Put the model where the work breaks

Most AI in marketing teams right now makes existing tasks faster. Drafting copy, summarizing a report, building the deck. That's useful, and it rarely changes the result. The same decisions get made by the same people on the same schedule, with less typing.

The bigger gains are in the steps where work breaks because a person can't be there, can't keep up, or can't remember. Those are the places we look first.

Three places it breaks

The first is volume that resets. A catalog of thousands of products, with prices and availability that change every day, will never be fully covered by a team writing by hand. Descriptions, feed attributes, and search terms go stale the moment they're finished. A model working from live data can keep up, and the team reviews the exceptions.

The second is timing. A home-service shop gets a call while every tech is on a roof. The call goes to voicemail and the customer calls the next shop. Faster slide-making doesn't touch that. An agent answering on the shop's own number does, because the failure was that nobody was available.

The third is memory. A team runs a test, learns something, and forgets it by next quarter. A model can draft hypotheses all day, and that makes the forgetting worse unless each one is written down with a prediction and read back against the result. The record has to come first. The model comes after.

How we decide

Before we build anything with a model in it, we ask what fails today when nobody is watching. If the answer is “nothing, it just takes a while,” the model is a convenience. If the answer is a lost job, a stale catalog, or a lesson nobody kept, the model belongs in the workflow.

Then we ask where it should stop. In Right On Call, the model runs the call. In Marketers Lab, it can suggest a test but can't mark one as won. deepstint has no model at all. Starting work is hard because of one decision too many, and a chat window would add another.