TL;DR. An AI persona becomes believable through the way it behaves in a conversation, not through the amount of background written about it. It needs a goal of its own, resistance at the right moments and a consistent position throughout the exchange. Personas that agree too easily make practice meaningless. Personas that never move are hardly any better. This article looks at the five elements a useful persona needs, how to set the right level of difficulty and who inside a company should build them.
Key Takeaways
- SUXXEED's persona library grew from around 70 AI buyer personas to more than 500, created by team leads without a development team or central administrator.
- Effective practice sits just beyond a learner's current ability, a defining property of deliberate practice (Ericsson, Krampe & Tesch-Römer, 1993).
- Participants at Würth rate Sleak an average of 4.76 out of 5. Their feedback frequently highlights the realism of the conversation itself.
- Personas and conversation types are versioned, with draft and published snapshots, and can be shared by team (platform capabilities C1 to C7 and D2c).
- Language plays a major role in credibility. At FEGA & Schmitt, using the company's own terminology helped the conversations feel like they belonged inside the business.
Many virtual counterparts have the same problem: they are too nice.
A procurement lead gives in after the second argument. An employee accepts criticism immediately. A customer agrees to the price without really challenging it. The conversation feels smooth, but that is exactly the problem. If you practise ten times with a counterpart that barely resists, you do not learn very much. You mostly confirm what already works.
The quality of an AI persona is therefore not a minor product detail. It shapes the quality of the exercise. A good virtual counterpart does roughly what a good sparring partner does: it creates pressure where pressure is useful, without becoming difficult for the sake of it.
That does not require pages of backstory. What matters is knowing what the person wants from the conversation, where they will hold their ground and what would make them change their position.
What makes an AI persona believable?
A convincing persona has an objective of its own, and that objective is usually different from the learner's.
A buyer wants the price reduced. An employee wants their side of the story to be heard. A customer wants to delay a decision. The tension in the conversation comes from that difference in interests.
Three qualities matter most.
First, the persona needs its own agenda. It should not simply follow the learner through the conversation.
Second, it needs resistance in the right places. Real people are rarely opposed to everything. They tend to become firm around specific issues. A useful persona should behave the same way.
Third, it needs to stay consistent. If a persona makes one claim in minute two and takes the opposite position in minute nine for no clear reason, the conversation stops feeling credible very quickly.
At Sleak, this kind of persona sits inside a Training Scenario. In Training Mode, the employee speaks with the virtual counterpart and then receives an evaluation from the AI Coach. The feedback points to specific moments in the transcript rather than offering general praise. Personas are versioned, can be tested as drafts and then published, and can be shared with specific teams. The practice itself runs through voice-based AI role plays.
Why do AI personas become too agreeable?
Language models tend to be cooperative. If a persona is only loosely described, it often turns into a polite counterpart that follows the learner's lead and accepts arguments too readily.
The obvious fix is to define a goal and specific points of resistance. But teams sometimes overcorrect and create the opposite problem: a persona that refuses every argument and never gives ground.
That is not useful practice either. The learner needs to see which behaviour has an effect. There has to be a clear reward path, meaning the conditions under which the persona is willing to move.
This is one of the easiest parts of persona design to miss. Experienced trainers often handle it instinctively during a live role play. With an AI persona, it has to be made explicit.
Language matters too. A persona can be factually accurate and still feel artificial if it uses words nobody in that company or industry would naturally use.
At FEGA & Schmitt, this turned out to be an important part of the experience. Martina Skibowski, Head of HR, People & Culture, described it this way: "What surprised me was how quickly Sleak picked up our terminology. After only a few short briefings, it felt like speaking with someone from our own company."
Five components are enough
A robust persona can be built around five elements: role, goal, resistance, reward path and language.
| Component | The question it answers | Common mistake |
|---|---|---|
| Role | Who is this person, and what can they decide? | The title is clear, but their actual authority is not. |
| Goal | What do they want from this conversation? | They have no objective of their own, so they simply follow the learner. |
| Resistance | Where do they hold firm, and why? | They resist everything, or nothing, without a clear reason. |
| Reward path | What would make them give ground? | It is never defined what a successful argument or behaviour should achieve. |
| Language | Which terms do they use, and which do they avoid? | The persona speaks in generic marketing language instead of the language of the trade. |
These five points matter more than an elaborate biography.
It is easy to start with age, hobbies, family background and a detailed career history. That may produce a rich character sheet, but it does not necessarily produce a useful conversational partner. For training, the more important questions are whether the person wants something, whether their reactions make sense and whether there are clear conditions under which they will reconsider their position.
That is also why every persona should be tested in a real conversation before it is published. On paper, it is surprisingly difficult to tell whether the five components actually work together.
How difficult should an AI persona be?
A useful training persona should be slightly harder than what the learner can already handle comfortably. It should stretch them, but not overwhelm them.
That follows the basic logic of deliberate practice: focused improvement happens when the task sits just beyond current ability (Ericsson, Krampe & Tesch-Römer, 1993).
