Meryem Ghatrif
Meryem Ghatrif

Contents

ghatrif.meriem@gmail.com

Pensées · UX & AI

Your AI agent has a personality. Choose it.

Six steps for agent builders: understand the context, choose its personality, give it what it needs, and test the hard moments.

8 min read

You ask a website’s chatbot if you can still return a parcel. It answers “Great question! 😊”, thanks you twice, then gives you the wrong date.

I saw the same thing with my own web host’s chatbot. It began with a list of general questions, as if my answers would help it understand my problem. It sounded like listening. It was only pretending.

In both cases, the personality was not a choice but a side effect, and it showed when things got difficult. An agent rarely pretends when things are easy. It pretends when it doesn’t know (it invents), when it should say no (it agrees), and when it can’t help (it acts as if it’s listening).

Every AI agent has a personality, whether someone designed it or not. The good news is that you can choose it, and make it useful. Here is how I would do it, step by step, from my work as a designer and from the Interaction Design Foundation’s course AI for Designers.

We see faces everywhere

The heartsease is the little flower of this site. In French, it is called la pensée, a thought. Look at one closely and you will probably see a face. The flower is not looking at you. You are the one who sees someone there.

A violet heartsease from a botanical plate: two dark petals above a bright centre, like a little face.
The heartsease, Viola tricolor. Do you see the face?

We do the same with machines. In 1966, Joseph Weizenbaum created ELIZA, a very simple program that turned your sentences into questions. His own secretary asked him to leave the room so she could talk to it in private. He was struck by how quickly people confided in it.

Today’s AI agents are much better at talking. People will see a personality in yours, whether you designed one or not. So you might as well choose it.

Too nice is a problem

Being warm is good. Flattering is not.

In April 2025, OpenAI rolled back an update to ChatGPT because it praised almost everything people said. OpenAI described it as “overly flattering or agreeable”.

A second risk is being taken for a friend. The same year, a study by OpenAI and the MIT Media Lab found that the heaviest users, especially those who saw ChatGPT as a friend, tended to feel lonelier. It shows a link, not a cause.

A good agent is kind, and it can still tell you when you are wrong.

Two sides of the same agent

When you build an agent, you design two experiences.

The person’s side. Who talks to it, in what state, how it answers, how it fails, and how a human takes over. This is UX.

The agent’s side. What it is given to work with: its instructions, the real information, its tools, and what it knows at the moment it answers. Builders now call this context engineering.

The personality lives on the person’s side. But it only works if the agent’s side is right. The opening scene shows both. “Great question! 😊” is a personality problem. The wrong date is usually a context problem: the agent did not have the real return rule in front of it.

“Context” also means two things. The person’s context (in a hurry, worried) changes the tone. The agent’s context decides whether the answer is right.

Here are the six steps I would follow, from the person’s side to the agent’s side, and back.

1. Understand the context

The person’s side.

Before writing a single word, read what people really ask. Emails, contact forms, phone notes, the questions your team answers every day.

From them, make two lists. The most frequent requests. And the difficult moments: a late order, a refund it can’t give, someone upset.

Then answer one question. Who will talk to it, and how will they feel? In a hurry, worried, curious? Their state will set its tone.

These lists feed every next step. Keep them close.

2. Choose its personality

The person’s side.

A personality is a voice and a character.

The voice. Choose three words, and give each one a limit: warm, never familiar; short, never cold; confident, never pretending to be sure. Start from how the organisation already speaks, and from what its visitors need. A bank and a bookshop won’t choose the same words.

The tone. The voice stays the same. The tone follows the person: lighter for a book idea, calmer for a lost parcel.

The character. It shows in the difficult moments from step 1. Answer five questions.

  1. What does it say when it doesn’t know?
  2. How does it say no? A good “no” explains why and offers another way.
  3. How does it fix a mistake? A good correction says “I was wrong” and fixes it.
  4. When does it pass the conversation to a person, and how?
  5. What will it never pretend? To be human, to have feelings, to be certain.

The words around it. Its name, its first message, its buttons, its error messages. They speak before the model does, and they are pure UX writing. Its face speaks too: a human name and a photo of a person already pretend.

3. Give it what it needs

The agent’s side.

An agent built on a language model already has a default character: polite, helpful, often a little too eager to please. And it knows nothing about your shop. Give it five things.

