What Investors Should Ask Management About AI

Management teams have learned that investors expect an AI narrative, and most can now deliver one. The difficulty for an investor is that a well-constructed narrative and a genuine operating capability sound similar in a two-hour meeting.

The questions below are useful because they are difficult to answer well without having actually done the work, and because the way they are answered is often more informative than the answer itself.

On what is actually running

• Which AI applications are in production today, and who uses them as part of their normal job?
• What happened to the initiatives that were started and are no longer active?
• What does the system do when it is wrong, and how do you find out?
• If we removed these systems tomorrow, what would change operationally?

A management team with real deployments will answer these quickly and with specifics, including failures. A team that is positioning will move toward roadmap language, vendor names, and future phases.

On data and systems

• Where does the data these applications depend on live, and who owns its quality?
• How much of your operational data is accessible without a manual export?
• What integration work was required, and what remains?
• Which parts of the estate would you not attempt to build on today?

On capability and dependency

• Who internally can maintain and extend this work?
• What is the dependency on specific individuals, contractors, or vendors?
• What contractual terms govern your model and platform providers, and how portable is the work?
• What is the current run-rate cost, and how does it scale with volume?

Concentration risk in AI capability is frequently held by one or two people whose names do not appear in the management presentation.

On governance and risk

• What data is being sent to third-party providers, and under what terms?
• How are permissions scoped for any system that takes action rather than producing output?
• What review occurred before customer-facing or regulated use, and who approved it?
• What would a material AI-related incident look like here, and what would it cost?

The most useful question

After the structured questions, one open-ended question tends to produce the clearest signal: what did you try that did not work, and what did you learn from it?

Teams that have genuinely operated this technology have a real answer, usually a specific and slightly frustrating one. Teams that have not will either have nothing to report or will describe a failure that reflects well on them. The difference is worth more than most of the material in the deck.