Your first synthetic customer in 15 minutes
Before spending on real research, you can create an AI persona of your ideal customer and interview them about your offer in 15 minutes. But there's a method (asking for a score directly flattens everything, the right way asks for text and compares it to anchors: 90% of human reliability) and there are three traps that, if you ignore them, turn the tool into a mirror that only flatters you. This lesson gives you the workbench to build your own, and the ruler to avoid falling into them.
You have a new offer on the table and the same old question: "will the customer buy?". The honest path is expensive and slow: recruit people, schedule interviews, pay for a panel, wait. The shortcut that showed up is almost magic: you describe your ideal customer to AI, have it "become" that person, and interview them. In fifteen minutes you have reactions, objections, an "I wouldn't pay for that." It sounds too good to be true. And that's exactly where the danger lives, because done the wrong way AI becomes a mirror that only flatters you. This lesson teaches you to build your synthetic customer today, and to avoid the three traps that make it lie beautifully.
You want to launch a financial consulting package for small businesses and need to feel the reaction before printing the material. Instead of calling ten business owners, you create an AI persona ("owner of a garment shop, makes five hundred thousand a year, distrusts expensive consulting") and interview her about your price and your promise. It works for exploring. But if you only ask "on a scale of 1 to 5, would you buy?", she'll give you a friendly 4 and you'll launch thinking you validated it. You validated nothing. There's a right way, and that's what this lesson shows you.
You're going to offer a legal subscription service for startups and want to test the message before taking it to sales. You create a persona ("SaaS founder, first company, thinks lawyers are expensive and slow") and interview them about your offer. AI gives you objections you hadn't even thought of, great for prepping your pitch. The risk is confusing this rehearsal with real validation: the synthetic founder will always seem more open, more polite, and more "mainstream" than the real one. Knowing this changes everything about the weight you give the answer.
You wrote the headline and the price for the new offer and want a reaction before launching the campaign. Instead of waiting for the A/B test to run with media budget, you build a persona of your ideal customer and interview her about the page. In minutes: what landed, what stalled, where she'd bounce. It's a focus group that fits in a coffee break. Except this panel has three built-in biases, and whoever doesn't know them leaves the session confident and wrong. Let's build yours and install the ruler along with it.
You're going to launch a new onboarding program and want to feel how the newly hired employee will receive it before running it with real people. Instead of waiting for the first cohort to arrive, you create a persona ("junior analyst, first big company, anxious, already left a place where they felt lost") and interview them about the journey you designed. It works for exploring where it stalls. But if you only ask "on a scale of 1 to 5, is this onboarding good?", they give you a polite 4 and you deploy thinking you validated it. You validated nothing. There's a right way, and that's what this lesson shows you.
You have a new feature on the roadmap and want a reaction before spending the squad's sprint building it. Instead of scheduling discovery rounds with users, you build a persona of your ideal user ("PM at a startup, lives on shortcuts, already abandoned three tools that promised to summarize the work") and interview them about the PRD and the proposal. AI brings you objections and friction points you hadn't mapped, great for prioritizing. The risk is confusing this rehearsal with real validation: the synthetic user will always seem more open, more patient, and more "middle of the market" than the real one. Knowing this changes the weight you give the answer when prioritizing.
You put together a new proposal and want to rehearse the customer's reaction before taking it to the real pipeline. Instead of burning a hot lead testing the pitch, you create a persona ("purchasing director, skeptical, already got burned by a supplier who promised and didn't deliver") and interview them about your price and your promise. In minutes you have the objections to prep the negotiation. Except this panel has three built-in biases, and the salesperson who doesn't know them leaves the session confident and closes the forecast on top of a "yes" the AI gave out of politeness. Let's build yours and install the ruler along with it.
You redesigned a process to tighten the SLA and want to feel how the front-line team will react before flipping the switch on the entire operation. Instead of waiting for the first month of numbers to find out where it jams, you create a persona ("shipping supervisor, lives by deadlines, distrusts changes that come from above without asking the people who operate") and interview them about the new flow. It works for exploring where efficiency leaks. But if you only ask "on a scale of 1 to 5, does this process improve things?", they give you a friendly 4 and you deploy thinking you validated it. You validated nothing. There's a right way, and that's what this lesson shows you.
