Most small firms already have a model somewhere near the investment decision. Someone pastes a deck into an assistant to get a summary before a first call. Someone asks it for the strongest case against a deal before the partner meeting. Nobody decided this; it arrived one convenience at a time. The question now is whether to make it deliberate: which step the model takes, what it sees, and how much weight its view carries.
That question has three parts. Is the model's judgment worth anything on this kind of material? Where does its judgment differ from the partners', and does the difference matter? And what does the firm give up by handing it founders' documents? I went looking for evidence on the first two before writing a policy, and found a study that tests them against investors directly.
Source
What the study measured
Felipe Csaszar (University of Michigan), Harsh Ketkar (University of Texas at Austin), and Hyunjin Kim (INSEAD) published "Artificial Intelligence and Strategic Decision-Making: Evidence from Entrepreneurs and Investors" in Strategy Science (volume 9, issue 4, pages 322 to 345, 2024). I read the post-review version on arXiv, dated August 1, 2024, which thanks the editor and referees and carries the published abstract. The figures below come from it, and the page and table numbers refer to it.
The paper is partly conceptual, but at its center are two studies, both run with GPT-3.5, the only model at the time with a context window large enough for the long prompts (page 10).
The first tested whether a model could write strategy. The authors took ten business plans (five accepted, five rejected) from the 309 submitted in 2021 and 2022 to a European accelerator that gave each startup 30,000 euros. They had the model write a new plan from each founder's own problem statement (pages 10 and 11). Then 250 U.S. evaluators with experience investing in angel syndicates, venture capital, or private equity rated the plans without knowing which were which. The evaluators averaged five years of investing experience and 16 years of work experience (page 11). To keep polish from deciding the result, the founders' plans were first run through the model to correct spelling and grammar (page 13).
The second tested whether a model could judge. The authors took plans submitted to ten cycles of a business-school startup competition, where judges were venture capitalists and angel investors. Of 731 plans, 138 were written fully in text rather than images; those had 541 evaluations from 137 judges, three to five per plan (pages 15 and 16). The model scored each plan on the same ten-part rubric the judges used, three times, and the three runs were averaged (page 16).
Source
What it found
The model wrote plans investors liked slightly more. Its plans were rated 0.14 standard deviations higher than the founders' plans. Evaluators were 5 percentage points more likely to recommend them for the accelerator, against a base of 57 percent. Stated likelihood of investing rose 2.55 points on a 100-point scale, against a base of about 47 (pages 12 and 13, Table 1).
That advantage sat almost entirely against weak plans. The evaluators' judgment carried real signal: they rated the plans the accelerator had accepted about 0.6 standard deviations higher than those it rejected (page 14, Table 2). The model's versions beat the rejected founders' plans clearly. Against the accepted ones, the difference cancelled out (pages 14 and 15, Table 2).
The model's scores tracked the judges'. Across the 138 competition plans, the correlation between the model's scores and the judges' was 0.52. It held for plans submitted after the model's training data ended as well as before (page 16, Figure 2). The model's scores explained about 29 percent of the variation in the judges' scores (page 17, Table 3).
It agreed with the judges more than the judges agreed with each other. Measured by intraclass correlation, agreement between the model and the average judge was 0.51. Agreement between individual judges was 0.25, a difference significant at the 10 percent level (page 16, footnote 6).
Agreement varied by what was being judged. It was highest on financials (0.57), the team (0.46), the execution plan (0.40), and competitive advantage (0.38). It was lowest on innovativeness (0.21), how well the product meets customer needs (0.24), and return potential (0.25) (page 18, Table 3).
Describe this figure
Two horizontal bar charts on the same scale, 0 to 1. The first shows overall agreement with the venture capital and angel judges of a business-school startup competition: the correlation between the language model's scores and the judges' across 138 plans was 0.52; the intraclass correlation between the model and the average judge was 0.51, the highlighted bar; between individual judges it was 0.25. The second shows agreement between the model and the judges by part of the rubric, highest first: financials 0.57, team 0.46, execution plan 0.40, competitive advantage 0.38, then return potential 0.25, how well the product meets customer needs 0.24, and innovativeness 0.21. The plans were scored by GPT-3.5 three times and averaged; the difference between 0.51 and 0.25 is significant at the 10 percent level; the comparison is with the judges, not with how the companies did. The figures come from the authors' post-review manuscript, page 16, footnote 6, and Table 3 on page 18.
