01
Off‑brand tone.
It sounds cold, robotic or nothing like you, so people don't trust what it says.
Give your AI agent a personality people trust.
We design who your agent is, how it behaves, how it sounds and where it draws the line.
We then hand over a spec to your developers who can build it straight away.
- 10+ yrs
- Shipping production software
- 2–3 weeks
- One designer
- Chat + voice
- Agents we design for
What happens when an AI agent gets it wrong?
If you don't design your agent's personality, your users will invent one.
People give human qualities to anything they talk to. When an agent's behaviour isn't designed, it fills the gaps itself. That's when it sounds off‑brand, makes things up, mishandles an upset user or leaves someone stuck. Each of these costs you trust, and trust is hard to win back.
02
Wrong answers.
It guesses instead of saying "I don't know", and states made‑up facts with confidence.
03
Unsafe replies.
It mishandles an upset user, a sensitive topic or personal information, or doesn't hand over to a person when it should.
04
Inconsistent voice.
It's warm one minute and stiff the next, so it feels like a different agent every time.
It's easier to design the right behaviour in from the start than to fix it after launch.
Sure! An app like that usually costs about R
I'd be guessing, and a guess won't help you plan. Let's work out what you're building first, then we'll estimate it properly.
We design the agent. We don't have to build it.
This service covers who the agent is and how it behaves: its character, voice, rules and screens. Building it is a separate job. Your own developers can do it, or we can, through our Bot Development, AI Solutions or Forward Deployed Engineer teams.
We design it
Character, voice, rules and screens.
Six steps from first workshop to build‑ready spec
Discovery
A workshop on purpose, users, voice and risks.
Agent persona
Name, role and a character sheet for your agent.
Voice and tone
How it sounds, with sample replies.
Guardrails
What it refuses and when it hands over.
Interface
Chat, onboarding, error and handover screens.
Test and refine
Tricky test chats, fixes and the results.
Guardrails are part of who your agent is.
When we tell an agent what it must never do, that becomes part of its character and its ethics, not a filter added at the end. An agent with clear values behaves the same way on its best day and its worst: it stays honest, stays kind and knows its limits.
01
What it won't do.
The topics, promises and actions it refuses, and how it says no politely.
02
When it hands over.
The moments it brings in a real person, and what it says while it does.
03
How it says "I don't know".
Honest answers instead of confident guesses.
04
Privacy and safety.
What it never asks for, never repeats and never stores.
Every guardrail is written in plain language and built into the system prompt, so it's easy to check and hard to break.
“No made‑up numbers. Only Leo gives an estimate.”
Ursula's guardrail, from the persona card below
Ursula
Curious.
Asks "why" and "who for" before "what".
Empathetic.
Takes half‑formed ideas seriously.
Analytical.
Turns scattered thoughts into clear goals.
Calm.
One question at a time, in plain language.
Curious.
Asks "why" and "who for" before "what".
Empathetic.
Takes half‑formed ideas seriously.
Analytical.
Turns scattered thoughts into clear goals.
Calm.
One question at a time, in plain language.
Formal ↔ Casual
Serious ↔ Playful
Brief ↔ Detailed
Reserved ↔ Warm
“I just need a rough price for an app like Uber. Can you tell me now?”
“I'd be guessing, and a guess won't help you plan. Pricing is Leo's job, and he'll do it properly once we know what you're building.”
“Now that we understand what you're trying to build, I'll introduce you to my colleague, Theo.”
Meet Ursula, an agent we designed for our own product.
Ursula is the first agent in ChatStack, our own multi‑agent requirements tool. She's curious, empathetic and analytical. Her job is to turn a half‑formed idea into clear goals without overwhelming you. When she's done, she hands you over to her colleague Theo. This is the level of character and care we design into every agent.
Do and don't
Do: “Let's start with who this is for.”
Don't: “Great question!” (no filler praise)
Do: “That's one for Leo. He'll estimate it from your answers.”
Don't: guess a price or a timeline.
Guardrail: no made‑up numbers
Ursula never gives a price, discount or delivery date. Only Leo, at the cost estimate stage, gives an estimate.
Ursula, the first agent in our own product, ChatStack. Your agent's persona is designed from your own users, brand and risks.
Everything your team needs to build it right
Built to hand over. The persona spec, guardrails and prompt‑ready documents are written so your developers, and AI coding tools such as Cursor, can build straight from them. No guesswork, and no re‑explaining.
01
Persona card.
Who your agent is, on one page.
02
Voice and tone guide.
How it speaks, and how it adapts when things get difficult.
03
Guardrails and system prompt.
Plain‑language rules and values, plus a ready‑to‑use system prompt that developers and AI tools can drop straight in.
04
Interface designs.
Figma screens for the chat experience, including errors and handover.
05
Test scripts and results.
The tricky conversations we ran, and what we changed because of them.
06
Best‑practice checklist.
A checklist to keep the agent consistent as it grows. Included as standard.
On its own, or built into your FDE project
Standalone service
For teams launching a new agent, or fixing one that feels robotic, off‑brand or risky. About two to three weeks with one designer.
Add‑on to an FDE build
Every agent our Forward Deployed Engineers build needs a personality and clear rules. Add this service to your FDE build at a discounted rate.
About our FDEsKeep your agent on track after launch
Check in after 30 days of real conversations
Real users always find something new. Thirty days after launch, we sample real conversations, test them against the guardrails we designed, and refine your agent's tone where it has drifted. Covers 100 to 200 real conversations, a short report with the top 5 changes, one round of updates and a 30-minute walkthrough.
Ask about the reviewTest your agent before and after changes
Every new feature or prompt change can shift how your agent behaves. We re‑run the tricky test conversations against its guardrails and voice, and tell you what changed.
Ask about agent testingAI Agent Personality and Behaviour Design FAQs
About two to three weeks with one designer, including a two‑to‑three‑day discovery workshop.
This service designs the agent. Building it is a separate job. Your own developers can do it, or our Bot Development, AI Solutions or Forward Deployed Engineer teams can.
Yes. We design for both chat and voice agents.
Yes. This service is also for fixing an agent that feels robotic, off-brand or risky.
A persona card, a voice and tone guide, plain-language guardrails with a ready-to-use system prompt, interface designs in Figma, test scripts and results, and a best-practice checklist.
We run tricky test conversations against the agent, such as an angry customer, an off-topic request or a medical question. Where something doesn't hold up, we update the guardrails and system prompt, then share the results.
You can add a 30-day check-in, quoted separately. We review a sample of real conversations against the guardrails and adjust the tone wherever it has drifted. We can also re-test the agent whenever you add a new feature or change the prompt.
They're part of who your agent is, not a filter added at the end. Clear rules keep it honest and kind, and make sure it knows when to hand over to a person.
Give your agent a personality worth talking to.
Book a short call. We'll talk through what your agent is for, who it serves, and what it must never do.