# The trust ladder

By Tommy Magee

Type: Framework
Canonical: https://tommymagee.com/resources/trust-ladder

Four rungs of authority, from drafting to acting alone, and the evidence that moves an agent up or down.

**Use this when:** An agent is doing useful work, and you need to decide how much to let it do without asking.

*This one looks ahead. It builds on one line from my essay on company-issued AI agents, that drafting a message well doesn’t earn the right to send it, and describes how I think trust in agents will have to work in large organizations.*

## Why it exists

Trust in an agent is usually all or nothing. Either every output is checked forever, or the agent is given a wide permission because the demo went well. Drafting a message well doesn’t earn the right to send it. Responsibility should grow as the work holds up, one step at a time.

## The framework

1. **Draft: Can it prepare while a person does?** The agent prepares the work, and a person does the action. Nothing leaves without the person. Most agents start here.
2. **Suggest: Can it recommend with its reasoning?** The agent recommends, shows how it got there, and a person decides. It starts to carry some of the thinking, while the person keeps the decision.
3. **Act with review: Can it act while a person reviews?** The agent acts, and a person approves it first or reviews it afterward. It works for actions that are easy to check or undo.
4. **Act alone: Can it act within limits?** The agent acts on its own inside defined limits. Everything is logged and a sample is checked. It should be the rarest rung.

## How to use it

1. Place every agent on a rung, and write down what it can do there.
2. Name the owner, who is willing to sign for the next rung.
3. Collect evidence at the current rung: reviewed volume, error rate, and how severe the errors were.
4. Move up one rung at a time, and only when the evidence and the reversibility of the action support it.
5. Move down when the conditions below appear, without treating it as a failure.

## What moves an agent up

- A large volume of reviewed work with few errors.
- A low error rate, and errors that were minor when they happened.
- An action that is easy to reverse.
- An owner willing to sign for the next rung.

## What moves it down

- An error with real consequences.
- A change in the source data or the guidance it relies on.
- The owner leaves.

## What it looks like

Take an agent that handles vendor email. At first it drafts replies and a person sends them. After a stretch of reviewed work, it suggests a reply with its reasoning, and the person decides. Later it sends routine acknowledgments by itself, while a person approves anything with a commitment in it. If the vendor’s terms change, it steps back down until the guidance is updated.

## Where it breaks down

The rungs are a way to talk about authority, and they don’t replace judgment. Evidence from one kind of task doesn’t transfer to another, so an agent can be on different rungs for different work. Reversibility matters more than volume. And an owner who signs without checking has only moved the risk somewhere else.

## Try it with AI

```
Help me decide how much authority to give an agent.

The rungs are: draft (it prepares, a person does), suggest (it recommends with its reasoning, a person decides), act with review (it acts, a person approves or reviews), and act alone (within limits, logged, sampled).

Ask me one question at a time about the agent: what it does, who owns it, how much of its work has been reviewed, how often it has been wrong and how badly, and whether its actions can be undone. Then place it on a rung, say what evidence would move it up, and what would move it down.
```

Topics: AI agents, AI governance, Delegated authority
