AI startup prediction is the use of machine learning models to score a young company’s chance of survival before an investor writes a cheque.
Quick Take
- SignalFire’s Beacon system tracks more than 8 million startups worldwide, watching hiring and code activity.
- Nearly half of EQT Ventures’ portfolio now comes from deals its Motherbrain software found first.
- About 90% of Indian startups shut within five years, which is the gap these models are built to close.
In This Article
An investor reading 40 pitch decks a week cannot check every claim in them. So a growing number of funds now let software read the company first. The software does not read the deck. It reads what the company does.
Between 2023 and 2025, more than 39,860 Indian startups shut down. In the first ten months of 2025 alone, 11,223 closed. Those numbers are why the pitch deck is losing its job as the main filter.
What Does AI Actually Look At in a Startup?
The models ignore the story a founder tells. They watch the traces a working company leaves behind without meaning to.
Hiring is the loudest signal. A company posting six engineering roles is building something. A company quietly losing its third senior engineer in four months is not.
Code activity is the second. Public repositories show how often a team ships. Steady commits suggest a real product. Long silences suggest a stalled one.
The third is money movement. Payroll size, vendor payments and app download curves all move before revenue does.
| Signal | What it suggests | Where the data comes from |
|---|---|---|
| Engineering hires rising | Product is being built, funding is likely near | Job boards, professional profiles |
| Senior staff leaving fast | Internal trouble ahead of any public news | Profile change tracking |
| Frequent code commits | Team ships regularly, product is alive | Public code repositories |
| App installs climbing | Real demand, not just marketing spend | App store rank trackers |
| Web traffic flat for months | Growth story may not match reality | Traffic estimate services |
None of these are secret. What changed is that software can now read all of them for millions of companies at once, every day, and rank the results.
There is a fourth group that matters more at seed stage. The models also score the founding team itself. Where the founders worked before, who they hired first, and whether they have built anything that shipped.
At an early company there is almost no revenue to study. So the team is most of what the data can actually see.
Which Investors Already Run These Models?
This is not a pilot project. Two firms have run it long enough to show results.
EQT Ventures built a system called Motherbrain in 2016. It scores companies on a scale of 1 to 340 and pushes the top names to human investors. Nearly half of EQT’s portfolio now comes from deals the software surfaced first.
SignalFire runs Beacon. It watches more than 8 million startups, tracking where engineers move and how often teams deploy code. The aim is to spot a company before a rival fund hears the name.
| Firm | System | What it does |
|---|---|---|
| EQT Ventures | Motherbrain, built 2016 | Scores companies 1 to 340, ranks leads for partners |
| SignalFire | Beacon | Tracks 8 million plus startups, watches talent and code |
Indian funds are not running systems at that scale yet. What they are doing instead is tightening what they check.
Diligence now leans on customer adoption, unit economics and whether a product is genuinely hard to copy. That is the same question the models ask, asked by people rather than software.
The money behind those questions is real. Peak XV Partners closed $1.3 billion across three funds in late 2025 and manages over $10 billion across 16 funds. Accel is raising a $650 million India fund.
Seed stage is where software would help most, because that is where the volume sits.
Blume Ventures was founded in 2010 by Karthik Reddy and Sanjay Nath. It writes cheques between $500,000 and $3 million and has backed more than 150 companies. A fund working at that cheque size sees far more companies than it can ever meet.
That is the real argument for these tools. They are not built to judge better than a partner. They are built to make sure a fund does not miss a company simply because nobody had time to look.
How Accurate Are the Predictions?
Better than most people expect, and worse than the marketing suggests.
In 2020, researchers at the University of St. Gallen tested an algorithm against 255 angel investors. Both picked from 623 real deals sourced from a large European angel network. The comparison ran against the decisions those investors actually made.
A separate model called CapitalVX predicts a startup’s exit outcome at over four times the accuracy of a successful venture capitalist.
Published accuracy figures vary widely. Two ensemble models reported 94.1% and 94.5% on their test data. Another study using random forest and gradient boosting reported 82% and 80%.
That spread matters. A model tested on companies that already succeeded or failed is grading a finished exam. Predicting a company that has not yet had its outcome is harder work.
What this means for you: treat any single accuracy number with suspicion, and ask what data the model was tested on.
Where Do These Models Get It Wrong?
Three failures show up again and again.
The first is survivorship. Models learn from companies that already finished their story. Winners are easy to study. The ones that quietly folded left far less data behind, so the training set is lopsided from the start.
The second is copying old bias. Suppose a fund’s past winners came from a narrow set of colleges and cities. A model trained on those deals will rank that same profile highly.
It is not finding talent. It is repeating an old habit and putting a number next to it. The number makes the habit look objective.
The third is the signal that can be faked. Once founders learn that hiring posts and code commits move the ranking, some will produce hiring posts and code commits. A signal only works while it is an accidental by-product of real work.
There is also a timing problem. Hiring and traffic move after a decision has been taken inside the company, not before. The model sees the shadow, not the moment.
Why Does This Matter More in India?
Because the failure rate here is higher and the data trail is thinner.
Around 90% of Indian startups close within five years. Startup shutdowns rose roughly 12 times to over 28,000 across 2023 to 2025. An investor filtering that volume by hand will miss things.
But the signals these models rely on are weaker in India. Many strong companies in tier-two cities do not maintain public code repositories. Plenty of profitable businesses run on WhatsApp and never appear in an app store chart.
A model trained mostly on American software companies will read that silence as weakness. It is not weakness. It is a different way of building.
So the honest position is this. The models are useful for narrowing a list. They are poor at understanding a market they were not trained on.
What Should Founders Do About It?
Assume a machine reads your company before any investor does. Then make sure it reads something accurate.
- Keep your team page and professional profiles current, with real job titles.
- Post job openings publicly, even when you also hire through your network.
- Keep at least one public repository active if you build software.
- Update your app store listing on a regular schedule.
- Make sure your website states what you sell and to whom, in plain words.
- Correct outdated headcount and funding entries on startup databases.
None of this replaces a real business. A model that reads a busy company with no customers will still score it poorly, and it should.
The wider shift is worth understanding before your next raise. For twenty years a warm introduction was the main way a young company reached a fund that mattered. That route still works, but it is no longer the only door into the room.
A fund running these systems can find a company that never asked to be found. That cuts both ways for founders outside the main startup cities.
About AI Startup Prediction
AI startup prediction is a set of machine learning methods that estimate whether an early company will survive, raise again or close. The models read public signals such as hiring, code activity, web traffic and app performance rather than pitch material. Investors use the output to rank which companies deserve a human conversation, not to make the final decision.
StartupFeed Insight
The interesting number is not the accuracy claim. It is that nearly half of EQT’s portfolio now arrives through software. That shifts what a fund is buying. It is no longer paying for a partner’s contact book. It is paying for coverage. For Indian funds the constraint is different, because the training data barely includes a profitable Coimbatore manufacturer with no GitHub account. Expect the first serious India-trained sourcing model to come from a fund with a large domestic portfolio to learn from, not from an imported tool. I would look for that within the next two funding cycles.
— Dr. Mayank Raj, Contributing Editor
Frequently Asked Questions
Published: 12 August 2026
Have a tip? Write to us at editorial@startupfeed.in.
