The useful question about AI in marketing is not what it can do. It is what you let it decide without asking anyone.
Almost every automation failure we have been called in to clean up was not a model failure. The model did exactly what it was told. Somebody had just handed it a decision that should have required a person.
The test: how fast can you undo it?
Two questions decide where a call belongs. How reversible is it, and how far does the damage travel? Cheap and reversible goes to the machine. Slow or expensive to reverse stays with a person.
| Decision | Reversible in | Blast radius | Owner |
|---|---|---|---|
| Shift budget between two ad sets | Minutes | One campaign | Automated |
| Pause a fatigued creative | Minutes | One ad | Automated |
| Update exclusion and negative lists | Minutes | One account | Automated |
| Launch a new angle in a regulated market | Sometimes never | The whole account | Human |
| Enter a new GEO | Weeks of spend | Budget and team | Human, AI-informed |
| Change positioning or brand voice | Months | Everything | Human |
| Kill a channel | Weeks | Revenue | Human |
Notice that spend size is not the variable. A $50 test that risks a licence is a human decision. A $5,000 budget shift inside guardrails is not.
Four decisions that never touch automation
- Anything a regulator reads. Claims, disclaimers, targeting in licensed markets. A model has no concept of a payment processor dropping you.
- Brand voice and positioning. Models regress toward the average of what already exists. Averageness is the one thing a brand cannot afford.
- Spend outside the agreed envelope. Inside the guardrails, automate freely. To move the guardrails, ask a person.
- Structural changes. New channels, new markets, restructuring accounts. Data informs these. It does not make them.
What should be automated, and usually is not
The failure we see far more often is the opposite one. Teams keep the mechanical work and automate nothing, then wonder why they cannot scale past a certain spend.
- Bid and budget nudges inside preset guardrails
- Creative rotation when frequency or CPA crosses a fatigue threshold
- Anomaly detection: tracking that stops firing, spend with no conversions, audience overlap climbing
- Reporting, pulled live from the platforms rather than retyped
- Exclusion hygiene: converters, employees, existing customers, bots
- Pacing to hit the month without a Friday scramble
If somebody on your team spends Friday afternoons updating negative keyword lists by hand, that is not a human decision. That is a person doing a machine's job while the actual thinking waits.
The gray zone, and how to write the rule down
Most decisions are not obviously one or the other. The fix is not a philosophy, it is a written threshold. Four things make the split real:
- A numeric guardrail. Ours is no more than a 20% budget change per campaign per day without approval. Pick your number, then hold it.
- A logged reason for every automated change. If you cannot reconstruct why budget moved on a Tuesday, you do not have automation, you have a black box.
- A kill switch anyone can reach. One action halts all automated changes on an account. If stopping the system requires an engineer, it is not safe to run.
- An escalation clock. When the system flags something, name who responds and within how long. Flags nobody owns are decoration.
Why this is a commercial argument, not a philosophical one
If a partner automates everything, you do not need them, you need their tool. If they automate nothing, they cannot run your spend at scale without adding headcount you end up paying for. The value is in drawing the line deliberately and being able to show you exactly where it sits.
So ask any agency pitching AI a specific question: which decisions does your system make on its own, and which ones does a person sign? An honest answer takes about two minutes. Watch what happens if they cannot give one.