  1. Its instructions. A short text the model reads before every conversation, often called the system prompt. Your answers to steps 1 and 2 go here. Chat tools, including AI plugins for WordPress, usually have a field for it.
  2. A few examples. Two or three sample conversations in its voice, including a difficult one: a refusal, an “I don’t know”. Examples often work better than adjectives.
  3. The real information. Opening hours, delivery, returns, the catalogue. Ask it to answer only from these. This is what makes an honest “I don’t know” possible.
  4. Its tools, with limits. What it may check or do, like finding an order. It helps when asked, and doesn’t jump in uninvited. Remember Clippy, the paperclip in Microsoft Office, popping up with “It looks like you’re writing a letter. Would you like help?” Microsoft turned it off by default in 2001 and removed it in 2007. And if the agent can act (book, order, send), it asks before anything that can’t be undone, then says what it did.
  5. The context of the moment. Who is writing, their order, the conversation so far. And only the data it needs, kept no longer than needed, somewhere you know. That is security, and it is sovereignty: you decide where the data lives and whether it trains someone else’s model.

For a bookshop’s website, the instructions could read:

You are the assistant of a bookshop’s website.
People write to you about orders and book ideas, often in a hurry.
Be warm, never familiar; short, never cold; confident, never pretending to be sure.
Answer only from the shop’s information and catalogue.
If the answer isn’t there, say so, and give the shop’s phone number.
To check an order, use the order tool. Never change an order yourself.
Never ask for card numbers or passwords.
Never pretend to be human or to have feelings.

Asked “Can I still return a book?”, with the shop’s return rule in front of it, it doesn’t answer “Great question! 😊”. It answers: “Yes, you can, within 14 days, with your receipt. Here is how.”

4. Put the guarantees in the interface

Both sides.

Some promises can’t depend on a model, which can forget a line or be talked out of it. Build them into the interface.

  • The AI notice, in the first message. In Europe, the AI Act has required telling people they are talking to an AI since 2 August 2026, unless it is obvious.
  • A button to reach a human, always in view.
  • A block on card numbers, before they reach the model.

The prompt asks. The interface guarantees.

If you build the chat window yourself, make it accessible. Put the messages in a container with role="log", so screen readers read each new reply aloud. If replies appear word by word, announce each one once, when it is complete. Keep the keyboard focus in the text field. And never let an emoji carry the meaning on its own.

5. Test the difficult moments

Both sides.

Write about twenty test conversations from your lists in step 1. The frequent requests, and the hard moments: a question it can’t answer, a request it must refuse, an angry message, someone trying to make it break its rules.

Run them before launch, and again after every change. For each one, check two things. Was the answer right? Did it sound like itself?

Then test it with real people. Frank Spillers, whose course on UX management I took at the Interaction Design Foundation, says who to choose: “angry, vulnerable and confused users”, not only patient ones.

His rule fits in one line: “Design the failure before the happy path.”

6. Tend it after launch

Both sides.

A pleasant voice is only the front of the service. Frank Spillers puts it bluntly: “A chatbot isn’t a service.” He notes that most AI work is designed at the interface, then breaks backstage. “Handoffs lose context, the AI is confidently wrong, and no one owns the recovery.”

Two things are best decided before launch.

  • What travels with the handoff. The person who takes over sees the whole conversation, so nobody has to start again.
  • Who owns the recovery. When the agent is wrong, someone is responsible for fixing it, and the customer knows who. In 2024, a Canadian tribunal held Air Canada responsible for what its chatbot said.

Then read the conversations every week. When it fails, fix the cause (the instructions, the information or the tools), not only the wording. A personality is tended, like a garden.

To remember

  • Every agent has a personality. Choose it on purpose.
  • Start from real messages, not from adjectives.
  • The voice stays. The tone follows the person.
  • Kind, yes. Flattering, no.
  • Give it the real information, or it will invent.
  • The prompt asks. The interface guarantees.
  • Test the difficult moments, before launch and after every change.

Look closer

Look at a heartsease long enough and you will see a face. It is not looking back.

An AI agent is the same: we are the ones who see a person. Design it so it never pretends to be one.

If you try these six steps with your own agent, I’d be glad to hear what changed.

Sources (10)