Simile Is Worth $2B for Pretend Users. Here's What You Should Trust to Them.

Two things landed in the same week that change how you learn from customers. A startup that sells fake users hit a $2 billion valuation. And a wave of voice AI is now built by studying real customer calls. One says skip the humans. The other says the humans are the whole point. You have to pick a side for real work, and soon. Let me catch you up.

The pitch that sounds too good

Simile just closed a $200 million Series B at a $2 billion valuation, five months after its Series A. The product is simulated users you can run marketing and product research against. Founder Joon Sung Park built his Stanford work on "Smallville," AI agents living fake lives, throwing fake parties. The stated mission is to simulate all eight billion people on earth.

The money is real, and the appeal is obvious. Synthetic users are fast and cheap. You can test a mock-up against a thousand fake personas before lunch. TechCrunch calls it vibe coding for research, and that lands. Aaru, a rival, also raised at a billion-dollar headline number.

But TechCrunch's own reporter calls the eight-billion goal "preposterous," and the reason matters. You run research because people are unpredictable, driven by feeling and reason at once. A model that predicts the average answer is not the same as a person surprising you.

The other bet: mine the real thing

While Simile sells simulation, another group is getting paid to study actual humans. Encore AI raised $30 million to build voice agents trained on real customer calls. CEO Dvir Ginzburg calls it "interaction mining." The platform pulls call recordings, emails, texts, and CRM data, then finds which parts of a conversation moved a deal and which killed it.

The detail that sticks: Encore's agents sometimes tell the same jokes the human relationship managers tell, because the system runs the playbooks it saw working. Annual revenue is up more than 5x in under 18 months, with 40-plus enterprise customers, mostly banks.

These are two opposite theories of where insight comes from. One generates plausible people. One studies the transcript of what a real person actually did. For your roadmap, that is the fork.

Fake voices break in ways that cost you

Voice is where the tradeoff gets concrete. Fish Audio raised $50 million with 8 million users and $21 million in revenue, offering 15,000-plus controls for expressive, steerable voices. Impressive, and it also ran into a mess: creators said their voices got uploaded without consent, and take-downs were slow.

Investor Oskue Honda drew the line plainly. A community-driven model "can only become a durable advantage if creators trust the platform," and consent, transparency, and attribution have to be built in, not bolted on. Synthetic output feels free until the provenance question shows up on your desk.

The German medical scribe writeup makes the failure mode physical. Speech models fabricate. They produce fluent, confident text from silence, a cough, or background noise. Harmless in captions. In a clinical note, that is an invented statement in a legal record. The team built a whole layer just to make low-signal audio produce nothing instead of a confident lie.

Synthetic works when you box it in

Here is where the medical piece flips the whole debate. Their claim is blunt: on German medical transcription, the cloud baseline hits 0.82 term recall, and a layered self-hosted pipeline beats it at 0.91. Synthetic data drove that result. They scored 5,504 synthetic transcriptions to map where models failed, then anchored it on five real consultations.

The win came from narrowing, not broadening. Loading a full medical dictionary made things worse, because the model started inserting words nobody said. A small, per-patient term set recovered 96 percent of otherwise-missed terms. Synthetic scale plus a real-world anchor, not synthetic instead of real.

That is the pattern worth stealing. Simulation is strong for coverage and stress-testing, for mapping where things break at volume. It is weak as the final word. The medical team never let the synthetic set stand alone. Neither should you.

The deep cut

The $2 billion number is not permission to swap your research team for a simulator. It is a signal that synthetic data is now good enough to do the boring, high-volume part: broad coverage, edge cases, first-pass reactions to twelve variants. The medical pipeline shows the shape that works, synthetic breadth with a real-human anchor holding it honest.

So split the job. Use synthetic users to widen your net cheaply and find where a design breaks. Use real signal, real calls, real transcripts, real interviews, to confirm anything you're going to bet the roadmap on. The failure you're guarding against is the same one that layer catches in the clinic: a confident, fluent answer generated from nothing. On a product decision, an invented insight costs you a quarter.

Three questions for your team

1. Which research decisions are we comfortable making on simulated users alone, and which ones need a real-human anchor before they hit the roadmap? Write the line down.
2. If we adopt a voice or synthetic-user tool, what is our check against fabricated output, the plausible answer built from noise? Who owns catching it?
3. We already have a pile of real customer calls and transcripts. Are we mining them like Encore does, or paying for simulated versions of insight we already own?