To scale simulation AI startups, prove the calibration first
AI that rehearses decisions on simulated customers, voters and patients is one of 2026's fastest-funded bets. The companies pulling ahead share one habit: before a client spends a dollar, they show the simulation would have called it right. Five of them, and how each earns that trust.
Simile · Unlearn · Aaru · Electric Twin · Listen Labs
$2B
Simile, a year after founding
8 / 10
earnings-call questions called in advance
15×
Listen Labs revenue growth in nine months
8,000+
respondents in Electric Twin's head-to-head
01 · THE BET
A simulation is a cheaper answer to a question someone was already paying to ask
Every simulation company is selling a substitute: for a focus group, a pricing pilot, a placebo arm, a poll. Two things decide whether that substitute sells. How much the real test costs, and how quickly the real answer shows up to check the simulation against. Put those on two axes and the market sorts itself into four very different businesses.
Cost of the real test →
COSTLY · SLOW TRUTH
Needs a referee
Huge savings, but the truth takes years. A regulator or independent study has to vouch for the model.
Clinical trials · public policy
COSTLY · FAST TRUTH
The sweet spot
Real tests are expensive and the answer arrives within weeks, so accuracy can be proven fast.
Pricing · product launches · earnings-call prep
CHEAP · SLOW TRUTH
Avoid for now
Cheap to test for real, and slow to verify. Little to save, little to prove.
Long-run brand sentiment
CHEAP · FAST TRUTH
Easy to prove, hard to charge
Accuracy is quick to show, but the real test is cheap, so buyers won't pay much to skip it.
Ad copy · message testing
How fast the real answer arrives →
MY READ
Most of the money in this space has gone to companies in the sweet spot or the referee box, and I think that's right. Message testing is where demos look best and budgets are thinnest.
If I were choosing a first market, I'd pick one decision in the top-right box and stay there until the accuracy record is public. A narrow track record is worth more than a broad promise.
02 · WHY TRUST IS THE BOTTLENECK
Buyers climb a staircase of belief, and revenue grows at each step
The sales question is never “is it smart?” It's “would it have given us the right answer?” A simulation earns that one step at a time, and each step unlocks a bigger budget: first a pilot, then repeat projects, then part of the research spend, and finally the real test itself.
Claims
A headline accuracy number
Back-tested
Re-runs past decisions
Scored
A confidence score on every answer
Simile
Side-by-side
Matches the client's own research live
Electric Twin · Aaru
Vouched for
A regulator signs off
Unlearn
Pilots that stallRepeat projectsResearch budgetRegulated contracts
Company placement is my reading of public information. Listen Labs sits off the staircase: it interviews real people, so it's the benchmark the others are measured against.
Illustrative. Both models are right on average; only one is right where buyers act.
The four numbers I'd put on the weekly dashboard
Calibration error, by use case
Use cases with a published accuracy check
Side-by-side match rate with clients' own research
Decisions where a simulation replaced the real test
MY READ
Simile has the feature I'd copy first: a confidence model that predicts how accurate each individual simulation is likely to be. That one number does more for trust than any headline accuracy claim, because it tells a buyer when not to rely on the answer.
03 · THE PLAYBOOKPROVEN SALES STRATEGIES · START HERE IF YOU'RE SHORT ON TIME
How simulation startups win deals: six plays with receipts
In self-learning AI, one company held most of the public proof. Here it's spread across all five, and each found a different way to make a buyer believe a simulated answer.
01
Start where the real test is slow and expensive
UnlearnCompany announcements
Unlearn gives every trial participant a digital twin that forecasts their outcome under placebo, so trials need smaller control groups. It goes one disease at a time — its 2026 deals are early-phase neurology studies such as ALS trials at VectorY and SOLA — and in January it moved upstream with TrialPioneer, for planning a trial before the protocol is locked.
Why it works: when a real test costs years and millions, a modest, provable saving already has a budget waiting for it.
1One costly decision→2One disease or category→3Then move upstream
Measure: cost or time saved, as measured by the clientAvoid: ten use cases before one is proven
02
Call something in public before it happens
Simile · AaruOn the record
Simile's CEO said on Bloomberg that it predicted eight of the ten questions analysts actually asked on a company's earnings call. Aaru's polling method correctly called the New York Democratic primary.
Why it works: a dated prediction anyone can check is the cheapest credibility a young company can buy, and it spreads far further than a case study.
1Fixed date, clear answer→2Publish before→3Publish the result either way
Measure: inbound in the 30 days afterAvoid: only publishing the hits
03
Turn a flagship customer into an investor
Simile × CVS HealthOn the record
CVS Health used Simile for five months in place of human focus groups, including work on whether patients complete their prescriptions. CVS Health Ventures then joined Simile's Series B, and Simile reports revenue grew 5× in the five months after launch.
Why it works: a customer with equity becomes your reference, your design partner and your voice inside its industry at the same time.
1Recurring research need→2Long pilot on live decisions→3Seat in the round
Measure: expansion inside the anchor accountAvoid: features only that one customer will ever use
04
Ride the channels that already own research budgets
AaruCompany announcements
Aaru counts Accenture, EY and Interpublic Group as partners, and Accenture Song committed to build Aaru's model, Lumen, into its work on product launches, marketing and customer service.
Why it works: consultancies and agencies are already in the room when research money gets spent, so each of their engagements becomes your sale.
