Last Wednesday, 11pm, I was hunched in a café rewriting surveys
14 customers said "we'll buy." Launch night: 11 went silent. The other 3 DM'd me asking for a discount. I've made this mistake before — treating "the survey says they'll buy" as proof, when all you're really measuring is "willing to do me a favor."
Small teams die from sample size and vibes.
Then I heard about Simile AI — a $2B unicorn
They do something counter-intuitive: use AI to simulate thousands of "fake users" who test new products, policies, and copy for companies. The founder, Joon Sung Park, authored the 2023 generative agents paper — that famous experiment where AI little people made friends, threw parties, and created drama in a virtual town.
How it works: recruit real humans for multi-hour interviews + scrape their purchase and browsing data + run randomized controlled trials, then train a "digital twin" for each person. The accuracy is wild — it nearly matches how consistently the same real human answers the same survey twice (85%). Clients include CVS pharmacy chains, running tens of millions of simulations per project.
The backers are legit: Fei-Fei Li and Andrej Karpathy both invested. GreenOaks and Index led a $200M Series B. This isn't a pitch deck — Fortune 100s are paying for it.
What it means for us small teams: product decisions might not require gambling on 14 survey responses anymore.
Replicate cost: don't get too excited yet
Money: Simile only takes big clients right now. Regular folks can't get in the door.
Time: Needs behavior data + research + model training. Not a few-week DIY.
Technical barrier: Core is behavioral science + LLM fine-tuning, not just clever prompting (the instructions you type into an AI).
But we can hack a poor-person version tonight:
Step one: open ChatGPT or Claude, paste this —
"You're my target customer. Background: [30-50 words describing your customer]. I'm about to pitch you a product: [one sentence]. Respond like a real human — push back, roast me, call it too expensive. Go."
Run 5-10 rounds with different personas. You'll instantly see: that "must-have" feature might be worthless in actual user eyes.
How to use this at each stage
Just starting (0 customers): I'd run 10 free AI persona chats first — saves a week of survey work. Remember, this is a supplement. Real human feedback is still the gold standard.
Got 1-2 paying customers: I'd pour energy into 3-5 deep interviews with existing customers. Their stories are worth more than 1,000 AI simulations — AI still can't replicate hesitation and trailing off mid-sentence.
Scaling up (product going wide): I'd keep watching for when Simile-like tools open up to small accounts. Short-term: A/B testing (showing half your users version A, half version B, and comparing the numbers) + real user data. This tool isn't for everyone — but if you're burning $10K+ monthly on user research, add it to your watchlist.