Why executive AI isn’t the same as frontline AI

What we see in the field

We deploy the same assistant in a 60-person SMB. Three months later, we look at actual usage.

The executive uses it 4 times a week, on average 8 minutes per session, for cross-cutting summaries. The middle manager, twice a day, on average 3 minutes, to prep meetings. The frontline employee, 15 times a day, on average 90 seconds, for precise micro-questions.

Three rhythms, three expected response formats, three tolerances for error. And a single assistant, configured by default for no one.

Executive AI

Looks for: the trend, the contradiction, the weak signal. “Are we billing better than two years ago?” “Which customers have turned toxic?” “Which past decision didn’t pan out?”

Tolerates: a long answer, with arguments and nuance.

Doesn’t tolerate: getting their time wasted with uncertain numbers. For an executive, an answer that’s “80% right” is a wrong answer.

Manager AI

Looks for: preparation. “Recap of the last three exchanges with this customer”, “Summary of how this deal has progressed”, “Upcoming deadlines”.

Tolerates: bullet-point answers, blunt, with numbers.

Doesn’t tolerate: having to rephrase three times. The manager wants a result on the first try.

Frontline AI

Looks for: the precise answer. “What’s this customer’s payment term?” “Where’s the quote from March 12?” “What’s the procedure for this case?”

Tolerates: a one-line answer, with a link to the source.

Doesn’t tolerate: rambling. Three lines of context before the answer is three lines too many.

How to configure for all three

Three practical options, in order of increasing complexity:

1. One assistant, three personas in the system prompt. The user picks their persona when opening the conversation. Works well up to 30 people.
2. Three distinct assistants, same sources, different prompts. Each user sees their own in their interface. Works well up to 200 people.
3. Three assistants, three subsets of sources, three prompts. The executive has access to HR; the frontline doesn’t. Beyond 200 people, it’s the only viable option.

The worst choice: one assistant, one prompt, and the hope that “it’ll adapt on its own”. It never does.