Where agents sit in the lifecycle, what records they may change, and how they are switched off. RevOps decides placement and write access; the function that owns each stage owns the result.
Boxes marked ↗ open the page for that part of the model.
How to read this
- The useful split is by write access, not by task. Agents may write activity, notes, drafts, enrichment and tasks. Anything that moves pipeline or the forecast stays with a person or a system rule.
- Outbound agents draft and a person sends. Inbound agents act live, but only inside a script, and hand off to a person the moment a conversation leaves it.
- An agent is judged on the same funnel metrics as the person it supports, plus an error rate from a weekly sampled audit, so a fast agent that writes bad records shows up.
- Each agent is its own CRM user with its own kill switch, and its instructions change through the same review as routing rules. That is what makes it auditable and safe to turn off.
Where it connects
Diagram source: 23-ai-agents.drawio 23-ai-agents.svg