ChatGPT in the boardroom: the questions founders now ask

Since the end of 2022 my conversations with clients have changed. Where AI used to appear on the agenda as "we should probably do something like that too", founders now arrive with specific questions. They have tried ChatGPT themselves. That changes the nature of the conversation.

I have noticed that behind very different questions there is the same fork in the road. It is worth naming it explicitly.

Three questions that come up most often

The first: "Can we build something like this on our own data?" This is a question about enterprise search, an internal assistant, a support chatbot. Technically - yes, it is possible. The question is what "our data" actually means and what state it is in.

The second: "Why can't our analysts answer questions as quickly?" This is not a question about AI. It is a question about data architecture and how accessible data is inside the company.

The third: "How much does it cost and how fast can we launch?" This is the hardest question, because the right answer depends heavily on what already exists.

What these questions really mean

Behind the first question is a hope for a quick result. Models that can work with text do exist. But between a public demo and an internal corporate tool there is a large gap. Corporate data is fragmented, unlabelled, and unstandardised. The model will be as chaotic as the data it works on.

Behind the second question is a real problem with data accessibility. If an analyst spends three days assembling numbers to answer one question from a manager - that is not the analyst's problem. That is an infrastructure problem.

Behind the third question is a misunderstanding of how much work comes before the model.

Where to actually start

I ask clients to work through three steps before talking about technology at all.

The first step is a data inventory. What data exists, where it physically lives, who is responsible for it. This is uninteresting work, which is exactly why it is almost never done.

The second step is a quality assessment. Not in the sense of "are the numbers right", but in terms of structure: are the same entities named consistently across different systems, is there history, are objects linked to each other.

The third step is defining a specific scenario. Not "let's do AI", but "this is the task, these are the users, this is the decision they need to reach faster".

Only after that does it make sense to talk about which technology might fit.

Questions to test readiness

Before moving forward, it is worth answering honestly:

• If a data specialist joined tomorrow, could they understand our data sources without a month of onboarding?
• Do we have documentation for the data we want to use for AI?
• Who in the company will be able to verify whether the AI system is answering questions correctly?
• What happens if the AI gives a wrong answer - who will catch it, and how?

If none of these questions have confident answers, starting with the model is the wrong move.