AI Agent Permission and Containment Review (Handoff)

AI Agent Permission and Containment Review (Handoff)

1. 01

   Inventory every sales automation agent with access to a client account

   List each connected worker (outbound prospecting agents, inbound qualifiers, conversational follow-up bots) alongside the CRM, mailbox, and dialer permissions it currently holds.

2. 02

   Map each agent to the minimum access its defined task requires

   An outbound sequencer needs send and read access to its own mailbox, not admin rights on the client's shared inbox or billing records. Strip everything beyond the task scope before the account changes hands.

3. 03

   Confirm human escalation paths exist for every conversation the agent can start

   Forrester's 2027 predictions put human connection at the center of customer success, and agencies erode client trust when automation replaces relationship-critical touchpoints rather than transactional ones.

4. 04

   Test containment with a deliberately out-of-scope request

   Send the agent a prompt that asks it to touch a resource outside its remit and confirm it refuses or escalates. Documented agent boundary failures at OpenAI, Anthropic, Google, and Meta show containment is a systemic risk, not a fringe scenario.

5. 05

   Set per-agent spend and volume ceilings before handoff

   Tag consumption by client so token and seat costs are billed against actuals rather than estimates, and cap daily send volume so a runaway sequence cannot burn a client's domain.

6. 06

   Record the permission matrix in the client handoff document

   Include agent names, granted scopes, escalation contacts, and the date of the last containment test so the client's own IT or compliance team can audit the setup without a call.

7. 07

   Schedule a 30-day permission re-review

   Access drifts as sequences get edited and new integrations get added. Put a recurring review on the account calendar and assign a named owner.