Every action an agent can take sits in one of three tiers: recommend, act with approval, or act and log. Anything unlisted defaults to the most restrictive. New agents start with everything on recommend and earn promotions one action type at a time.
Once you decide to let an automated agent act on your marketing, you need a way to decide what it may do on its own and what needs a person. Deciding case by case does not scale. The three action tiers are a simple structure for making that decision once, clearly, for every type of action.
The three tiers
Tier 1: Recommend
The agent never executes. It writes down four things — the trigger, the evidence, the proposed change and the rollback step — and then stops. A person decides whether to act.
Tier 2: Act with approval
The agent prepares the change, captures the state before the change, and sends a clear description of the difference to a person. Then it waits for an explicit yes before executing.
Tier 3: Act and log
The agent executes immediately, inside its guardrails, with the before-state captured so the change can be reversed. The action is logged and visible, but nobody has to approve it first.
Every action in exactly one tier
Every type of action the agent can take is assigned to exactly one tier. An illustrative assignment for a paid media agent might look like this:
| Action type | Tier | Why |
|---|---|---|
| Add a negative keyword | Act and log | Small, common, easily reversed |
| Pause an ad group spending with no conversions | Act and log | Protective and reversible |
| Raise a daily budget one step | Act with approval | Spends more money |
| Shift budget between campaigns | Act with approval | Reversible, but affects several campaigns |
| Change a bidding strategy | Recommend | Large effect, disrupts learning |
| Launch new ad copy | Recommend | Reaches buyers; needs human voice |
Your assignments will differ. The point is that the decision is made in advance, written down, and applied consistently.
Anything unlisted defaults to recommend
If an action type is not explicitly assigned, it defaults to the most restrictive tier. This matters because agents encounter situations nobody anticipated. The safe response to the unexpected is to propose rather than to act.
Two rules that make it safe
The reversibility rule. If you cannot write down the rollback step for an action, it drops to recommend — regardless of which tier it would otherwise be in. This is the operational form of if you cannot undo it, do not automate it, and it keeps one-way doors in human hands automatically.
The ramp rule. A new agent starts with every action type on recommend. Action types are then promoted one at a time — recommend to act-with-approval, then act-with-approval to act-and-log — based on what the log shows. If the agent’s recommendations for adding negative keywords have been consistently right for a month, that action type has earned a promotion. If its budget suggestions have been hit and miss, they stay where they are.
The ramp rule is how trust is extended on evidence rather than on faith. It mirrors how you would treat a capable new colleague: start supervised, and widen responsibility as they demonstrate good judgement.
Silence is never consent
One more rule, and it is more important than it looks. When an action is waiting for approval and the request expires without an answer, that counts as a rejection, not an approval.
The temptation is to design it the other way — “if nobody objects within four hours, proceed” — because it keeps things moving. But it quietly converts act-with-approval into act-and-log whenever people are busy, which is exactly when mistakes are most likely to go unnoticed. An approval tier only means something if approval actually has to be given.
Where to go next
Bounded autonomy explains the operating model the tiers support. Guardrails that make speed safe sets the limits every tier operates inside. And trigger, action, impact is the record the ramp rule reads when deciding on promotions.