The Team Brain · Module 1: The questions
Sort the list
With the tally done, sort. Group the phrasings into distinct questions. Then, for each, ask three things. Does it have a stable answer, or does it depend on judgement? Is the answer written down anywhere, or only in your head? And how costly is a wrong answer?
The assistant is for questions with stable answers, ideally already written down somewhere, where a wrong answer is embarrassing rather than dangerous. "What is the expenses limit for meals" qualifies. "Should I escalate this client complaint" does not, and should stay with a human.
This sort usually leaves you with ten to twenty questions for the first version. That is the right size. Build for those, get them right, and extend later.
| Question type | Assistant or human |
|---|---|
| Stable answer, written down somewhere | Assistant, first priority |
| Stable answer, only in your head | Assistant, once you write it down (Module 02) |
| Depends on judgement or context | Human. The assistant can say who to ask |
| Wrong answer would be dangerous | Human, always. The assistant should refuse and redirect |
In practice. A finance team's tally included "which cost code do I use for X" (stable, documented, assistant) and "can I approve this invoice" (judgement, human). The assistant was built for the first and instructed to answer the second with "that is an approval decision; ask your budget holder." Both were correct outcomes.
At your desk. Sort your tally into the four types. List the questions for the first version of the assistant. Write down, separately, the questions it must redirect to a human.
Write it down.
- Questions in version one:
- Questions that must go to a human:
- The one I was surprised to find was stable: