Ask a general chatbot whether a startup is a good investment and you get a polite, balanced answer: strong team, large market, some execution risk. It reads well and tells you nothing, because the model only saw the deck, and the deck is the founder’s best version of events.
Evaluating a startup takes more than one prompt. At Kuanta, nine specialist agents work on each evaluation. One reads the financials, another checks the patents against the registry, another compares traction claims with public data, and one looks for contradictions between them. Every claim ends up with a source or a flag.
Try it yourself
Take a pitch deck you know well, give it to a general chatbot and ask whether the startup is a good investment. The answer will be polite, balanced and oddly familiar: the team is strong, the market is large, there is some execution risk, and further due diligence is recommended. Now try it with a deck you know to be weak. You will get almost the same answer.
Why the answer is always so agreeable
There are two reasons, and neither is a flaw you can prompt your way around. The first is that the model has only seen the deck, and a deck is the founder’s best version of events. A summary of it, however well written, carries the same blind spots.
The second is that evaluating a startup is several different jobs. Someone has to turn the story into a list of claims that can be checked. Someone has to look those claims up in the patent registry, on the company’s own website and in public filings. Someone has to work out who the real competitors are, including the ones the founder did not mention. Someone has to rebuild the financial logic and see whether it holds together. And someone has to read all of it side by side and notice where the pieces contradict each other. A single prompt skips all of that and writes a fluent paragraph about a document.
What the nine agents do
Kuanta splits the work the way a good investment team would. Nine specialist agents each take one part of the evaluation. One turns the deck, the financials and the cap table into structured claims. Another maps the market and the competitors from live sources. Another rebuilds the unit economics and tests the assumptions about burn and runway. And one has a single task, which is to be sceptical: it looks for gaps between what the company says and what the outside world shows, and it turns those gaps into questions for the founder.
The company is then scored on 645+ criteria, in a framework chosen for its sector and stage, so that a biotech is judged as a biotech and a marketplace as a marketplace.
What about made-up numbers?
Anyone who has used a language model knows it can state something false with complete confidence, and in due diligence one invented number can undermine a whole memo. We do not claim to have made that impossible, and you should be wary of anyone who does.
What we do is make it easy to catch. Every finding in a Kuanta report points to where it came from, whether that is a slide in the deck, a registry entry or a public page. When a claim cannot be verified, the report says so and labels it as the founder’s assertion. A wrong statement with a source attached gets found quickly, whereas a wrong statement inside a smooth paragraph can survive all the way to the investment committee.
Does it work?
An independent master’s thesis at Rotterdam School of Management compared Kuanta with a jury of human analysts on 150 startups and then tracked what happened to the companies. Overall accuracy was comparable, Kuanta identified more of the startups that later raised funding, and the best predictor of all was the two agreeing with each other.
That is how we see the role of this kind of tool. The analyst still makes the call, with a second read next to them that has done the checking and can show its work. If you want to see the difference, run one deck through a general chatbot and through Kuanta, and compare what comes back.
Questions people ask
Can ChatGPT evaluate a startup? A general chatbot can summarise a pitch deck, but it only sees what the founder wrote and does not check claims against outside sources, so its assessments tend to be generic.
What is agentic startup evaluation? An approach in which several specialised AI agents each handle one part of due diligence, such as claim extraction, market mapping, financial logic and contradiction checks, with every finding traced to a source.
L. Meijer (2026), Man versus machine, master’s thesis, Rotterdam School of Management · Kuanta product documentation
Kuanta Engineering Team
Published on Aug 24, 2026 · Updated September 2026 · Part of the Kuanta Research & Venture Decision Science series.