AI · 8 min read
Agents need boundaries, not personalities
Every failed AI deployment we've inherited had a charming assistant and no permission model.
Takeaways
- Scope, permissions and escalation beat persona work every time.
- Run six weeks in shadow mode before the first automated action.
- Every agent needs a human route it cannot refuse to take.
01
The failure is architectural
Teams spend weeks on tone and minutes on scope. Then the agent confidently refunds a customer it shouldn't have, and the project is quietly shelved.
Define the blast radius first: what can this thing read, what can it write, and what must always route to a person.
“Define the blast radius before you define the tone of voice.”
02
Shadow mode is non-negotiable
Run the agent alongside the humans for a month. Log every decision it would have made. Compare. Only then hand over the keys, one request type at a time.
03
Make every decision reversible
If you cannot undo an agent action in one click, it should not be automated yet. Reversibility is what buys you organisational permission to keep going.
Keep reading