In practice, that usually means creating several levels for the same situation rather than one universal persona.
A new sales rep might first practise a price objection with a buyer who accepts a solid factual response. Later, the same rep can face a buyer who pushes back several times and introduces a competitor's price on the third objection.
This kind of progression does more than make the training harder. It also helps identify where someone starts to struggle.
A person may handle the first objection well but lose confidence once the buyer challenges them repeatedly. In that case, the problem may not be argument quality at all. It may be the ability to stay composed under pressure and keep the conversation on track. A single persona with one fixed level of difficulty rarely shows that distinction clearly.
Personas that are too difficult create a different problem. They are often built around the idea of preparing people for the worst possible case from the start. That may sound rigorous, but it can reduce the number of repetitions. Someone who feels they had no chance in the first attempt is less likely to want another one.
A useful rule of thumb for an entry-level persona is simple: an average prepared learner should be able to succeed on the second attempt. Not immediately, but not only after five failed runs either.
How do you test a persona before publishing it?
A good test does not begin with your strongest argument. It begins with a deliberately weak one.
In the first pass, argue badly on purpose. If the persona still gives in quickly, the resistance is too weak.
In the second pass, use the strongest argument available. If the persona still refuses to move at all, the reward path is probably missing or too narrow.
The third pass is about consistency. Pick up something the persona said earlier in the conversation and return to it a few minutes later. If the persona contradicts itself for no clear reason, the solution is usually not more biography. More often, the goal, position or boundaries need to be stated more clearly.
There is also a fourth check outside the conversation: does the persona actually trigger the behaviour the Scorecard is designed to assess?
If a Scorecard evaluates how well someone handles a price objection but the persona never raises the issue of price, the exercise is misaligned. Scores may then come out low even though the learner never had a meaningful opportunity to demonstrate the required behaviour.
Versioning matters here. At Sleak, personas are maintained as drafts and published snapshots, so changes can be tested before they affect an ongoing practice series.
Who should build the personas?
The people closest to the real conversations usually know them best.
A sales team lead knows which objections a particular customer segment actually raises. They know the wording, the timing and the points where customers genuinely become difficult. Those details are easy to lose when persona creation is handed entirely to a central content team.
At SUXXEED, the persona library grew from around 70 AI buyer personas to more than 500. Team leads created them themselves, without a development team or central administrator. Employees and candidates have since completed almost 15,000 simulated conversations, and objection handling scores rose by more than 42.4%.
For sales organizations, that suggests a useful division of responsibility. Evaluation standards can be defined centrally. The actual counterparts should be created as close as possible to the teams that know those conversations from daily work.
More realistic does not mean harder
Realism and difficulty are easy to confuse in persona design.
A realistic persona reacts in a way that fits the situation, including agreeing when a good argument deserves agreement. A hard persona simply agrees less often.
A counterpart that never yields is therefore not automatically more realistic. In fact, it can feel just as artificial as one that accepts everything.
There are cases where a deliberately difficult persona makes sense, particularly in assessment. But the purpose is different. Practice is meant to develop behaviour. Assessment is meant to establish whether a standard has already been reached.
When those two purposes are mixed into one persona, the line becomes blurred. Practice needs an achievable reward path. Assessment can set a much higher bar.
There are also situations where a human sparring partner remains the better option. That is especially true when the challenge depends heavily on a specific relationship, for example a conflict with one particular colleague. An AI persona can model a behaviour pattern. It cannot reliably predict how that exact person will respond in the real conversation.
What participants notice about good personas
At Würth, participants rate Sleak an average of 4.76 out of 5. What stands out in the feedback is how often people talk about the conversation itself rather than the technology behind it.
One sales representative put it simply: "I really liked the highly realistic conversation flow with different customers."
The phrase "different customers" is worth noticing. What the participant highlights is not one unusually detailed persona, but the range of counterparts. That fits the broader lesson from persona design: credibility often comes from variety across several different goals, objections and reaction patterns, not just from adding more depth to one character.
Another participant from the same rollout described feeling more confident in real conversations. Someone else compared the experience to a computer game.
Those reactions are not contradictory. Good practice needs to feel serious enough to matter and light enough to repeat.
FAQ
How many personas does a team need to start?
Two or three per conversation type are usually enough at the beginning, ideally at different levels of difficulty. The important point is not the number itself. The personas should differ in goal, resistance and behaviour, not just in name or industry.
How detailed does an AI persona's background need to be?
Usually less detailed than people expect. Role, goal, points of resistance, reward path and vocabulary carry most of the conversation. Extra biographical detail can add flavour, but it often contributes less than a well-run test conversation.
Can an AI persona represent a specific real customer?
It can represent a behavioural pattern, not predict a specific person. When preparing for one particular meeting, the persona is best used as a model of plausible reactions rather than a digital copy of the real customer.
How do you keep a persona stable over time?
Through versioning. At Sleak, personas are maintained as drafts and published snapshots, so edits do not automatically affect an ongoing practice series. If scores are meant to be compared over time, the persona should not be changed halfway through that series.
