten minutes, and people won’t watch it the whole way. The design has to work for someone who glances at it, leaves, and comes back.

People can’t hand over a task well to something whose limits they don’t know. Before the first task, say what the agent can reach and what it will do with it, in plain words: “I can search and book flights and hotels, and pay with your saved card. I can’t change bookings made somewhere else.”

An empty box that says “Ask me anything” promises more than any agent can do, and gives people nothing to start from. Offer a few real tasks instead, written the way someone would ask for them. They show the range, and they show what a good request includes.

Be modest about what it’s good at. An agent that sometimes does more than people expected earns trust. One that often fails at what it seemed to promise loses it quickly. Google’s People + AI Guidebook calls this calibrating trust: helping people trust the system enough to use it, without trusting it so much that they stop checking.

For anything longer than a few steps, show the plan before starting: a short list of what the agent intends to do, in the order it will do it. It is the cheapest moment to catch a misunderstanding. Correcting “search flights from Manchester” costs one edit before the task, and a cancelled booking after it.

Ask the questions that matter at the same time. If the agent needs to know whether a checked bag is included, or which card to use, ask now, so the rest of the task can run while the person does something else. Don’t ask for details it can reasonably work out, and don’t start on a guess about something it can’t.

For a short task, skip the plan and just do it. Nobody needs to approve a plan for checking tomorrow’s weather.

A spinner is fine for two seconds. For two minutes it leaves people guessing whether anything is happening. Show the steps instead, mark the one the agent is on, and say what it is doing in plain words: “Comparing 12 flights”, not “Calling search_flights”.

Let people leave. Tell them they can close the window, then let them know when the task is done or when the agent needs them.

Here is the whole loop for a short booking: a plan, the steps ticking off, a pause before payment, and a confirmation at the end.

Not every step needs permission. Sort what the agent can do by what’s at stake.

Looking around, such as searching, reading a page or opening a file, can just happen. Changes that are easy to undo, such as drafting an email or adding a reminder, can happen too, as long as undo is easy to find. Anything that costs money, can’t be undone, or speaks for the person to someone else should stop and ask: paying, deleting, sending, posting, sharing data with someone new.

Asking too often is as bad as not asking. If every step needs a tap, people learn to tap without reading, and the one confirmation that mattered gets approved with the rest. Nielsen Norman Group makes the same point about confirmation dialogs in ordinary software, and it applies twice over to an agent that can take dozens of actions in a minute.

When the agent does ask, name the action, the amount and who it affects: “Pay £184 for 1 flight to Lisbon? Charged to Visa ending 4421.” A button that says “Pay £184” is clearer than one that says “Allow”.

A Stop button should be on screen for the whole task, not hidden in a menu. Stopping should leave things in a known state and say what was finished, what wasn’t, and whether anything needs tidying up: “Stopped. I had booked the flight but not the hotel.”

People should be able to steer without starting over. If someone realises halfway through that they want an aisle seat, they should be able to say so and have the agent fold it into the rest of the task.

Offer undo for everything that can be undone, for as long as it can be undone. Where something can’t be reversed, such as an email that has gone, say so in the confirmation before it happens, not afterwards.

When the task ends, “Done!” is not enough. People need to check the work, and the report should make that quick. Start with what happened overall, list each action with the detail needed to check it, and say clearly what the agent couldn’t do.

Here is the same finished task, reported two ways. The first asks the person to take it on trust. The second lets them check it in seconds.

Agents fail in ordinary ways: a website changes, a login expires, a seat sells out between one step and the next. The failure itself is rarely the problem. The problem is an agent that retries quietly for ten minutes, or reports success when it only got halfway.

Say what happened, what the agent had already done, what is left, and what the person can do now. Keep everything it had found or entered, so finishing by hand doesn’t mean starting again. If a request is ambiguous in a way that matters, ask rather than guess.

Many agents save details between tasks, such as a home airport, a seat preference or a card. That saves time, and it can also surprise people who never knew it was kept. Say when the agent saves something about the person, and give them one place to see, change and delete what it knows.