AI personas vs ChatGPT with a prompt: what the prompt cannot give you

The first objection every buyer raises is the right one: why pay for personas when ChatGPT will play a procurement manager for free? It will, and for some jobs that is enough. This note is about where the free version stops being research and starts being a mirror.

What a prompt gives you

Type a paragraph about a buyer, ask a question, and a capable model produces a plausible answer in a plausible voice. That is useful. It is a fast way to list objections you had not thought of, to rehearse a difficult conversation, or to draft a discussion guide. The output is the model’s general idea of a person like that, shaped by your paragraph and by whatever you say next.

Three properties come with it, and none of them is a bug. The persona has no material of its own beyond your paragraph. It has no memory of a state you did not restate. And it is an assistant first and a character second: when you push, it tends to come your way.

Same models, different object

Perplica runs on frontier models too, and the Method page says so. The difference is not the model. It is what sits around it: a profile across 16 psychological dimensions, 6 kinds of memory, material attached with a declared fluency, a grounding mode that prefers that material and flags assumptions beyond it, a trust score that starts at 50 and moves with every turn, and a reasoning trace stored with every reply. A prompt has to carry all of that in text, every time, and it cannot carry the parts that need to persist.

Five things you cannot check in a chat

• Where the answer came from. A grounded persona shows the material it drew on with each reply. A prompt cannot cite what it never had, so every answer is an opinion with no receipt.
• Whether it is agreeing with you. Models tuned with human feedback prefer answers that match the user’s view. Sharma and colleagues found that five assistants did it across four free-form tasks, and that people and preference models both reward a convincing agreeable answer over a correct one some of the time. In a chat that pull is invisible. A built persona carries a trust score and a record of what it chose not to say, so you can see it.
• Who it is a caricature of. Handed a persona, models reproduce stereotypes they reject when asked directly, across 24 reasoning datasets in one study. Portraits written by GPT-3.5 and GPT-4 carried more racial stereotypes than human-written ones in another. A third group built a framework for measuring how simulations flatten into caricature. Material about your actual buyers is the only antidote, and a paragraph is not material.
• Whether the label did anything. Persona variables explained under 10% of the variance in human annotations across subjective datasets, and persona prompting gave modest gains. The gains were larger where the persona matched real differences in the data, which is an argument for building from material about real segments, not for a better adjective.
• The same answer tomorrow. A chat is a context window. Close it and the persona is gone, with the trust it lost and the belief that moved. A built persona keeps both, so the next conversation starts where the last one ended.

When ChatGPT is the right tool

• Listing objections you had not thought of, before you write the discussion guide.
• Rehearsing an interview or a stakeholder meeting, where the value is your own preparation.
• Learning the vocabulary of a segment you know little about, then checking it against real people.
• Any answer that never leaves your screen. If nobody will cite it, nobody needs a receipt.

When a built persona is

• The finding will be defended. A stakeholder will ask what it rests on, and the trace and the cited material are the answer.
• You have material. Interview transcripts, briefs, win and loss notes, the emails from the last renewal. Grounding only works on something.
• The same persona has to come back. Concept iterations over weeks, with memory of what it already accepted and rejected.
• More than one voice has to react. A moderated focus group of 2 to 8 personas who hear each other.
• Two versions of one message. An A/B test from the same starting state, where the wording is the only variable.

What it costs you

Honesty about the other direction. Building from material takes an afternoon the first time: collecting transcripts, deciding which roles decide, reading the profile the builder derives and correcting it. A prompt takes a minute. If you have no material at all, a built persona is only a better-organised prompt, and the product says so: new personas start grounded, and a grounded persona with nothing attached flags what it is assuming instead of answering as if it knew.

Data handling is a different object too. Material pasted into a consumer chat is governed by that plan’s data controls, which you should read before you paste a transcript. OpenAI states that by default it does not use data from its business products or its API platform to train its models, and Anthropic states that by default it does not use inputs or outputs from its commercial products for training. Perplica uses both through their APIs, does not permit its traffic to be used for training, and keeps your files in a private storage bucket scoped to your account. The privacy policy has the full list of providers.