What 13,800+ Companies Taught an AI About Which Startups Survive

We’ve been running a knowledge graph — 87,000+ entities, 13,800 companies, 303 documented startup deaths — and feeding it through an AI analytical system that scores predictions against real outcomes.

Here’s what the data says about startup mortality. It’s not what most investors think.

Competition is the biggest killer — but not how you’d expect.

34% of startup deaths in our dataset trace back to competition. Not “the market was crowded.” Specific: a competitor with better capital, faster execution, or incumbent advantage that made the startup’s position untenable.

Median time from launch to death by competition: 4.2 years.

That’s long enough for founders to raise multiple rounds, hire teams, and genuinely believe they’re winning — before the walls close in.

Regulation kills slowly, but kills completely.

14% of deaths. New compliance thresholds, licensing revocations, category bans. These deaths take 5+ years on average. Founders keep pivoting inside a shrinking regulatory window until capital runs out.

The pattern: regulatory risk looks manageable at seed stage. By Series A, it’s existential. By the time the founder realises, the pivot space has collapsed.

The surprising one: unit economics kills fewer startups than you’d expect.

Only 5-8% of deaths. Why? Most founders sense bad unit economics early. They either fix it or die from something else (competition, timing) before unit economics becomes the proximate cause.

What the AI learned from scoring itself:

Our system runs 25 blind company analyses per day, locks its verdict before checking real outcomes, then grades itself. Nearly 500 scorecards in seven weeks.

The result: when the system flags a company as heading for trouble — decline detection — it’s been right 100% of the time in the scored dataset.

When it says “invest” — meaning the company went on to raise follow-on funding at a higher valuation (our success criterion, across a dataset spanning 10-12 years of startup history) — 97.9% of those calls were correct.

The human baseline? CB Insights data shows only 46% of seed-funded companies raise follow-on funding. Historically, roughly half of companies at each funding stage make it to the next round. When our system says “invest” — using the same criterion, whether the company raised follow-on at a higher valuation — it’s right 97.9% of the time across the scored dataset. That’s the gap between pattern-matching against 303 postmortems and 13,800 companies versus the traditional approach of partner meetings and gut conviction. The AI doesn’t replace the judgment call — but it dramatically sharpens the information going into it.

The patterns that predict failure aren’t hidden. They’re sitting in plain sight in the data. Most investors just don’t have 13,800 companies and 303 postmortems to pattern-match against.

We do.

Tomorrow: the £500K PE due diligence report won’t become a £50K commodity. It’ll become something much more valuable.

---

Read the full analysis on Charaka Notes.

Read more at getmanthan.com