We build AI for private markets, so people assume we think AI can do everything. We do not. There's a misconception that if a language model is smart enough, it can run due diligence.
Intelligence isn't the problem. Evidence is.
The fastest way to show you the line between "useful AI" and "AI that will lose you money" is to walk right up to it and poke it. So we took a real diligence scenario, handed it to a plain consumer chatbot, and asked it to do an analyst's job.
Dear Reader, it tried so hard. And that was the problem.
The experiment
We gave a general-purpose chatbot a partial set of company materials, the kind of incomplete pile you actually get in a real deal, and asked it to do what an analyst does: summarise the financials, assess the competition, and flag the risks. We did not warn it that some information was missing, because nobody warns you in real life either.
And the output is as follows, lightly anonymised and entirely real in spirit.
Finding 1: It invented revenue
One of the financial documents was incomplete. The latest annual revenue simply wasn't there.
A human analyst would stop, mark the gap and ask management for the missing figure. That's how due diligence works. Missing information isn't an inconvenience, it's a finding.
The AI took a different approach. It produced a precise revenue figure that appeared nowhere in the documents.
The number wasn't absurd. In fact, it looked perfectly reasonable. It followed the historical trend and fit the profile of the business.
It just wasn't real.
That's what makes hallucinations dangerous in an investment workflow. The invented number and the genuine numbers looked identical. There was nothing in the response to tell the reader where one ended and the other began.
Finding 2: It conjured a competitor that doesn't exist
Next, we asked the model to assess the competitive landscape.
It identified three competitors.
Two were legitimate. The third wasn't.
It had a convincing name, a believable market position and even a description of the products it supposedly offered. As far as we could verify, the company simply didn't exist.
Again, this wasn't irrational behaviour from the model.
Language models are designed to predict the most likely next token, not to distinguish between verified facts and plausible possibilities. When evidence is missing, they don't naturally stop. They continue generating what appears most likely.
For most writing tasks, that's a strength.
For investment research, it isn't.
If you put that in an IC memo, you are now defending a competitive analysis built partly on a ghost.
Finding 3: It treated uncertainty as certainty
This was the most interesting result.
The model asked almost no follow-up questions. It rarely acknowledged that information was missing. Instead, it produced something that looked complete.
That's exactly what a general-purpose AI has been trained to do. Its objective is to be helpful and coherent, not to leave blank spaces.
Due diligence has the opposite objective. The blank spaces are often the most valuable part of the analysis.
Every unanswered question becomes a request to management. Every inconsistency becomes a follow-up investigation. Every missing document becomes a potential risk.
A report that looks complete too early should make investors nervous, not confident.
Generic AI predicts. Due diligence verifies.
This is the real distinction.
It isn't about whether ChatGPT, Claude or Gemini are "good enough."
They're solving a different problem.
A general-purpose language model is optimised to produce fluent, plausible responses from the information available.
Due diligence is optimised to separate facts from assumptions.
When the evidence doesn't exist, the correct answer isn't the most probable one.
It's simply:
"The information required to answer this isn't available."
Those are fundamentally different optimisation goals.
Why this happens (it is not a bug, it is the design)
None of this means the model is bad. This isn't about whether ChatGPT, Claude or Gemini are "good enough." The problem is asking it to do a job it wasn't designed for.
Generic LLMs are built to produce fluent, plausible, helpful-sounding text. When the information is there, they are genuinely excellent at pulling it out. When it is not there, they do not stop. They generate the most probable-looking thing, because refusing is not their default behaviour. They would rather be confidently wrong than admit a gap, in the same way a nervous candidate keeps talking instead of saying "I do not know."
In a brainstorming session, that tendency is harmless and occasionally delightful. In diligence, where the entire point is to separate what you know from what you are assuming, it is exactly the wrong instinct.
What diligence-grade AI does differently
The difference isn't a smarter model. It's a different system.
A diligence platform should work from one simple principle:
Every conclusion should be traceable back to evidence.
That means:
- Every answer is grounded in the documents provided, not the model's memory.
- Every claim links back to its source.
- Missing information is explicitly flagged instead of inferred.
- Contradictions across documents are surfaced instead of smoothed over.
- Confidence comes from evidence, not eloquence.
That's why askRIA wasn't built as a chatbot wrapped around an LLM.
Its Due Diligence Agent treats uncertainty as a feature rather than a flaw. If a number isn't in the data room, it says so. If two documents disagree, it highlights the conflict. If something can't be verified, it refuses to invent an answer.
That behaviour isn't flashy. It's exactly what's required.
That sounds boring next to "AI writes your whole memo", and boring is precisely the point. In diligence, the tool that admits what it does not know is worth ten that sound impressive.
So, should investors use ChatGPT for due diligence?
Absolutely, but only for the parts it's good at.
- Summarising documents.
- Explaining accounting concepts.
- Rewriting notes.
- Brainstorming management questions.
- Drafting emails.
Those are all excellent uses of a general-purpose AI. What it shouldn't become is the analyst of record.
Investment decisions depend on evidence, not probability. Any number, claim or conclusion that influences a deal should always be traceable back to a source document.
That's true whether the work is done by an analyst or an AI.
The best investment teams won't replace judgement with AI.
They'll use AI to make evidence easier to find, easier to verify and harder to misinterpret.
That's a much more valuable role.
We love this technology. We just love it more when it tells the truth, including the truth that it does not know.
Keep reading
- How investors use AI to screen Pitch Decks
- Is it safe to upload a Data Room to AI?
- How to Build an Investor Data Room
*Want AI that cites its sources and admits the gaps instead of inventing them? Run your first deal free in askRIA and see grounded diligence in action.*
FAQs
- Can ChatGPT perform due diligence?
ChatGPT can summarise documents, explain concepts and identify themes, but it should not be relied upon to perform end-to-end due diligence. Generic language models are designed to generate plausible responses, not to distinguish between verified information and missing evidence.
2. Why does ChatGPT hallucinate during financial analysis?
When information is incomplete, a language model predicts the most likely answer based on patterns in its training rather than refusing to answer. In financial due diligence, this can result in invented figures or unsupported conclusions if outputs aren't verified against source documents.
3. What is grounded AI?
Grounded AI generates responses using only approved source documents and links every conclusion back to supporting evidence. Rather than filling gaps with assumptions, it explicitly identifies missing information and uncertainty.
4. Is specialised AI better than ChatGPT for due diligence?
Purpose-built diligence platforms are designed around verification rather than generation. They cite evidence, flag inconsistencies, identify missing documents and separate confirmed facts from assumptions, making them more suitable for investment workflows.
5. Can investment firms use AI safely?
Yes, provided the AI is designed for evidence-based analysis. The safest systems don't replace analyst judgement, they accelerate document review while ensuring every conclusion can be traced back to a verifiable source.