Good first candidates
- Drafting data-quality and QA checks for review.
- Summarising reports and anomalies.
- Generating documentation and test cases.
- Preparing, not sending, campaign drafts.
Guardrails
- Read-only access first; add write actions deliberately.
- Approval steps for anything customer-facing.
- Allow-listed tools and scoped credentials.
- Audit logs and rollback plans.
What to avoid
- Giving an agent broad production credentials.
- Skipping evaluation because a demo looked good.
- Automating decisions nobody owns.
Frequently asked questions
Is MCP required for agentic AI?
No. It is one way to connect models to tools in a standard manner; the guardrails matter more than the protocol.
Can agents run journeys on their own?
They can assist, but customer-impacting actions should keep human approval and clear limits.
Last updated: