Lean on it for: which one is better
0.85: the correlation between predicted and real effects in a study of 70 US survey experiments published in Nature, as its authors report it. Language models are good at saying which wording will land harder.
Then ship it to the ones who do. Cast of Thousands is building crowds of AI-simulated users and audiences, shaped like yours, to test whatever you’re about to launch.
Try your…
where it brokewhat came throughWe have no product yet. This is the kind of answer we’re building towards.
812 get what they came for. The other 188 fail in four distinct ways, each one a fix you can make before a real customer finds it.
Testing AI agents170 finish. 370 stop at the same step, “connect your calendar”. That is the screen to redesign before launch.
Product and UX testingB wins clearly. Trust the order more than the exact split: simulation ranks options well and guesses percentages badly.
Message and ad testingThree entries pull ahead. Third and fourth are too close to call, and a ranking should say so instead of hiding it.
Ranking a large pileA simulated person is more than a prompt that says “act like a customer”. Each one has a goal, a temper, habits and a history, drawn from a description you write or from your own data: transcripts, survey answers, analytics.
Then a thousand of them use your product, read your message or talk to your agent, all at once. The answer is meant to come back in minutes, with every one of them able to say why.
Then the part most simulation skips. We’re building it to compare what the crowd did with what your real users did, so every result comes with how far to trust it.
Call sheet5 of 1,000 · illustrative
| No. | Who | Wants | What happened |
|---|---|---|---|
| 0031 | Returning customer, in a hurry, charged twice | A refund without repeating herself | Resolved in four turns |
| 0417 | First-time user, on a phone, on a train | To book one appointment | Quit at “connect your calendar” |
| 0562 | Procurement lead, wary of sales calls | To know what leaving would cost | Chose headline B |
| 0788 | Second-language speaker on a bad line | To change a delivery address | Misheard twice, hung up |
| 0903 | Judge who prizes feasibility over polish | The entry most likely to deliver | Picked #214 over #187 |
Lean on it for: which one is better
0.85: the correlation between predicted and real effects in a study of 70 US survey experiments published in Nature, as its authors report it. Language models are good at saying which wording will land harder.
Don’t lean on it for: the exact percentage
12.4: the average miss, in percentage points, when Pew Research Center had AI twins of its own panellists answer nearly 300 survey questions. A simulated poll is not a poll.
The same holds for testing AI agents. A study with 451 real participants found that simulated users are an “easy mode” that inflates agent success: too cooperative, too alike, never properly frustrated.
So we won’t sell you a crowd on trust. Use it to rank options and to find where things break. For anything else, we’ll show you how the simulation compared with reality before you rely on it.
There’s no product to sell you yet. We’re building it with the first people on this list, and what you write here decides what gets built first.