Building an AI customer persona for sales simulation: 8 fields

Short answer

A realistic AI customer profile is built from eight fields: customer type, their goal, their behaviour, their objections, what they know and do not know, the conversation language, when the conversation ends, and a description of a good conversation with this type of customer. The first seven define how the AI plays the customer; the eighth defines what the employee is scored against.

The most common mistake is writing the profile as one sentence: "a customer who objects to the price". With that, the AI either plays a generic "difficult customer" or behaves differently every time. Below is what to write in each field, plus a filled-in template.

Persona versus practice profile

In marketing, a buyer persona is a general portrait of the target audience: age, role, interests, channels. A practice profile is a role written for one conversation. It does not need age or hobbies; it needs this: what the person wants on the call, what worries them, what they will ask, which answer they will accept and which answer will make them end the conversation.

In other words, a profile is a description of behaviour. You can start from your persona document, but take only the parts that affect the conversation.

Eight fields

1. 11. Customer typeWho they are and what situation they are in: "a young family buying their first flat", "a restaurant owner opening a second branch". One sentence, but specific.
2. 22. GoalWhat they want from this conversation: to learn the price, compare two offers, buy time, reduce risk.
3. 33. BehaviourHow they talk: short and rushed, suspicious, friendly but undecided, keen on technical detail. Write down tone and pace.
4. 44. ObjectionsWhich objections they raise and in what order. Write the real reason behind each — for example, behind "too expensive", "I don't understand what I'm paying for".
5. 55. KnowledgeWhat the customer knows (a competitor's price, a past experience) and does not know. The AI customer should not invent facts it has not been given.
6. 66. LanguageWhich language the conversation is in and whether the customer mixes languages.
7. 77. End conditionWhen they agree (to a meeting, a second call, an order) and when they politely end the conversation. Without this, the AI either gives in too quickly or objects forever.
8. 88. Good-conversation standardWhat a good conversation with this type looks like: which questions to ask, which facts to state, how to end. Scoring relies on exactly this.

A filled-in template

The profile below is illustrative and does not describe a real customer.

• Type: director of a 40-person accounting firm, replacing office furniture.
• Goal: compare three suppliers' offers and get delivery during the summer months.
• Behaviour: speaks briefly, short on time, wants numbers, dislikes general praise.
• Objections: 1) "another firm quoted 15% less" (reason: does not see the quality difference); 2) "installation takes too long" (reason: afraid work will stop).
• Knowledge: knows the competitor's price; does not know your warranty terms.
• Language: the local language, sometimes with technical terms in another.
• End condition: agrees to a meeting to see samples if they hear that installation can be done at the weekend and about the warranty; if only a discount is offered, says "I'll think about it" and ends.
• Standard: first ask about office size and time constraints; explain the difference through materials and warranty; propose an installation plan; end with a dated meeting.

Writing objections in layers

Real customers do not say all their objections at once. Write objections into the profile in sequence: the first is surface-level ("too expensive"), the second reveals the real concern ("the last one we bought broke quickly"). If the employee does not ask a clarifying question at the first objection, they never reach the second — and that is exactly where the value of practice lies. Writing the objection scenario is a separate topic; the profile is its foundation. The practice routine shows how profiles connect to weekly topics.

How many profiles you need

Three or four profiles are enough to start: the customer you meet most, the customer you lose most, and one or two hard cases. Test every new profile with one question: "What does this teach that the other profiles do not?" If there is no answer, you do not need a new profile — raising the difficulty of an existing one is enough.

• Simple profile: for new hires, one objection, friendly behaviour.
• Medium profile: two layers of objection, a customer who wants information before deciding.
• Hard profile: a competitor's offer, time pressure, an unclear decision-maker.

How to test a profile

1. 1Play it yourselfLet the profile owner hold two conversations: one good, one bad. Does the AI customer react differently?
2. 2Check the factsDoes the AI customer invent facts not in the profile — getting your price wrong, for example?
3. 3Check the end conditionDoes it agree after a conversation that meets the standard, and not after one that does not?
4. 4Give it to an experienced repAsk them: "Do real customers talk like this?"

Common mistakes

• Writing the profile as one sentence.
• Leaving out the reason behind each objection — the AI just repeats the words.
• Setting no end condition — the conversation is either too easy or endless.
• Copying a real customer's name, phone number or conversation into the profile.
• Opening the profile for scoring before the standard is written.

Limitations

Even the best-written profile does not reproduce all of a real person's unpredictability: the AI customer relies on the profile, while a real customer arrives with their mood, their day and their history. However precise the profile, the AI can sometimes step out of role — so test a new profile before giving it to the team, and take employees' "that's not real" feedback seriously.

Profiles in Vexvon AI Training

In Vexvon AI Training the customer profile fields are exactly these: name and description, customer type, language, customer behaviour, an objection list (up to 50), sales notes, and a "good call script" — the company's standard for that type. A test call is scored only if the profile has a standard. If a company has no profiles yet, the system can generate one with AI from the company's own prompt and knowledge base; when nothing is known about the company, it does not invent one.

Next step

Fill in the eight fields for the customer type you lose most and play the profile yourself twice. More articles are in customer simulation and scenarios; to build your profile together, get in touch.

Further reading on this topic: objection handling role play, indecisive buyer role play, sales enablement knowledge base.