Implementation worksheet · 5 min read
A Human Approval Boundary for Marketing Agents
Draw the boundary by blast radius and reversibility, not by how clever the action is. Auto-approve reversible, bounded actions: adjusting send time inside an approved window, choosing between approved message variants, pausing an underperforming variant. Require one human for anything that reaches new people or changes what is said: new audience definitions, new copy, channel additions, budget or volume increases. Require two for the irreversible and the regulated: anything touching suppression or consent, anything with a legal or financial claim, and any send above a volume threshold you set in advance. Write the table, name the approver per row as a role rather than a person, and log every automated action with enough context to reconstruct why it happened.
Agentic marketing tools are adopted for speed and then quietly constrained after the first surprise. Drawing the boundary in advance keeps the speed where it is safe and puts a human exactly where the cost of being wrong is high.
Put it into practice
1. Classify by blast radius and reversibility
Two questions per action: how many people does it reach, and can it be undone after it lands? An email cannot be unsent — that single fact sets most of the table.
2. Auto-approve the bounded and reversible
Send-time optimisation inside an approved window, variant selection among approved copy, pausing a losing variant. These are the actions worth automating and the reason to adopt an agent at all.
3. Require a human for new reach or new words
New segments, new copy, new channels, higher volumes. The agent proposes; a person approves. This is where most of the value and most of the risk sit together.
4. Require two for irreversible and regulated
Suppression and consent changes, legal or financial claims, sends above your volume threshold. Two approvers is not bureaucracy here — it is the only control that survives one person having a bad afternoon.
5. Log every automated action with its reason
Actor (which agent), timestamp, what changed, what triggered it, and what it would have done otherwise. Without the reason, an agent's decision is unauditable and you will eventually be asked to explain one.
Approval boundary
Copy this structure into your review document and record your observed result for each row.
| Action | Blast radius | Reversible? | Gate |
|---|---|---|---|
| Send-time shift in window | same audience | yes | auto |
| Variant selection (approved copy) | same audience | yes | auto |
| Pause losing variant | fewer people | yes | auto |
| New audience definition | new people | no | one human |
| New copy | existing audience | no | one human |
| Add a channel | new surface | no | one human |
| Suppression / consent change | compliance | no | two humans |
| Send above volume threshold | large | no | two humans |
A failure worth checking
The silently widened audience: an agent optimising for reach relaxes a segment rule, the audience triples, and the send goes out before anyone notices — because 'audience definition' sat in the auto column on the grounds that it was only a filter change. Reach is the one dimension where automation should always stop for a human, regardless of how small the edit looks in the diff.
Common questions
Does this slow the agent down to uselessness?
The three auto rows are the ones that run constantly; the gated rows happen weekly at most. Structured this way, you keep almost all of the speed and almost none of the exposure.
What if we have only one marketer?
Then the two-human rows need a second approver from outside marketing — a founder, or whoever owns compliance. If a second approver genuinely does not exist, do not enable the irreversible actions.
Basis and scope
This is a proposed implementation method using illustrative examples, not a measured benchmark or a customer case study. Prepared with AI assistance. Validate product-specific behavior against current documentation and your own test environment.