Most investment teams now use AI somewhere in due diligence. The open question is which kind of tool to use: a generic LLM (ChatGPT, Claude, Gemini), a data platform with an AI layer on top, or a dedicated evaluation engine built for the job. Each has a distinct failure mode, and the differences show up exactly where diligence matters most: verifying founder claims.
The short version: generic LLMs are excellent thinking partners and terrible verifiers. Data platforms are strong on the numbers they own and blind outside them. Dedicated evaluation engines like Kuanta exist to close that gap, with every claim cross-checked against external sources, scored against 645+ industry-specific criteria, and backed by citations you can audit.
Generic LLMs: great analyst, unreliable witness
ChatGPT, Claude and Gemini are genuinely useful in diligence: summarizing a data room, drafting question lists, stress-testing an investment thesis. Used that way, they save a great deal of time.
The failure mode appears when they are asked to evaluate. A generic LLM does not systematically check claims against outside sources, and it benchmarks against nothing. It will confidently score a pitch deck using the deck’s own numbers, which is circular. And because these models are trained to be helpful, they tend to tell you what you want to hear about a deal you already like.
“An LLM that evaluates a pitch deck using the deck’s own claims gives you the founder’s narrative, restated fluently.”
Data platforms with AI layers
PitchBook, Crunchbase and similar platforms have added AI assistants over their datasets. These are strong at what the underlying database is strong at: retrieving and summarizing the financial and firmographic data they own.
Their limit is scope. They can evaluate what is in the database. What the founder claims about market size, technology, competitive position and team is the substance of early-stage diligence, and it sits largely outside firmographic data, so the AI layer cannot verify it.
Dedicated evaluation engines: verification as architecture
A dedicated engine is built around the diligence workflow itself. Kuanta ingests a pitch deck or company name, extracts every material claim, and cross-checks each against external sources. The source is cited next to the finding, and unverifiable claims are explicitly flagged.
Scoring adapts to sector and stage. A biotech is evaluated on clinical pipeline, regulatory pathway and IP, and a hardware startup on manufacturing readiness, supply chain and unit economics, all drawn from 645+ industry-specific metrics. The result is benchmarked against real peer startups instead of a generic average, and delivered as a structured report covering team, market, product, financials, traction and risk, each scored, with an executive summary and a full audit trail.
Independent validation matters here. In an independent master’s thesis at Rotterdam School of Management (Erasmus University, 2026), Kuanta and a jury of human analysts scored the same 150 startups, and the outcomes were tracked afterwards. Overall predictive accuracy was comparable, Kuanta identified more of the startups that later raised funding, and the strongest predictor of all was the two agreeing with each other.
“An independent master’s thesis at Rotterdam School of Management (2026) compared Kuanta with human analysts on 150 startups: comparable accuracy overall, and more of the later-funded startups identified by Kuanta.”
Claim verification
Every founder claim is cross-checked against external data sources, and unverifiable claims are flagged instead of repeated.
Sector-adaptive scoring
645+ industry-specific metrics. The framework changes with the startup’s sector and stage instead of forcing one scorecard on everything.
Real-peer benchmarking
Scores are put in context against genuinely comparable startups, drawn from a validated database of 3.9M+ companies.
Auditability
Every finding carries its source citation, so an investment committee can trace any number in the report.
Choosing for 2026
Use generic LLMs for what they are: fast thinking partners for drafting and summarizing. Use data platforms for the financial records they own. When the decision depends on whether the founder’s claims are true, which is what due diligence means, use a tool whose architecture is verification: cited, benchmarked, sector-aware evaluation. That is the standard Kuanta was built to, and the simplest way to test it is to run a deck you know well and audit the citations yourself.
Kuanta Intelligence Team
Published on Sep 21, 2026 · Updated September 2026 · Part of the Kuanta Research & Venture Decision Science series.