The pitch for AI in an agency is speed. Drafts in minutes, reports in seconds, research overnight. And the speed is real. The problem is what happens when that speed reaches a client unchecked: a claim the brand isn't allowed to make, a tone that's slightly wrong, a number pulled from the wrong account.
None of these are dramatic failures. They're small, and they're exactly the kind of mistake a good account lead would have caught. Clients rarely leave over one of them. They leave over the feeling that nobody is really looking.
Speed is the reason, not the excuse
A human checkpoint is often treated as the thing that slows automation down. It shouldn't be. If the AI does the reading, the checking and the first draft, the person at the end isn't doing the work again. They're making one decision, with everything they need already in front of them.
That changes what the checkpoint looks like. Instead of a reviewer opening a document cold, they see what the system found, the exact text it flagged, what it changed and why. Approving takes seconds when the reasoning is visible. Disagreeing takes seconds too.
Decide what runs on its own, in writing
Not every step needs a person. Logging a lead, filing a draft, updating a sheet: these are low-risk and reversible. The rule we use is simple. Anything that reaches a client waits for a named person by default. Anything else can run on its own, but only once it has been agreed with the agency, in writing.
- Client-facing: always waits for a person.
- Internal and reversible: can run alone, once agreed.
- Anything involving client data leaving your tools: named and approved before it goes live.
Every decision should teach the system
The best part of a well-designed checkpoint is what happens after the decision. When an editor keeps the original instead of the AI's fix, that's information: the flag was wrong for this brand. When they send a draft back, the reason is a rule waiting to be written down.
Capture those decisions and the system gets sharper with use, brand by brand. Skip them and you have an automation that makes the same mistakes forever, with a person quietly fixing them each time.
The goal isn't to remove people from client work. It's to make sure the time they spend there is spent on judgement, not on copying and checking.
If you're automating anything client-facing this year, start by deciding who signs it off, and what they need to see to do that in under a minute. The rest of the build gets easier from there.