The authors draw two practical implications: investors could use a model to screen ideas at scale, such as the first round of a competition, and founders could use its feedback to improve early plans (page 17). They also warn that with current models, generated ideas are likely to cluster around areas well represented in the training data (page 21).
A second study points the same way on method. Doshi, Bell, Mirzayev, and Vanneste, in the Strategic Management Journal (2025), compared model rankings of business models with human experts' in two samples of 60. They found the model's single evaluations often inconsistent and biased, but its aggregated rankings tended to resemble the experts' (abstract).
Synthesis
What the study does not tell an investor
It compares the model with judges, not with outcomes. Agreement with experienced investors is a meaningful bar, but no one in the study checked which plans became good companies. A model that scores like the judges also inherits whatever the judges get wrong.
It is a written-plan test. Only the 19 percent of competition plans written entirely in text could be scored (page 16), and the task was reading a 1,500-word plan against a rubric. A first meeting, a reference call, or a founder's answer under pressure is not in it.
It used an older model. GPT-3.5 is well behind what firms use now. That cuts both ways: current models may agree more, but nothing here measures them.
Synthesis
What that means for the firm
A model is a credible second reader of written material. Agreement of 0.52 with experienced judges, and closer agreement with the average judge than one judge has with another, is good enough to make a model useful as a consistent first pass on inbound plans. It is not good enough to make it the decision.
Use it where it agrees and look hardest where it does not. The model tracked investors most closely on financials, team, and execution, and least on innovativeness, customer need, and return potential. Those last three are where early-stage judgment earns its keep. I would let a model summarize and score the parts of a plan it reads well, and keep the partner's attention on the parts it reads least well.
Run it more than once and average. The authors averaged three runs; Doshi and colleagues found single answers inconsistent and averaged answers closer to experts. A single model score on a deck is a noisy number. The firm's process should say how many runs make a score, and show the spread.
Expect it to help most at the bottom of the pile. Its advantage in writing plans was against weak ones, not strong ones. The same pattern suggests a model is better at saying "this is not ready" than "this is exceptional."
Use it to argue, not only to agree. The paper shows a model producing 15 contrarian arguments against a strategic plan, and notes that a model does not hesitate to challenge a senior person (page 8). A devil's advocate pass on the investment memo, before the partner meeting, is a low-risk use.
Decide what it may see. Founders send decks in confidence. Before any of them go into a model, the firm should follow every copy of that material and know whose authority the connection uses, as I set out in before an AI assistant can access firm data, and decide whether founders are told.
Proposal
How I would decide and prove it
Start with a blind comparison on the firm's own pipeline. Take 30 to 50 decks the firm has already decided on, and have the model score each three times against the firm's own first-look criteria. Compare its average with what the partners decided and, where enough time has passed, with how the companies did. Record the correlation, and list the companies where the model and the partners disagreed most. That list is more informative than the correlation.
Then set the model's place in writing. It reads and scores every inbound plan against the firm's criteria, and the record shows the average and the spread across runs. It writes a devil's advocate note on each memo that reaches a partner meeting. It does not decline companies; a partner does, and can see the score while doing so. Any change to the model, the prompt, or the criteria restarts the comparison.
Proposal
What I would test next
I would run the blind comparison before changing anything in the workflow. A few hours with past decks will show whether the model agrees with the partners on this firm's deals, where it disagrees, and whether those disagreements cluster on the novel companies the firm most wants to see.
Open question
Should the partner see the score first?
The authors raise, as a question for research, whether showing a model's view before the humans give theirs could reduce some of the biases in group decisions (pages 25 and 26). It could also anchor the partners on the model. For a small partnership, where one voice already carries a lot of weight, I do not know which effect would win. Trying it both ways on a quarter's first looks, and comparing the discussions, is the cheapest way I know to find out.