Grade actions, not agents. Let low-impact, reversible actions run at full speed within tested limits, and concentrate human review on the few actions that move money, change records others rely on, or cannot be undone. Speed comes from not reviewing everything. Control comes from reviewing the right things and checking every action automatically at the moment it happens.
Why do speed and control seem to conflict?
Because many teams have only one control: a person looks at everything. Then every gain in speed is a loss of control, and every new check slows the work. The answer is to separate two kinds of control.
Automatic checks run on every action, in milliseconds: is this within the agent's scope, under the limit, from a trusted source, with a current certificate? These cost almost nothing in speed.
Human review is slow and expensive. Spend it where an automatic check cannot judge, such as unusual cases, high values or irreversible actions.
A system prompt is not either kind. Instructions in the prompt are advice to the model, not a check on the action; see why a system prompt is not enough to control an AI agent.
What are the steps, and who acts?
- Process owner: lists the agent's actions and grades each. Reversible or not, value at stake, who is affected outside the business. Evidence: the graded list.
- Process owner and risk: set the depth per action. Read and draft for most, execute within bounds for routine reversible work, hold for the rest. Evidence: the depth table, signed.
- Platform: enforces the checks on every action. Role scope, limits, sources and certificate are checked automatically. Evidence: a filed verdict per action.
- Approvers: review held items only. Their queue is the slow lane. Evidence: approval records.
- Process owner: tunes the split monthly. If an action is always approved and the agent has passed testing on it, consider moving it to the fast lane. If errors appear, move it back. Every change goes through change control for AI agents. Evidence: the change log.
- Risk: records the control evidence. Results of the checks feed your control library; our guide on recording control evidence in your GRC platform shows how. Evidence: control test results.
What goes in the fast lane and what in the slow lane?
| Action type | Lane | Typical depth |
|---|---|---|
| Reading and summarising | Fast | 1 |
| Drafting a message a person sends | Fast for drafting, person sends | 2 |
| Routine updates within limits, reversible | Fast, checked automatically | 3 |
| Moving money, closing accounts, deleting records | Slow, held for a person unless evidence supports more | 4 if earned |
| Actions the platform withholds | Never | None |
Access design follows the same split. Traditional role-based access control gives a role a set of permissions. For agents, go one step further and grant depth per action within the role.
A worked example
Take an insurer's claims intake agent. Before, every claim it logged waited for a handler to check it, and logging took a day.
The claims manager grades its actions. Logging the claim, requesting missing documents and setting the initial reserve under 2,000 are reversible and routine. After testing on past claims, they run at depth 3. Setting reserves above 2,000, declining cover and paying anything stay with handlers. Logging time falls from a day to minutes. Handlers now see a fraction of the volume and spend their time on the claims that need judgement. Every action, fast or slow, is checked and filed.
What usually tips the balance wrong?
- Holding everything at launch and never revisiting. The agent adds work instead of removing it, and people start rubber-stamping.
- Moving actions to the fast lane under pressure, without test evidence.
- One depth for the whole agent, which forces the slowest action's control onto every action.
- Control that lives in the prompt, which feels fast and is not control.
Which controls let you go faster safely?
- Automatic checks on every action, so speed never bypasses the bounds.
- Depth earned on evidence, so the fast lane is backed by results.
- Fast tightening. Any owner can slow an action instantly if something looks wrong.
- A record of everything, so fast actions can be sampled after the event.
Our answer on how to control what an AI agent is allowed to do covers the bounds in detail, and the topic guide, agent governance is a bound, not a model, explains why the check sits outside the model.
Questions people ask
Does automatic checking slow the agent down?
Not noticeably. Checking an action against scope and limits takes a tiny fraction of the time the model takes to draft it.
Who decides when an action moves to the fast lane?
The process owner, on test evidence, through change control. Risk should be able to challenge the move.
Can speed targets be part of the agent's goals?
They can be measured, but never traded against bounds. An agent that is rewarded for speed should still meet every check on every action.