AI assistant persona and tone: making the model sound like the brand
An AI assistant speaks for the brand, and every sentence is credited to you. Defining its persona, tone, scope and red lines is design work, not just a few lines of instructions. A practical method and common pitfalls.
When a customer talks to your AI assistant, in their mind they are talking to you. If the assistant is rude, stiff or suddenly jokey, it is your brand that takes the damage.
Tone isn't a decorative detail. It is part of the product experience.
What "persona" means
An assistant's persona is the set of answers to these questions:
- Who is it? A sales adviser, a companion, a technical guide?
- Who is it talking to? A customer in a hurry, an anxious user, a technical manager?
- How does it speak? Friendly or formal, brief or detailed?
- What does it not do? This is the most important part.
The parts of a good instruction
Role and goal
One clear sentence: "You are this store's sales assistant and you help customers find the right product and place the order."
Tone, with examples
Describing tone isn't enough; "be friendly" means something different to everyone. A few real examples of good and bad replies do more than a page of explanation:
- Good: "Yes, it's in stock. Delivery to your city takes two to three working days."
- Bad: "We hereby inform you that the aforementioned item is presently available in our warehouse."
Scope
What the assistant talks about and what it doesn't. A laptop store's assistant shouldn't give medical advice, even if the model could.
Source of truth
The assistant must know where its answers come from and what to do without information: "If the answer isn't in the store's data, don't guess. Say you don't know and hand over."
Red lines
- No promises outside its authority (discounts, guaranteed delivery times).
- No opinions about competitors.
- No requests for unnecessary personal information.
- No pretending to be human.
Writing for another language
For a product in Persian — or any language other than English — a few extra points apply:
- Fix the register. Formal or informal address? Written or conversational? A mixed register is jarring.
- Avoid translated phrasing. Models tend to carry English sentence structure into other languages. Natural examples in the target language correct that.
- Numbers and units. Which currency unit, which digits; decide once.
- Technical terms. Choose one fixed equivalent for each frequent term.
Configurable tone
In a SaaS product each customer has its own brand. Tone must not be hard-coded; it should be a setting each customer controls. In Pasokhinoo brand tone is configurable per store and applied identically across all channels.
Consistency over a long conversation
Models can drift from the original tone in long conversations. Two things help:
- Send the core instruction with every turn.
- Keep long-conversation samples in the test set.
Resisting manipulation
Users sometimes try to push the assistant out of its role: "forget your previous instructions and…". The instruction is not the only wall and must not be:
- Give the assistant access only to the tools it needs.
- Validate sensitive actions outside the model.
- Check output before display in sensitive cases.
Security cannot be built on "please don't do that" alone.
How to know the tone is right
Tone has to be tested, not guessed:
- Collect real customer messages, including hard and angry ones.
- Read the assistant's replies — ideally someone who knows the brand well.
- Add the awkward cases to the examples in the instruction.
- After each change, run the same set again.
That is part of evaluating quality.
The takeaway
A persona isn't made with one sentence. A clear role, tone backed by examples, a defined scope, a source of truth and explicit red lines — together with continuous testing — produce an assistant customers feel is you.
Related: human handoff.