Agentic AI1 min read
Useful AI agents keep a human in the loop
A simple four-rung ladder for deciding what an AI agent should be allowed to do on its own, and what it must leave to a person.
“Agentic AI” is everywhere, and so are demos of agents doing everything on their own. In real businesses, the agents that last are the ones people trust. Trust comes from being clear about what the agent may and may not do.
We use a simple four-rung ladder.
| Rung | What the agent does | Who is accountable |
|---|---|---|
| 1. Capture | Gets data in with less typing: reads an email, fills a record | The person, assisted |
| 2. Surface | Finds and summarises on request | The person asks, the agent retrieves |
| 3. Interpret | Explains a result that rules or calculations produced | The agent narrates and never sets the value |
| 4. Decide | Chooses a course of action | Always a human |
Most of the value sits on rungs 1–3. Rung 4 is where an agent can propose and rank options, but a person approves.
Guardrails we build into every agent
- Security-trimmed: an agent only sees what the user it acts for can see.
- Explainable: every output cites the records or documents it used.
- Approval on writes: nothing important is committed without a confirmation step.
- Auditable: AI-assisted content is marked as such, with the confirming user.
- An off switch: every feature can be disabled per environment.
This is the same approach we take with the AI features in Kaizen PPM, and with the agents we build for clients.
Have a process you think an agent could take over? Tell us about it.