Every investor with a pulse is now using LLMs to read decks faster. Most are doing it badly. They paste a deck in, ask the model what they think, and treat the answer as analysis.
The model is not analyzing. It's summarizing. There's a difference, and getting it wrong wastes the time the tool was supposed to save.
What LLMs are actually good at
Reading speed. A model can ingest a fifteen-page deck, a financial spreadsheet, and a market research report in seconds and pull out the structural facts. Who are the founders. What's the round size. What's the claimed growth rate. What's the market sizing methodology. These are extraction tasks, and they're now nearly free.
Pattern matching against the model's training data. If a founder claims a market size based on a methodology that's known to be inflated, a good prompt will surface that. If the financial projections rest on assumptions that contradict the team's claimed traction, a good prompt will flag the inconsistency.
Drafting follow-up questions. After reading a deck, an LLM can generate a tight list of the ten most important questions to ask the founder. This saves the partner from staring at a blank document for twenty minutes.
What LLMs are bad at
Judging a founder. A model can summarize a LinkedIn profile, but it can't tell you whether this specific person is the right one to bet $500K on. The thing that makes founders investable is mostly outside what's written down anywhere.
Reading a room. The signal in a partner meeting comes from how the founder answers, what they avoid, what they get visibly excited about. None of that survives a transcript well, and no LLM is going to surface it from a deck.
Recognizing when a TAM is a fantasy. Models can spot known-bad methodologies, but they will happily summarize an inflated market size as if it were a real estimate, because they're not equipped to do the original research that would falsify it.
The model can answer "what does this say." It cannot answer "should I believe it."
How to actually use it
Use the model for the parts of diligence that don't require judgment. Document extraction. Cross-referencing claims against publicly available data. Drafting initial summaries that a human will then evaluate. First-pass financial sanity checks.
Reserve human attention for the parts where judgment is the entire point. The founder meeting. The reference calls. The market thesis that requires you to disagree with the consensus.
A reasonable workflow: model does the first pass on every inbound deck in twenty seconds. Surfaces the structural facts. Flags inconsistencies. Drafts follow-ups. The partner spends their thirty minutes on the meeting, not on summarizing the deck.
The trap
The trap is treating the model's output as the analysis. Founders are getting better at writing decks that LLMs will summarize favorably. If you outsource your judgment to the model, you'll fund the founders who optimized for the model, not the founders building the best companies.
The model is a tool, not an investor. Use it to free your attention, not to replace it.
What it means
Adopt the model for the boring half of diligence. Protect human time for the half that matters. The asymmetry — machine reading the documents, human reading the people — is the right division of labor.
Investors who get this right will move faster on the early-stage volume while still being deliberate on the close. Investors who get it wrong will either be slow and traditional, or fast and unhelpfully credulous. Neither extreme wins.