Digital sales expert agent

Agent for a sales and marketing team: answers from an agreed knowledge base and asks for approval before changing work systems.

Problem

The team needs an AI assistant it can trust: answers come from an agreed knowledge base, and the agent changes nothing without approval.

Value

The team can take the tested agent into use one part at a time. Data and connections open only after approval.

Role

The agent’s parts and the rules for what it may do. Agents wrote most of the code.

  • Five clearly separated parts. The first, content design, has its instructions and sources ready.
  • Airtable actions need approval: only a narrow set of operations, and no deletes.
  • The agent’s instructions are built the same way every time.
  • Built to fit the organisation’s way of developing in a test environment first (Dataverse).

Limits

  • Not yet in use in the organisation: not published, not connected to M365 content or Airtable.
  • Instructions for four of the five parts are still to be written.
  • Employer-related system, described only in general terms. Code and material stay private.

Sources

  • 8 design decisions, 13/13 tests, behaviour and package checks