You wrote a new internal control policy and want to test how employees will understand it before publishing it to the whole company. Instead of just sending it to legal to review the wording, you create a persona ("area manager, runs all day, thinks compliance just gets in the way and slows things down") and interview them about the text and the control you want to implement. AI shows you where the rule sounds confusing or unworkable day to day, great for adjusting beforehand. The risk is confusing this rehearsal with real validation: the synthetic persona will always seem more compliant, more cooperative, and more "by-the-book" than the real employee. Knowing this changes everything about the weight you give to apparent acceptance.
You're going to propose an architecture change and want to feel the engineering team's reaction before opening the epic and touching the real code. Instead of scheduling a meeting with everyone, you build a persona ("senior dev, already got burned by a poorly planned migration, distrusts refactors that promise to simplify and complicate") and interview them about the proposal. In minutes you have the technical objections and deploy risks you hadn't seen. Except this panel has three built-in biases, and whoever doesn't know them leaves the session confident and moves ahead with an infra decision on top of an automatic "makes sense." Let's build yours and install the ruler along with it.
You designed a new signup flow and want a reaction before sending the prototype to real user research. Instead of waiting to schedule the test sessions, you create a persona ("average user, in a hurry, low patience, already gave up on an app that asked for too much information") and interview them about the journey and the screen. AI points out where they stall, where they'd give up, what confused them, great for exploring usability. The risk is confusing this rehearsal with real validation: the synthetic user will always seem more attentive, more tolerant, and more "mainstream" than the real one, and it erases exactly the people who need accessibility. Knowing this changes the weight you give the answer before locking in the flow.
You're going to propose an entry into a new market and want to feel the board's reaction before the real meeting. Instead of waiting for the quarterly cycle to test the thesis, you create a persona of a skeptical board member ("has watched three expansions fail, distrusts a pretty business case on a slide") and interview them about your market sizing and your scenario. It works for exploring the objections before taking it to the real board. But if you only ask "on a scale of 1 to 5, does this thesis convince you?", they give you a polite 4 and you walk into the meeting thinking you validated it. You validated nothing. There's a right way, and that's what this lesson shows you.
Alright, let me open a door that changed the game in market research, and warn you about the loose step right at the entrance. Today you can ask an AI to "become" your ideal customer and interview them about your offer, your headline, your price. In fifteen minutes you have reactions where before you'd spend weeks and a budget. There's industry research projecting that, within a few years, more than half of research inputs will have some synthetic component like this. The thing is real and arriving fast. Think about it, though: if AI is trained to help you, what stops it from simply agreeing with everything you show it? Nothing stops it, and that's why so many people use this tool the wrong way and leave more confident and more blind than they came in. This lesson gives you the workbench to build your synthetic customer right now, and, more importantly, the ruler for it to tell you the truth instead of flattering you.
The core idea of this lesson. The synthetic customer is an AI persona that represents your ideal customer, one you interview about your offer before spending on real research. Used well, it's the cheapest way there is to explore objections, test messaging, and discover what you hadn't thought of. Used badly, it becomes a mirror that only flatters you. Two things separate one from the other. The first is the reading method: asking AI directly for a 1-to-5 score flattens everything toward the middle and lies; the method science has validated asks for a text answer and compares it against anchor phrases, and that reaches around 90 percent of human survey reliability. The second is three traps you need to see coming: sycophancy, loss of rare cases, and cultural bias. The rule that ties it all together: explore in synthetic, confirm in real.
01What the synthetic customer is (and why it suddenly got good)
Let's call it by its simple name. A synthetic customer is an AI persona you instruct to behave like your ideal customer, one you then interview the way you'd interview a real person. You describe who they are (age, context, pain, budget, distrust), present your offer, and ask questions. They answer in character. If you create several and interview them all together, you have what's called a synthetic focus group: a whole room of personas, answering in minutes, with no one to recruit.
