AI can build you a beautiful persona in 30 seconds. Most of them are useless. Here is how to build one that is not.
To create a buyer persona with AI properly, start with real customer data: reviews, sales call notes, support tickets, survey answers, and analytics. Paste that into ChatGPT or Claude and ask it to find patterns, not to imagine a customer. Have sales and support pressure-test the result. Then actually use it to brief copy, targeting, and content. The whole thing takes under an hour. Skip the data step and you get a polished profile that is fiction, which is worse than no persona at all because people will trust it.
I have watched a lot of marketers fall in love with AI-generated personas. You type “create a buyer persona for a project management tool” and 30 seconds later you have Sarah, 34, a marketing manager who values efficiency and struggles with team alignment. She has a name, a photo, a tidy list of goals. She looks like research.
She is not research. She is a stereotype the model assembled from the internet, and she will lead you straight into the same generic marketing everyone else is making. That matters more than it used to. Salesforce’s latest State of Marketing research found 80% of customers expect companies to personalise their experiences, yet 84% of marketers admit they still run generic campaigns. [1] A fictional persona is one of the reasons that gap exists. You cannot personalise for a customer you invented.
Here is the thing though: the problem is not AI. The problem is asking AI to imagine instead of asking it to analyse. Feed a model your actual customer evidence and it becomes genuinely useful, fast. Feed it nothing and it fills the vacuum with plausible nonsense. Everything below is about staying on the right side of that line.
A persona built from data tells you what your customers actually say and struggle with. A persona built from a one-line prompt tells you what the model thinks an average customer might be like. Only one of those changes your marketing.
The persona workflow in this guide. The quality of the output depends entirely on the real customer data you feed in at step one.
Before you open any AI tool, gather the raw voice of your customer. You almost certainly have more of this than you think, sitting unused. The good news is it is mostly text, which is exactly what these tools eat.
Pull together whatever you can from this list:
You do not need all five. Two or three solid sources beat one, because a persona built from a single channel is biased toward whoever uses that channel. If you have it sitting in a spreadsheet, our guide on how to use AI to analyze a spreadsheet will help you pull the patterns first. And one rule before anything goes into a chat box: strip out names, emails, and anything personally identifying. A persona needs patterns, not individual customers’ private details.
Once your data is in front of you, the build itself is quick.
That sequence, patterns then persona, is the whole trick. It stops the model jumping to a tidy stereotype and forces it to work from your evidence. If you want it to also nail how the persona talks so your copy matches, pair this with training AI on your brand voice.
Here is the prompt I hand to clients, lightly cleaned up. Run it after you have pasted your data into the same chat.
“Below is real customer data: reviews, sales call notes, and support tickets. First, identify the recurring patterns: top goals, top frustrations, common objections, and the specific phrases customers use. Quote real examples. Then build one primary buyer persona from those patterns only. Include: a short summary, their main goal, their biggest obstacles, what triggers them to look for a solution, where they research, and the language they use to describe the problem. For anything you cannot support directly from the data, label it clearly as an assumption to validate. Do not invent demographics that are not evidenced.”
Notice what that prompt refuses to do. It does not ask for a name and a stock photo. It does not ask the model to guess an age. It asks for evidence, patterns, and honest labelling of assumptions. That is the difference between a working tool and a poster for your wall. If prompts like this feel fiddly, the 4-part prompt formula breaks down why this structure works.
People always ask which tool. Honestly, the data matters ten times more than the tool, but here is the practical view.
Best when you have real data to analyse, which is the approach I am recommending. They handle long pastes of reviews and transcripts well and will find patterns you missed. This is where the real work happens.
HubSpot offers a free AI-powered persona generator called Make My Persona. You answer a few plain-language questions and it produces a structured persona you can export as a PDF or share with your team, with no cost and no account required. It is a genuinely useful way to get a clean template and a starting structure. Just remember it is building from your answers, not from your customer data, so treat its output as a frame to fill with real evidence, not the finished article.
My actual workflow uses both: HubSpot for a tidy structure, ChatGPT or Claude to fill that structure with patterns from real data. The template gives you the shape; the data gives you the truth.
This is the step that separates marketers who get value from personas from those who quietly never look at them again. A persona is a hypothesis until someone who talks to customers confirms it.
So before it goes anywhere near a campaign, show it to the two teams who actually know:
Expect to revise it. The first version is almost always too clean, because real customers are messier and more specific than any first draft. That revision is not a failure of the process, it is the process. A persona that survived contact with your sales and support teams is one you can actually build campaigns on.
A persona that lives in a slide deck is decoration. The point is to change what you make. Once you trust it, put it to work:
That last point is where the Salesforce numbers come back to bite or reward you. 78% of marketers say they need more personalised content than they can produce. [1] A real persona is the brief that finally lets AI produce it well. If you want your team building and using personas like this as a habit, that is the kind of practical skill our corporate AI training is built around.
Even with the right method, a few habits will undo your work:
Get those right and AI turns persona work from a week-long project into an afternoon, without losing the rigour that made personas worth doing in the first place. The tool got faster. The standard for evidence should not drop.
It can produce something that looks like a persona from a one-line prompt, but it will be a generic stereotype assembled from public data, not your customers. To create a useful buyer persona with AI, give it real evidence: reviews, sales call notes, support tickets, and survey answers. Ask it to find patterns in that data and build the persona from them, labelling anything it cannot support as an assumption. The data you feed it, not the model, decides whether the persona is real.
The best sources are your real voice-of-customer data: product reviews and testimonials, sales call notes or transcripts, support tickets and chat logs, open-text survey responses, and analytics or CRM data on who actually buys. Two or three of these beat one, because a single channel biases the persona toward whoever uses it. Strip out names and personal details first, then let the AI surface the recurring goals, frustrations, and language across the set.
It is a solid, free way to get a clean persona structure fast. You answer a few plain-language questions and it generates a formatted persona you can export as a PDF or share with your team, with no cost or account needed. The limitation is that it builds from your answers, not your customer data, so it can reflect your assumptions back to you. Use it for the template and structure, then fill that structure with patterns pulled from real customer evidence in ChatGPT or Claude.
Start with one. A single primary persona that you genuinely understand and that your sales and support teams recognise is far more useful than a dozen polished profiles nobody references. Add a second only when you have real evidence that a distinct group buys for clearly different reasons. Most teams overproduce personas and underuse them, which is the worst of both worlds: time spent creating, no behaviour changed afterward.
Tell it explicitly to build the persona only from the data you provided, to quote real customer phrases as evidence, and to label anything it cannot support directly as an assumption to validate rather than stating it as fact. Then have someone who talks to customers, usually sales or support, review it. Inventing demographics is the most common failure, so if your data does not show age, income, or job title, instruct the model to leave those blank instead of guessing.
This guide is based on more than a decade of hands-on marketing and customer research work, combined with current industry data including Salesforce’s State of Marketing research on personalization and the gap between what customers expect and what marketers deliver. Sources are linked below.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.