Don’t approve every action, and don’t approve none. Approve the boundary once, let the system act inside it, and audit what it did. That only works if limits, visibility and reversibility are all present together.
Every team that automates marketing eventually hits the same question: how much should the system be allowed to do on its own? And most teams land on one of two answers, both of which fail.
Failure one: approve everything
The cautious answer is to require human approval for every automated action. It feels safe. It also rebuilds exactly the bottleneck you were trying to escape.
Picture it concretely. A campaign starts bleeding money at 2am because a competitor changed their bidding or a landing page broke. The automation spots the problem immediately and proposes the fix. The fix then sits in an approval queue until someone logs in at 9am. Seven hours of spend, gone. The automation did not help you. It generated a to-do list.
Failure two: approve nothing
The ambitious answer is to let the system do whatever it decides. It feels modern. It produces a black box: something that spends your money, makes changes you cannot see, cannot explain its reasoning, and cannot cleanly undo what it did.
When it works, you do not know why. When it fails, you find out late — usually from a spend report or an angry question — and you have no idea which of its hundreds of decisions caused the problem.
The third option
Bounded autonomy sits between the two. The system reasons and acts on its own, but only inside limits you set in advance, and every move it makes is visible and reversible.
The reframe that makes it work is this: you are not approving every action. You are approving the boundary once, then auditing inside it.
That is not a new idea. It is how organisations already delegate to people. A new manager gets authority to spend up to a certain amount without asking, is expected to report what they did, and knows that anything irreversible or unusually large needs a conversation first. Nobody approves every purchase they make. Nobody lets them spend without limit either. Bounded autonomy applies the same logic to a system.
Three properties, together
Bounded autonomy only works when three properties are all present:
- Defined limits. What the system may do, how much, how fast, and what it must never touch. These are the guardrails.
- Full visibility. A readable record of every action — not “optimisations applied”, but what changed, when, why, and with what result. See trigger, action, impact.
- Reversibility. Every action can be undone cleanly, because the state before the change was recorded. See if you cannot undo it, do not automate it.
Drop any one and you are back to a black box in a nicer jacket:
- Limits without visibility — you cannot tell whether the limits held.
- Visibility without reversibility — you get to watch a mistake you cannot fix.
- Reversibility without limits — you can undo the damage, but only after it has been done at scale.
Setting the boundary
Start narrow. A new automation should begin with tight limits: small changes, low spend, a short list of permitted actions. Then widen the boundary based on evidence — what the log shows the system actually did, and whether it did it well.
That progression is formalised in the three action tiers: everything starts as a recommendation, and individual action types are promoted to more autonomy one at a time as the record justifies it. Trust is extended the way it is with people — gradually, and on the basis of what has been demonstrated.
Why it is harder to build
If bounded autonomy is so clearly better, why is it rare? Because it is harder to build than either alternative. A recommendation engine only has to suggest. A rules engine only has to follow instructions. A fully autonomous system only has to act. A bounded system has to act, check its limits every time, record what it did in a form a person can read, and preserve enough information to reverse it. That is more engineering, and it is the reason so few tools genuinely offer it. The spectrum of options is laid out in the execution-depth spectrum.
Where to go next
The execution-depth spectrum shows where bounded autonomy sits among the alternatives. Explain, bound, reverse turns the three properties into questions you can ask any tool. And Campaign Automation is the practice built on this model.