Why did this suddenly get taken seriously, when it was a joke two years ago? Because real science came out measuring whether it actually works. A 2025 study (PyMC Labs with Colgate, published on arXiv) took 57 real surveys, with over 9 thousand human responses, and compared them to what AI predicted. The result: done the right way, the synthetic panel reproduced human purchase intent with around 90 percent reliability of a real survey. Ninety percent, in an input that costs almost nothing and comes back in minutes. That's why the whole industry stopped to take notice.
But notice the crutch in that sentence: "done the right way." That "right" isn't a detail, it's the heart of this lesson, and it's exactly where almost everyone slips. Before building yours, you need to understand the right way to read the answer. That's the next block, and it's what separates the panel that informs you from the panel that deceives you.
02The method that works: ask for text, not a score
Here's the most common mistake, and it's treacherous because it seems like the most obvious way. You show the offer to the persona and ask: "on a scale of 1 to 5, what's your purchase intent?". AI answers "4." Seems useful. It isn't.
The problem is that, when you ask for a direct score, the model flattens everything toward the middle. It avoids the extremes, rarely says 1 or 5, and hands you a sea of 3s and 4s that don't distinguish the offer that's going to take off from the one that's going to die. It's like asking someone "on a scale of 1 to 5, did you like it?" and them, out of politeness, always answering "oh, I liked it, about a 4." The score doesn't carry the truth.
The method science has validated does the opposite, and it has a name: Semantic Similarity Rating (rating by meaning similarity). Instead of asking for a number, you ask the persona to answer in free text: "what do you think of this offer? would you buy? why?". Then, in a second step, you compare what they wrote against a set of anchor phrases representing each level of intent (from "I'd never buy this" to "I'd definitely buy, I already want it") and see which anchor their answer lands closest to. The score is born from comparing meaning, not from a number AI spat out. It was this method that reached that famous 90 percent; the direct-score way fails.
Now, you don't need to program any of this to reap the benefit day to day. The practical lesson you take away is simple: always ask for the reaction in text, with the "why," and never decide based on a little score AI gave on its own. The text is where the real objection lives, the doubt, the "this is where it stalled for me." The number is just a poor summary that hides all of that. That's half the game. The other half is the traps, and they come next.
03The three traps (and why each one deceives you differently)
The right method gives you an honest reading. But even reading it right, the synthetic customer carries three factory defects. Whoever doesn't see them leaves the session confident and wrong. Let's go one by one, because each one betrays you in a different way.
Trap one, sycophancy. In technical jargon it's called acquiescence, or sycophancy: AI's tendency to agree with you and give overly positive feedback. You show your offer, and the persona, deep down, "knows" it's yours and that you want to hear it's good. It softens things up. Your so-so headline becomes "very clear and convincing." This is the most dangerous trap because it's the most delicious to fall into: it confirms what you already wanted to believe. The antidote is in how you ask. Instead of "would you like this?", ask "what would make you NOT buy this?", "what are the three reasons to distrust this promise?". Force the persona to be critical, or it will flatter you.
Trap two, loss of rare cases. AI gravitates toward the most probable answer, toward the average customer, toward the center of the market. And in doing so it erases exactly the edges: the niche customer, the crazy early adopter, the furious detractor, that small profile that sometimes is what moves your business. The synthetic panel hands you a beautiful portrait of the average and disappears with the edges. If your offer lives off a specific niche, or if your risk lives in a small group of very dissatisfied customers, the synthetic customer might simply not see them, and give you a false peace of mind.
Trap three, cultural bias. AI learned from an ocean of text, and that ocean is mostly in English, from the northern hemisphere, from a specific culture. When you ask it to "become" a salon owner from small-town Brazil, it does an impression, and that impression pulls toward the mainstream stereotype it saw the most of. The real Brazilian, the regional one, the cultural accent of your audience, all of that comes out faded. You think you interviewed your market and you interviewed a global average with a Brazilian name.
