Agentic AI
Where people step into an agentic flow
An agent works in steps. At each step, a person can set limits, check, pause, correct or stop it.
The flow, and where people step in
Read the flow as a list
- Set limits, before the goal.
- Approve the plan.
- Pause or stop while it acts.
- Review the output before release.
- Override and correct the effect.
- Escalate the effect.
- Feed back from the effect to the goal.
Seven ways to step in
-
1 Set limits
Decide what the agent may do alone.
Amira · Product Owner
A written list of allowed actions in the team agreement.
-
2 Approve the plan
Check the agent's plan before it acts.
Jannik · Developer
An approval step before any change goes live.
-
3 Pause or stop
Halt the agent safely mid-task.
Anyone in the team
A visible stop button that leaves things in a safe state.
-
4 Review the output
Read it before a citizen does.
Lena · UX/UI Designer
Citizen-facing messages wait for review.
-
5 Override and correct
Replace a wrong result and record why.
Sabine · Caseworker
An override with a short reason field.
-
6 Escalate
Hand the case to the right person.
Sabine · Caseworker
An agreed route, with a named person and a time frame.
-
7 Feed back and improve
Turn corrections into better rules and learning.
The whole team
Overrides are reviewed regularly and go to the shared library.
How close a person stays
- Suggests only
- Person approves
- Person watches and can stop
- Fully autonomous
Many Hands default for citizen-facing steps
For anything that affects a citizen, the team chooses one of the first three, and writes it down.
What the EU AI Act expects
For high-risk AI, the people overseeing it must be able to:
- understand its limits;
- interpret its output;
- stay aware of over-trusting it (automation bias);
- decide not to use its output;
- intervene, or stop it with a “stop button” or similar.
EU AI Act Service Desk, Article 14 (external site)
Application dates for high-risk rules are being phased in.
Each of these is a skill. Role paths teach it, role by role.
Why it matters
In my research, staff saw AI errors but had no way to stop them. Agents act faster, so the moments to step in must be designed and practised, not assumed.