Almost every time I show Cast to a founder, an investor or a corporate team, someone asks the same question within ten minutes: "Couldn't I just ask ChatGPT (opens in a new tab) or Claude (opens in a new tab) to pretend to be my customer?"
It is a good question, and I asked it myself before we built Cast. I use both tools every day, and I think they are extraordinary. So I tried. For months, across real decisions at Enhance Ventures (opens in a new tab), I tried to get customer research out of a chat window with ever cleverer prompts.
Here is what I learned: the model was never the problem. The process was.
What happens when you ask a chat assistant to be your customer
Try it yourself. Describe your product, describe your customer, and ask the assistant to answer as that person. You will get a thoughtful, articulate, plausible reply. It will probably like your idea, with one or two sensible caveats. Ask it to be twelve different customers and you will get twelve slightly reworded versions of the same thoughtful, articulate, plausible person.
It feels like research. It is not. Here is why.
1. One model is playing every role, and it knows your hopes
In a chat, the same model reads your pitch, invents the customer, asks itself the questions, answers them and summarizes the result. It knows what you are hoping to hear because you told it in the first message. Researchers at Anthropic (opens in a new tab) showed in Toward Understanding Sycophancy in Language Models (opens in a new tab) that AI assistants tend to tell people what matches their stated views. In customer research, that is fatal.
In Cast, the roles are separated and each is blind to something. The Casting Agent never sees your questions. The AI user never sees your hypothesis. The Director never sees the AI user's hidden backstory. And the Critic, a separate reviewer, checks every line. I explain the reasoning in Why Cast uses four AIs instead of one.
2. A chat assistant is trained to be helpful, not to be a customer
General assistants are tuned to be agreeable, polite and useful. Those are wonderful qualities in an assistant and terrible qualities in a research participant. Real customers are distracted, skeptical, short on budget and loyal to whatever they use today. Getting a model to hold that stance consistently, across a whole interview, under follow-up questions, is not a matter of adding "be critical" to a prompt. It takes a casting process that gives each person a reason to doubt you, and a reviewer that sends back answers that are too eager to please.
3. One invented persona is a stereotype, not a population
When you ask a chat assistant for "a 35-year-old operations manager in Dubai", it fills the gaps with the most typical version it can imagine. That is the average of a stereotype. Studies like Whose Opinions Do Language Models Reflect? (opens in a new tab) from Stanford (opens in a new tab) show how default model voices lean toward some groups and away from others.
Cast does two things a prompt cannot. First, the Casting Agent researches each person's market, city, prices and constraints, and keeps the sources so you can check them. I describe that in Grounded, with sources. Second, we apply Fuzziness, our proprietary technology: we cast the people around your profile across eleven default dimensions and optional ones you add, so your findings come from a varied population rather than one voice repeated. That is the idea I wrote about in Why we needed synthetic users and invented Fuzziness.
4. Everything leaks in one conversation
In a single chat, the fifth "customer" has read the answers of the first four. Each reply quietly shapes the next. It is the AI version of a focus group where the loudest voice sets the tone, the problem Solomon Asch (opens in a new tab) documented in his conformity studies.
Cast runs every interview as its own conversation, in parallel. Nobody hears anyone else. Twelve people means twelve independent conversations, which is what makes it meaningful when eight of them raise the same objection.
5. Nobody checks the answers
A chat assistant grades its own homework. Research from Google DeepMind (opens in a new tab), Large Language Models Cannot Self-Correct Reasoning Yet (opens in a new tab), found that models struggle to fix their own mistakes without outside feedback. In Cast, the Critic reads every line against the AI user's backstory and sources, and rejects answers that are out of character, unsupported or simply too nice. Every decision is logged so you can see it.
6. A chat gives you text, not research
Even if every answer were perfect, a chat window leaves you with a long scroll. Cast turns interviews into findings, patterns across audiences, quotes linked to the transcript they came from, reports and decks. And because AI users keep their memory, you can come back days later with a follow-up question and get an answer in character.
So is it the prompt or the process?
People sometimes imagine Cast as a very long prompt. It is not. It is a pipeline: research, casting, Fuzziness, separate interviews, deliberate skeptics, independent review and synthesis, each step with its own instructions and its own information. We have spent a long time testing and tuning how those pieces fit together, and that is where the quality comes from.
You could, in principle, rebuild parts of it yourself with enough engineering. But no amount of clever wording inside a single chat gets you there, because the problems above are structural. They come from one model doing everything in one place.
When a chat assistant is the right tool
I want to be fair. For brainstorming a questionnaire, drafting interview questions, or getting a quick first reaction to an idea, a chat assistant is excellent, and I use one for exactly that. The line is when you need to rely on the answer: when you are choosing which product to build, which customer to target or what price to charge. That is when you need research, not a conversation.
What I would suggest
- Use a chat assistant to sharpen your questions.
- Use Cast to run them across a varied, researched, independently checked cast.
- Take the strongest patterns to real customers before you commit serious money.
That combination is faster and cheaper than anything I had in my consulting or investing years, and far more honest than a chat window that is quietly trying to please you.
Isn't Cast just a wrapper around ChatGPT or Claude?
No. Cast uses leading language models as engines, but the results come from the process around them: researched casting with sources, Fuzziness, separate parallel interviews, deliberate skeptics and an independent Critic on every line.
Can I get the same results with a better prompt?
Not in a single chat. The main problems, one model playing every role, answers leaking between personas and no independent check, are structural, so they cannot be solved by wording.
Which AI models does Cast use?
Cast uses current large language models from leading providers and assigns different roles to separately instructed instances. The method, not any one model, is what produces the results.
Should I stop using chat assistants for research?
Keep using them for brainstorming and drafting questions. Use Cast when you need answers you can rely on, and confirm the biggest bets with real customers.
Read next: why we needed synthetic users and invented Fuzziness.
