If you have ever tried to price a new product, you know the ritual. You ask people what they would pay, they say a number to be nice, and three months later the sales data laughs at you. So it is no surprise that the first question founders ask us is: "Can Cast tell me my price?"
The answer is "it can tell you a lot about your price," and the research explains the difference.
What the studies found
At Harvard Business School (opens in a new tab), James Brand (opens in a new tab), Ayelet Israeli (opens in a new tab) and Donald Ngwe (opens in a new tab) have spent three years on exactly this question in Using LLMs for Market Research (opens in a new tab). Their current conclusion is refreshingly direct: willingness-to-pay estimates from language models are "sometimes comparable to estimates from human studies, but are often inaccurate and in some cases wrong-signed." Fine-tuning on past surveys helps for new features in the same category, but not for brand-new categories. They frame AI as a supplement to human studies, not a substitute.
Perceptions are a different story. Peiyao Li (opens in a new tab), Noah Castelo (opens in a new tab), Zsolt Katona (opens in a new tab) and Miklos Sarvary (opens in a new tab) tested whether models can map how consumers see brands, in Marketing Science (opens in a new tab). INFORMS (opens in a new tab) reported agreement between human and AI-generated data of 75% to 85% (opens in a new tab), while the authors warned that fully automated research "may not be accurate for all product categories."
Humans plus AI beat either alone
Our favorite result comes from Neeraj Arora (opens in a new tab), Ishita Chakraborty (opens in a new tab) and Yohei Nishimura (opens in a new tab) in the Journal of Marketing. They replicated two real studies from a Fortune 500 food company and had experts judge the work. Human and AI hybrids beat both human-only and AI-only analysis. In the AMA's summary (opens in a new tab): "LLMs and humans bring unique, complementary insights that managers should leverage."
A word about vendor numbers
You will see big claims in this space. Qualtrics (opens in a new tab) says synthetic respondents cut research time "from weeks to minutes" (opens in a new tab) and that its fine-tuned model is far closer to human answers than general models in its own test (opens in a new tab) of 518 matched respondents. Ipsos (opens in a new tab) reports eliciting over 250,000 AI responses (opens in a new tab) for conjoint research and calls for "a rigorous examination of its predictive reliability, biases, and limitations." We agree with Ipsos. Treat any company's self-run test, including ours, as a starting point.
How Cast handles price
Here is how we put the research to work. We never anchor an AI user with your number first. We ask what they pay today, then walk a price ladder, then ask what would make them walk away. And we report the result as a range and a ranking ("fair at 260, walk-away at 320, subcontractors most price sensitive"), not a precise forecast. Use that to choose two or three prices worth testing with real buyers. That test will be cheaper and sharper because you ran Cast first.
Why stated prices fail, with or without AI
Asking people "what would you pay?" has a long and unhappy history. People are polite, they have no budget in front of them, and nothing is at stake. Economists call the gap between what people say and what they do hypothetical bias, and it shows up in almost every stated-preference study. An AI respondent inherits this problem, because it learned how people talk about money, which is not the same as how they spend it.
The second problem is the anchoring effect (opens in a new tab). Show someone a number first and their answer drifts toward it. Language models are, if anything, more suggestible than humans here. If your question says "would you pay $49?", you have mostly measured your own number.
Better questions about price
Good pricing research rarely asks for one number. It asks for boundaries and trade-offs. The Van Westendorp Price Sensitivity Meter (opens in a new tab) asks four questions: at what price is it too cheap to trust, a bargain, getting expensive, and too expensive to consider. Conjoint analysis (opens in a new tab) makes people choose between bundles, so price competes with features instead of floating alone.
Both techniques translate well to Cast because they ask about reasoning, not just a figure. When an AI user says "anything over 200 a month and I'd rather keep my spreadsheet", you learn the alternative you are really competing against.
A pricing study you can run this week
- Ask what each AI user pays today for the problem you solve, including the cost of doing nothing.
- Walk the four Van Westendorp boundaries without mentioning your price.
- Offer two or three packages and ask which they would pick and why the others lost.
- Ask what would make them cancel after three months.
- Take the price range and the top objection to five real buyers before you publish a pricing page.
Talk about price before you build, not after
Most founders treat pricing as the last step: build the product, then pick a number. Madhavan Ramanujam and Georg Tacke of the pricing consultancy Simon-Kucher argue the opposite in Monetizing Innovation (opens in a new tab), whose subtitle says it all: How Smart Companies Design the Product Around the Price. Their point is that willingness to pay is a product question. If nobody will pay for a feature, the cheapest time to learn that is before anyone writes the code.
That is where early synthetic interviews earn their keep. They are not good at giving you a final number, as the studies above show. They are very good at telling you, cheaply and early, which parts of your offer people talk about as worth money and which parts they talk about as nice extras.
A small pricing glossary
Willingness to pay (WTP). The highest price a person would accept for a product. Economists have a long literature (opens in a new tab) on measuring it, and almost all of it warns that asking directly overstates it.
Reference price. The price a buyer has in their head when they judge yours, often what they pay today or what a rival charges. A reference price (opens in a new tab) is why "cheaper than hiring a part-time bookkeeper" lands better than "only 300 a month".
Gabor-Granger. A method developed in the 1960s by Clive Granger and André Gabor. The Gabor-Granger method (opens in a new tab) asks about purchase intent at one price, then moves the price up or down depending on the answer, to find the highest price each person still accepts.
Becker-DeGroot-Marschak (BDM). An incentive-compatible method (opens in a new tab) from experimental economics, first published in 1964, where people state a bid and a random price decides whether they actually buy. Because real money changes hands, overstating costs them something.
Decoy. An option added mainly to make another one look better. In Predictably Irrational, Dan Ariely describes an old Economist offer where print-only cost the same as print plus web. In his experiment with students, 84% chose print plus web with the decoy present. Remove the decoy and 68% chose the cheaper web-only option (decoy effect (opens in a new tab)).
A worked example
Imagine a founder building scheduling software for independent physiotherapy clinics. She has three packages in mind and a hunch that 120 a month is right. Here is how a sensible first week might look.
- She runs a Cast study with two audiences: solo practitioners and clinics with three to six therapists. She never mentions her packages or her price.
- She asks each AI user how they book appointments today and what that costs them, counting reception time and missed appointments.
- Only then does she show the three packages, in random order, and asks which they would choose and which feature they would drop first to save money.
- The pattern she sees (still hypothetical): solo practitioners compare her against a free calendar app, while larger clinics compare her against a receptionist's hours. Two different reference prices, so probably two different packages.
- She takes that hypothesis to real clinic owners, with an actual quote they can accept or refuse.
Notice what Cast gave her: a reason to split her pricing, and better questions. Notice what it did not give her: the number. That still comes from people with real budgets.
What to do after the synthetic round
The strongest follow-up is anything where real money or real commitment is at stake: a paid pilot, a pre-order, a signed letter of intent, or a pricing page that shows real prices to real visitors. If you want a structured study, a BDM-style exercise with actual purchases is far more honest than a survey. Synthetic interviews narrow the options. Commitment tests choose between them.
Questions founders ask about pricing with AI
Can I quote a Cast price range in my investor deck?
Quote it as a hypothesis, not as market data. "Synthetic interviews suggest a 150 to 250 range; we are validating with ten design partners" is honest and strong.
Is the AI biased toward cheaper answers?
Sometimes. Models can be cautious about spending other people's money. That is why we look at the ranking between packages and the reasons given, which is more stable than the absolute level.
What about B2B prices that depend on negotiation?
Ask about budget ownership and the approval process instead. Who signs, at what threshold, and what proof they need is often more useful than a number.
How do I price a SaaS product using customer interviews?
Ask about current spending and alternatives first, then about trade-offs between packages, and only then about price boundaries. Use the answers to pick two or three prices, and test those with offers people can actually accept.
What is the difference between willingness to pay and ability to pay?
Ability to pay is about budget: can they afford it at all? Willingness to pay is about value: would they choose to spend it on you? An AI user can tell you a lot about both, but only real buyers prove either.
Read next: replaying psychology’s greatest hits on a machine.
