Explaining AI's Real Limitations to a Skeptical Executive, What Worked

An executive opened a call by asking whether AI could just replace the reporting team entirely — generate the weekly decks, answer stakeholder questions, the works. Saying "no, that's not realistic" would have landed as pessimism from the AI consultant, which is the worst possible position to argue from. So I didn't lead with no.

I led with a live example instead. I pulled up their actual reporting data and asked an LLM, live on the call, a question a stakeholder had asked the week before — one with a slightly ambiguous definition buried in it. The model answered confidently and wrong, because it had resolved the ambiguity in a way that happened to misalign with how their org actually defined the metric. Nobody in that room needed to be told AI has limitations after watching that happen in real time.

Don't tell an exec that AI has limitations. Show them the specific wrong answer it gives on their own data.

What the conversation actually needed

Once the limitation was visible and specific, the conversation shifted productively on its own — not to "should we use AI" but to "where exactly does it earn trust and where does it need a human checking its work." That's a far more useful conversation than an abstract debate about AI capability, and it's the one that actually gets budgeted sensibly.

• Generating a first draft of a report from clean, well-defined data — strong use case, low risk if a human reviews before sending.
• Answering an ambiguous stakeholder question without a human validating the interpretation first — weak use case, demonstrated failure mode.
• Flagging anomalies in a dataset for a human to investigate — strong use case, plays to what the model is actually good at.
• Making a final call on a number that goes in front of the board unreviewed — not yet, and probably not for a while.

The exec didn't leave that call anti-AI. They left with a specific, defensible plan for where it got deployed in the reporting workflow and where a human stayed firmly in the loop. That's a far better outcome than either extreme — blind enthusiasm or blanket skepticism — and it only worked because the limitation was demonstrated, not asserted.