Chapter Four / Step One
Decide
Agree what the AI should do before anything is built. In plain language. Signed by whoever owns the process.
Skip it and the AI builds to one person’s assumption. Fixing it costs more than doing it right for the first time.
Why it matters
Without a definition, the agent does what someone assumed the process was.
Two people each assume the other owns a step. An exception nobody wrote down never reaches the build. The agent then does exactly what it was told, which is not what the process is.
What good looks like
One page stating what triggers the process, which rules apply, which exceptions exist, and what a correct output looks like, specific enough that two people reading it would build the same thing. It is signed by whoever owns the process: operations, finance, compliance. The signature turns a set of assumptions into a decision someone has made.
It is not a requirements document, and nothing in it should name a system or a model.
One page. Plain language. Signed. Written before anything is built.
A fifty-page PRD. A backlog. A diagram nobody signed.
Figure 4.1
What the agreement looks like
Without it
DROPPED
Both teams believed the other owned the threshold rule. Neither was wrong, because nobody had written it down.
In your firm
Your AI approved 340 vendor invoices last month. Nobody told it what “approved” means.
Finance set a $10K threshold for manager sign-off. Procurement set $5K. The AI was built on Finance’s rule, so for two months invoices between $5K and $10K went through unapproved.
The AI did exactly what it was told. Nobody had agreed what to tell it.