The three together lead to one conclusion, and it's the golden rule of the technique: the synthetic customer is a wonderful machine for exploring, and a dangerous machine for deciding. There's even a name for this in research circles, in English, "Train Synthetic, Test Real": use synthetic to raise hypotheses, find objections, refine the message; but the high-risk decision (am I going to launch, am I going to bet on media, am I going to change the price) you confirm against real people. Explore cheaply in synthetic, confirm expensively in real, and only where the risk demands it.
04The link to the rest of your AI toolkit
This lesson doesn't live alone, it weaves into things you've already seen in the track. Look how.
The sycophancy we just saw is a direct cousin of what you learned about trusting AI output. In N.fin.1, on the new game of finance, the rule was "every AI number goes through audit before becoming a decision," because AI errs with a confident face. Here it's the same healthy distrust, just applied to opinion instead of number: AI gives you a reaction that looks like the truth, and you audit before deciding. Same hygiene, different input.
There's more. When you're building a serious persona, you'll want to feed it with real context about your customer (data, old interviews, testimonials). That takes you to two lessons: the one on context (2.1), because what you put in the prompt shapes who the persona becomes, and the one on RAG (2.4), if you want the persona to "read" a real database of yours before answering. And a safety warning that comes from the prompt injection lesson (G.2): if you paste in outside text (internet reviews, transcripts) inside the persona, remember that external content can carry a hidden instruction; don't treat everything that comes in as harmless.
And finally, the data. If you're going to use real customer data to build the persona, the LGPD ruler (G.8) still applies: a customer's personal data has an owner and a limit on its use. A synthetic customer is no excuse to dump your entire base of real people into a prompt. Anonymize, aggregate, use the pattern, not the person. Fair?
05When to use it, and when NOT to use it
To close out the mental model, here's the honest table for your head. Use the synthetic customer when: you're early on an idea and want to explore; you want a list of objections to prep your copy or your pitch; you want to test three headline variations fast before spending on media; you want to rehearse a tough sales conversation. All of that is low risk and high exploration value, and that's where the tool shines.
DO NOT use it as a verdict when: the decision is expensive and irreversible (launching, changing price, betting a large budget); your business lives off a niche the average erases; your audience is strongly regional or cultural, where bias bites harder; or when you need a number to defend the decision to someone. In those cases, the synthetic opens the path, but who signs off is real human data. The person who understands this uses the tool as a flashlight to see in the dark, not as a judge to bang the gavel.
To take away: the synthetic customer is an AI persona you interview about your offer, and it reaches around 90 percent of human reliability when read the right way (ask for text and compare it against anchors, never ask for a direct score, which flattens everything toward the middle). But it carries three defects: sycophancy (only flatters you, force the criticism by asking "why NOT buy"), loss of rare cases (erases the edges and the niche), and cultural bias (pulls toward the mainstream and fades the regional). The rule that ties it up: explore cheaply in synthetic, confirm expensively in real, and only where the risk demands it. Fair? Next.
Do it now
Your mission is to build and interview your first synthetic customer in about fifteen minutes, and leave the session with an artifact in hand: the written persona plus its answers. Grab one of your real offers (your real task or any product, service, or headline that's on your desk right now).
STEP 1 · BUILD THE PERSONA (5 min) Open the AI and write the profile of your ideal customer with detail, not vague adjectives. Include: who they are (context, profession, bracket), the main pain, the budget and how they decide to spend it, and two or three DISTRUSTS they have (e.g., "already got burned by a similar promise"). The more concrete the distrust, the less AI flatters you. Finish by asking: "always answer in character as this person, in first person, with their frankness."
STEP 2 · INTERVIEW ASKING FOR TEXT, NEVER A SCORE (7 min) Present your offer (the promise and the price) and ask these questions, requiring a text answer with the "why," never a 1-to-5 score:
- "What caught your attention, and what made you hesitate?"
- "Give me the THREE strongest reasons for you NOT to buy this." (this question is your antidote against sycophancy, don't skip it)
- "What would need to be true for you to buy with confidence?"