1Map who sells to your buyer→2Fit inside their deliverable→3Let them present
Measure: partner-sourced revenue per quarterAvoid: a partner who owns the client and the data
05
Publish a fair fight against the old way
Electric TwinPublic numbers
Electric Twin ran its synthetic audiences against eight identical surveys covering the industry's main methods — panels, river sampling, microtask platforms — with more than 8,000 respondents, and reported its predictions were indistinguishable from conventional research.
Why it works: buyers don't compare you with perfection. They compare you with what they buy today, and a fair head-to-head answers that directly.
1Same study, both ways→2Publish the method→3Repeat yearly
Measure: how often prospects cite it on a first callAvoid: a test rigged so you can't lose
06
Get the first answer back in hours
Listen LabsPublic numbers
Listen Labs runs AI-led interviews with real people and grew annualized revenue 15× to eight figures in nine months, serving hundreds of enterprises. Microsoft cut its wait for customer research from weeks to hours or days.
Why it works: speed is the benefit a team feels on day one. Once answers come back the same week, research becomes a weekly habit instead of a quarterly project.
1Self-serve first study→2Searchable answer library→3Easy second purchase
Measure: days from sign-up to first finished studyAvoid: fast answers nobody trusts enough to act on
A 90-day plan for one new account
Weeks 0–2
Pick the costly decision and agree, in writing, how accuracy will be judged.
Weeks 2–6
Back-test on the client's past decisions before touching a live one.
Weeks 6–8
Run side-by-side with their own research on a live decision.
Weeks 8–12
Publish the head-to-head with their permission.
Quarter 2
Expand to every repeat decision of the same kind.
MY READ
None of these plays sells the simulation itself. They sell evidence it can be trusted: a dated prediction, a regulator's sign-off, a fair head-to-head, a customer willing to invest. The model gets you in the door; the proof closes the deal.
With one quarter and a small team, I'd run plays 2 and 5 together — one public call and one head-to-head — because between them they answer the first thing every buyer asks.
04 · WHAT PROTECTS THEM
Five simulation AI leaders, and the moat behind each
Four things make a simulation company hard to copy: data from real people, proof of accuracy, distribution, and trust from regulators. No one has all four.
Unlearn
Pharma and biotech sponsors running smaller, faster trials
Charges
Not disclosed
Valuation
Not disclosed
Revenue
Not disclosed
Real-human data
Proof of accuracy
Distribution
Regulatory trust
Hardest to copy: patient-level historical trial data plus a European Medicines Agency qualification. $130M+ raised.
Simile
Enterprise teams testing products, brands and earnings calls on simulated people
Charges
Annual or multi-year subscriptions
Valuation
$2B (Jul 2026)
Revenue
5× since launch
Real-human data
Proof of accuracy
Distribution
Regulatory trust
Trains on interviews with real people — an asset that grows in value as the web fills with AI-written text.
Listen Labs
Product and research teams interviewing real customers with an AI interviewer
Charges
Not disclosed
Valuation
$500M+ (Jan 2026)
Revenue
Eight figures
Real-human data
Proof of accuracy
Distribution
Regulatory trust
A panel of 30 million pre-qualified people and over a million real interviews.
Aaru
Brands, consultancies and campaigns forecasting how populations respond
Charges
Not disclosed
Valuation
~$1B headline
Revenue
Not disclosed
Real-human data
Proof of accuracy
Distribution
Regulatory trust
Distribution through Accenture, EY and Interpublic is its edge, and its dependency.
Electric Twin
Media and consumer brands testing messages before launch
Charges
Not disclosed
Valuation
Not disclosed
Revenue
Not disclosed
Real-human data
Proof of accuracy
Distribution
Regulatory trust
The most transparent accuracy work in the group. $14M raised, led by Atomico.
Bars are my assessment from public information.
MY READ
The moat I'd bet on is fresh data from real people. Every simulation drifts unless it's re-anchored to how humans actually answer, and the text it learned from is getting less human every month.
That makes Listen Labs the most strategically interesting name here, even though it simulates nothing. A few million real interviews is exactly what the simulators need. I'd expect it to either sell that ground truth to them, or start simulating on its own.
05 · WHAT SHIPS NEXT
Product roadmap predictions
Mid 2027
Simile60%
Published accuracy for each use case — earnings prep, pricing, patient adherence — shown next to every result
Because: it already predicts each simulation's accuracy with a confidence model
Watch for: an accuracy page broken out by use case
2027
Aaru55%
Lumen sold as a standard module inside consultancy offerings, not one-off projects
Because: Accenture Song has committed to build Lumen into its offering
Watch for: Lumen named in a packaged service
Research incumbents50%
A major market-research firm buys a simulation startup
Because: synthetic research is pulling budget from panels and surveys
Watch for: an acquisition announcement
End 2027
Electric Twin50%
A dedicated product for governments testing policy
Because: it already tests policy decisions, and its founders led parts of the UK's pandemic response
Watch for: a public-sector contract
Listen Labs45%
Synthetic follow-up answers built from its library of real interviews
Because: teams already search its interview repository over time
Watch for: simulated respondents in the product
Unlearn40%
Digital twins in a late-stage Phase III trial
Because: its next raise likely depends on late-stage proof
Watch for: a Phase III sponsor naming Unlearn
Odds are my own estimates, grounded in what each company has said or shipped.
If you're building in simulation, spend your first year proving one answer was right, not making a thousand answers possible.