Can AI personas replace customer interviews?

No. An LLM playing your buyer can help you prepare for a customer interview. It can't tell you why a real buyer moved, because it has never bought anything. Use AI personas to sharpen your questions. Use calls with real buyers, and what those buyers actually did, as your evidence.

Why is everyone asking this now?

Because a lot of money is betting that software can predict people.

On Oct 6, TechCrunch's Rebecca Bellan reported on Mirror Particle, a San Francisco startup building what it calls a world model of human behavior. The same piece lists the money around it: Simile raised $200 million at a $2 billion valuation, Aaru raised $88 million at a $1 billion valuation, and Humans& launched Persimmon after a $480 million seed round.

The interesting part is Mirror Particle's own argument. CEO Abhivyakti Ahuja says the common approach, prompting or fine-tuning an LLM to role-play a target demographic, is broken. LLMs model written language. Her company focuses on what TechCrunch calls revealed behavior: what people actually do, over what they say in surveys.

Mirror Particle has a commercial reason to say that about its rivals' method. The point still holds for B2B: a model trained on text knows what people write about buying, not what they did.

What does an AI persona actually give you?

The average of what people have written about a role.

Ask a chatbot to be "a VP of RevOps at a 200-person SaaS company" and you get a fluent, reasonable answer. It sounds like every LinkedIn post about RevOps, because that's what it learned from. It will say it cares about data quality, forecast accuracy, and tool sprawl. All true. All useless on a sales call, because every competitor's persona says the same thing.

What it can't give you:

• The trigger that made this buyer start looking this quarter.
• The spreadsheet workaround they built and hate.
• The name of the person who killed the last purchase.
• The exact words they use for the problem.

Those come from people who lived it.

Where does an AI persona help?

Before the call. Never instead of it.

I use persona prompts for three jobs:

1. Question prep. Have it argue with your interview script. Which questions are leading? Which ones invite a polite lie?
2. Objection practice. Rehearse the hard answers so you aren't hearing "we already have a tool for that" for the first time live.
3. Copy stress test. Paste your homepage and ask what a skeptical buyer wouldn't understand. That's a grunt test check, not proof the message sells.

Every output is a hypothesis. Mark it that way in your notes.

What counts as evidence?

What a buyer did, then what they said about it.

Mirror Particle's pet food pilot shows why. Per TechCrunch, the brand asked which imagery to put on its packaging. The answer was that imagery wasn't the problem. Shoppers saw the brand as mass market and cheap. A better answer to the wrong question would have changed nothing.

B2B has the same trap. You ask "which feature should we lead with?" when the real issue is that buyers don't believe a startup can pass their security review.

Behavior you can actually see:

• Closed-won and closed-lost deals. What changed at the account before they bought. A win/loss interview gets you the buyer's version.
• Product usage. What trial accounts did in a tool like PostHog, versus what they said they'd do. Here's how to read PostHog for GTM signals.
• How they found you. A free-text self-reported attribution field.
• Past attempts. On a call, "What did you try last time?" beats "Would you use this?"

How do I run interviews that beat a persona?

Ask about the past. Don't ask people to predict their future.

1. Recruit five buyers who bought something in your category in the last year, from you or anyone else.
2. Ask about the last time: what broke, who noticed, what they tried, what it cost, who signed.
3. Write down their words verbatim. Don't translate them into your vocabulary.
4. Look for the same trigger in three of the five. That's a pattern worth testing in outbound.
5. Then ask the AI persona to poke holes in what you heard.

"Would you pay for this?" gets a yes from nice people. "What did you pay for last time, and why?" gets a fact.

What's the risk of skipping real calls?

You build a message that sounds right to a model and lands with nobody.

I've lived the gap between sounding right and selling. One of my posts did 28,000 organic views. 4 people visited the website. Zero sales. The crowd's reaction felt like validation, and none of it was buying behavior. A persona chat is that crowd compressed into one prompt.

This is a Signals problem first. If the input is synthetic, everything downstream inherits the guess: the outbound, the page, the ads.

What should I do this week?

1. List the three beliefs your messaging rests on.
2. Next to each, write where the evidence came from: a call, a deal, usage data, or a prompt.
3. Anything that came from a prompt goes back on the list as a hypothesis.
4. Book five past-tense buyer calls in the next two weeks.
5. Use an AI persona only to tighten the script before them.

An AI persona can tell you what sounds right. Only a buyer can tell you what happened.

Start Signals, Convert, Grow

FAQ

Are synthetic users the same as AI personas?

Mostly. Both use a model to stand in for a customer. Some startups, like Mirror Particle, say they build their own behavior models from client data and other sources instead of prompting a chatbot. Treat any prediction as a hypothesis until real deals back it up.

Can I use AI to summarize customer interviews?

Yes. Summaries and tagging save hours. Keep the verbatim quotes next to the summary, because the buyer's exact words are the part you'll reuse in outbound and on your site.

How many customer interviews do I need?

Start with five past-tense calls with recent buyers. If the same trigger shows up in three of them, test it. If nothing repeats, talk to five more before you change your messaging.

When is an AI persona good enough on its own?

For low-stakes drafts, like checking whether an email is confusing. For any decision about who to target or what to say to them, get evidence from real buyers first.