Separate preparation from authority
Mark every research item VERIFIED, INFERRED or UNKNOWN. Require a human review before sending outreach, agreeing terms or making a commercial commitment.
- Map the sales stages
- Assign one preparation job
- Supply approved evidence
- Review the draft and risks
- Record quality issues before expanding
Define one preparation job before choosing a model
Choose a bounded task such as preparing an account brief from approved public sources or drafting follow-up notes from an authorised meeting record. Define the input, output fields, permitted sources and reviewer. Do not start with an instruction to 'run sales': that hides several different decisions with different consequences.
Use VERIFIED for a sourced fact, INFERRED for an interpretation and UNKNOWN where evidence is missing. Require the source and date beside important claims. A fluent sentence is not proof, and an old company page may not establish who holds a role today.
Illustrative example: a reviewed follow-up draft
A salesperson supplies an approved meeting summary and asks for a draft covering the agreed problem, open questions and next step. The assistant must not add an unapproved price, delivery date or customer promise. If those details are missing, it should flag them for the salesperson instead of making the message sound complete.
The salesperson checks the draft against the meeting record before sending. Keep sending authority outside the drafting task. The same separation applies to changes in a CRM: preparing a proposed update and applying it are different permissions.
Test messy inputs before extending access
Try a normal example, a contradictory source and an input missing a commercially important detail. Record omissions, unsupported assertions and correction time. Compare the reviewed output with the existing process; a fast first draft is not useful if checking it takes longer than writing it.
Use approved data-handling arrangements and the minimum information needed for the task. Avoid copying customer records or private negotiations into an unapproved AI service. Expand the role only when you understand its failure modes and the reviewer has a practical way to detect them.