How to survive a pitch to investors who use AI to evaluate you.

Torsten Steiner · 07 October 2025 · 4 minute read

Most data rooms have two readers at work. One sits at a desk; the other is one of the large AI models. Investment managers use AI to create smooth sounding pitches (more so than pre-AI). Investors point their own systems at the pitch files and data rooms to test the IMs claims.

What this really changes

Machine-generated analysis now appears on both sides of the table. That creates two common problems. First, slight differences in data cuts or model settings lead to very different conclusions. (Most AI users will not know what settings they use by the way.)

Second, it can be hard to trace confident statements back to a specific paragraph in a specific document. As the standard AI is not doing this in the default mode, yes there are exceptions, but still many AI claims are difficult to trace back.

How investment managers can adjust their pitch

1. Clearly structure the pitch, no heavy formatting: Make every material claim traceable to page and line. Summaries help the AI to understand documents and sections, also stay clear of heavy formatting.

2. Start with a one-page fact sheet: Company, product, market, revenue model, key KPIs, accounting policies, reporting period, data cut-off, version.

3. Use clear, consistent headings: Stable H1/H2 hierarchy and plain section names investors query: Revenue recognition, Customer concentration, Seasonality.

4. Make data machine-readable: Real tables (not images), one metric per column, one time unit per row, ISO dates, units, plus a brief column dictionary.

5. Cite sources and methods inline: Page-and-line references or data-room links beside each claim, with a short note on the method you used.

6. Add an AI support pack: “AI notes” appendix (glossary, caveats, cohort rules), all the metadata you would maybe not include as a human analyst would understand everything without, but this is crucial for an AI.

   
   

How LPs can screen better

You do not need to be an AI expert to get reliable answers from a pitch deck. Ask for a few basics and you will cut through most confusion.

1. Get the numbers in real tables. If a key table is a picture, you cannot sort or re-calculate it. Request the data as a simple CSV or Excel file and a brief explanation of the column names. This lets your team test the figures without guesswork.

2. Spot-check a few claims: This is so crucial for any AI user! Choose three to five important statements (for example churn, customer concentration, covenant headroom) and yes manually trace each one back to the page and line in the source document. If two checks fail compared to the AI, check your prompt, your model and maybe speak to an expert.

3. Set a clear window for follow-ups: Agree a short period for changes after any data-room update and then freeze the pack. Moving targets create unnecessary work and make comparisons harder. (This was also true pre-AI)

4. Respect any clearly marked “human-only” annexes: If some documents are not meant for automated tools, say you understand and request a short call or supervised sampling instead. You will usually get cleaner answers and fewer redactions.

AI is your new audience, let's adjust.

AI is now part of any pitch, whether we like it or not. If you write/present/prepare for both audiences, the person and the model, you reduce misunderstandings, speed up checks and keep the discussion on what matters. If you prefer to limit automated analysis for specific annexes, say so plainly and provide a simple route for human verification. Clarity about structure, sources and intent will do more for AI readability than fancy images which are more for the human audience. So maybe less work for marketing, that might be a good thing.