Paste a deck into ChatGPT or Claude and ask for an analysis. The answer comes back in under a minute, well structured, with the company’s strengths first and a few generic risks at the end. One analyst told us these answers are “always too optimistic.” The bias has three causes, and the prompt can correct much of it.
The deck is the only source, and it was written to persuade
A deck exists to raise money. Its numbers are selected, rounded and shown in their best light, which is normal for a sales document. When the model has nothing but that document, it summarizes it. Every claim comes back in the answer in the founder’s words: “market leader,” “strong traction,” “€40B market.”
The gap between what the company states and what an independent source confirms then disappears. In an investment memo, that gap carries much of the judgment. An MRR of €80k with no customer ledger or bank statement does not weigh the same as a figure reconciled with filed accounts.
Ask the model to sort every figure into three groups: stated in the deck, with the page; confirmed by an outside source, with the link; or calculated, with the formula. Forbid it from repeating an adjective from the deck without the evidence behind it. The answer gets shorter, and every figure keeps its status.
Assistants lean toward the user
Anthropic researchers showed in 2023 that five commercial assistants tailored their answers to the user’s stated views, and that human raters more often preferred answers that agreed with them. Sharma et al., Towards Understanding Sycophancy in Language Models, 2023. In April 2025, OpenAI rolled back a GPT-4o update it described as overly flattering, explaining that it had given too much weight to short-term user feedback. OpenAI, April 29, 2025.
The wording of the question already steers the answer. “What are this company’s strengths?” calls for a list of strengths. “Analyze this deck” usually yields a summary faithful to the document, and therefore favorable.
Start with the case against. Ask for the three strongest reasons an investment committee would turn the deal down, each backed by a page of the deck, then for what would have to be true for it to pass. Keep the business summary for last, once the critical read is done.
The model does not know your fund
Without your thesis, the assistant judges the company in a vacuum. It may see “a great opportunity” in a €15M Series B when your fund writes €500k pre-seed checks, or in an online betting company your fund rules exclude. With the thesis in front of them, an analyst sets these deals aside in seconds.
Paste the thesis as checkable criteria: stage, check size, geography, sectors, exclusions, usual reasons for passing. Ask for one answer per criterion, met, not met or unknown, with the page or source. State that missing information counts as unknown, never as met.
A prompt you can copy
Here is a starting prompt. Adapt the thesis to your fund and review the answer with the deck open beside it.
Prompt
You are an analyst at a venture capital fund. Our thesis: [stage, check size, geography, sectors, exclusions, reasons for passing]. Read the attached deck, then answer in this order.
- List every figure in the deck with its page. Label it: stated by the company, confirmed by an outside source (name it) or calculated (show the calculation).
- For each criterion in our thesis, answer met, not met or unknown, with the page or source. Missing information counts as unknown.
- Give the three strongest reasons not to invest, each backed by a page of the deck.
- Write what would have to be true for the deal to go to committee, and which document would prove it.
- Then summarize the business in five lines.
Do not use any adjective from the deck (“leader,” “unique,” “strong traction”) without the matching evidence. If you do not know, write “unknown.”
What the prompt does not fix
Without web access, the model verifies nothing, whatever the prompt says. With search turned on, it may cite the company’s own press release as confirmation, although a page that repeats the company is not an independent source. Our guide to external research in due diligence covers which registers to check and what each one lets you conclude.
The work also stays in one person’s conversation. Each analyst pastes their own version of the thesis, the answer never reaches the fund’s deal file, and the next deck starts from scratch.
Finally, check your fund’s policy before pasting a confidential deck into a personal account. Depending on the plan and account settings, conversations may be used to train models.
Casian applies these principles to every incoming deck: the fund’s thesis checked criterion by criterion, a source for every figure, the gaps between the deck and public sources, and the opinion kept in the team’s deal file. You can see the result on one of your own decks with the free screening, no account needed.