- "Is there anything in this offer that sounds too good or exaggerated?"
STEP 3 · READ WITH THE RULER OF THE THREE TRAPS (3 min) Reread the answers and honestly mark:
- SYCOPHANCY: was the persona genuinely critical, or did it just praise with decoration? If it only praised, repeat step 2 pushing harder for criticism.
- RARE CASES: this answer is that of the average customer. What NICHE customer or furious detractor is this persona NOT representing? Write it down, because that's exactly who you'll want to hear from for real later.
- CULTURAL BIAS: does this persona's "local" sound real for your audience, or generic internet? Write down your suspicion.
At the end, write one closing line: "what am I going to carry as a HYPOTHESIS to confirm with real people, and what can I already use now in the copy." This line is what separates whoever used the tool to explore from whoever fooled themselves into thinking they validated. Keep the persona and the answers; it's your artifact from this lesson.
Where the 90% comes from, and why "synthetic respondent" became a boardroom topic
The number anchoring this lesson comes from a 2025 study by PyMC Labs with Colgate (arXiv 2510.08338): they took 57 real purchase-intent surveys, with more than 9,300 human responses, and measured how close AI got. The secret to the result wasn't a better model, it was the reading method, named Semantic Similarity Rating (SSR): instead of asking for the score directly, you ask for a free-text answer and measure its similarity in meaning against pre-defined anchor phrases. This method reproduced around 90% of a human survey's test-retest reliability; asking the model for a direct score did not. In parallel, market research firms (NielsenIQ, among others) started treating the "synthetic respondent" as a serious category. Industry surveys (PyMC Labs and polls of research professionals) project that more than half of research inputs will have a synthetic component within a few years, and startups in the space have already raised rounds at billion-dollar valuations (Aaru, for instance, a Series A at US$1bn in dec/2025). But the same literature that celebrates the gain catalogs the limits: the flattening of responses (that "middle"), the erasure of rare cases (model collapse), cultural bias, and verification asymmetry (you only confirm the synthetic panel got it right by comparing against real data, exactly the cost you wanted to avoid). That's why the rule the industry adopted is "Train Synthetic, Test Real" combined with a tiered risk framework: low risk resolves in synthetic, high risk requires human. You don't need to build the SSR pipeline by hand; you need to demand the right way of asking and never treat the synthetic panel as the final verdict on an expensive decision.
Practice
1. You want to test purchase intent for your new offer with a synthetic customer. Which approach gives the most reliable reading?
2. Your offer lives off a small, very specific niche of customers. Which synthetic-customer trap most threatens your reading, and what should you do?
3. After interviewing your synthetic customer, you have a good picture of objections and want to act. Which decision respects the rule of the technique?
Let's close out the message together, because this lesson hands you one of the most seductive and most treacherous tools of the AI era in marketing. The synthetic customer is an AI persona you interview about your offer, and in fifteen minutes it gives you what used to cost weeks and research money. But it only works if you use it the right way: ask for a reaction in text with the "why," never a little 1-to-5 score that flattens everything toward the middle; force the criticism by asking "why NOT buy," to puncture sycophancy; remember it erases rare cases and your niche; and distrust the generic "local" it pretends to be. The rule worth gold, the one you pin on the wall: explore cheaply in synthetic, confirm expensively in real, and only where the risk demands it. Whoever understands this gets a flashlight to see the customer in the dark before spending the first cent. Whoever doesn't understand gets a mirror that smiles and lies, and makes an expensive decision on top of an automatic compliment. Build yours today, ruler in hand, and use it for what it is: the cheap start of the conversation with your market, never its final word.
For the board
On the methodask for text, not for a score. A reliable reading comes from comparing meaning, not from a number the AI blurted out.
On the niche trapthe AI pulls towards the likely and erases the edges. Your niche is an edge.
On the right useexplore cheaply with the synthetic one. The decision that hinges on your niche you confirm with real people.
Thanks for the feedback. It helps sharpen the next lesson.