The moment I remember most from the Herman Scheer engagement wasn't in a concepting session. It was a Thursday evening, walking past a producer's desk on the way out. She had a retailer spec sheet open on one monitor and a folder of approved packaging hero shots on the other, and she was exporting, renaming and uploading them one at a time into a retail partner's portal. Different crop for every placement. A file naming convention that only that retailer uses. A form field for each asset that asked for the same SKU she had already typed forty times.
I asked how often she did that. "Every launch," she said. "It's the part I leave until everyone's gone." Nobody in the room had thought to point AI at it, because nobody thought of it as the work. It was the stuff around the work.
That team went on to an estimated $4.5 million in year-one savings, and I'd argue a real share of the early momentum came from exactly this kind of task. Not because it was hard. Because everyone hated it, and the hate was costing hours nobody tracked.
The misallocation I see in almost every team
When a creative team starts with AI, they aim it at the thing they're proudest of. Concepts. Hero images. Headlines. That's understandable. It's also where the team has the strongest opinions, the highest bar, and the most resistance. Your best art director is going to fight the model on composition for a week before trusting it.
Meanwhile the work that drains the calendar sits untouched. Resizing a master into 30 placements. Renaming exports to match a client's DAM. Filling out the trafficking sheet. Pulling usage terms out of a talent contract and pasting them into a tracker. Checking that every link in a launch email actually goes somewhere.
None of that is craft. All of it is time. And because nobody bills against it cleanly, nobody measures it, so nobody fights for it to change.
Agents that click are the missing piece
This is where the conversation has shifted in the last few months. Every's Context Window ran a piece arguing that most people are sleeping on computer use, the ability for a model to drive a browser or desktop app the way a person does. Their point, which I agree with, is that the value isn't in hard tasks. It's in annoying ones. Once that's available, the reflex becomes, in their words, "I'll ask the agent to do that."
For a creative team, that reflex matters more than it does for most. So much of production lives inside portals and tools with no API: retailer upload systems, ad platforms, client asset libraries, vendor booking forms. For years the answer to "can we automate this?" was "not without an engineer." Now the honest answer is often "probably, if someone writes down the steps."
That last clause is the whole game.
Write the steps before you hand them off
Here is what we did with that producer, and what I now run in the first week of most engagements.
List the avoided work. I ask every person on the team one question: what do you leave until last? Not what's hardest. What do you put off. The answers are remarkably consistent across in-house teams and agencies. Exports, uploads, renaming, form filling, status updates, chasing approvals.
Pick one and write it like a brief. The producer wrote out the retailer upload as if she were handing it to a new assistant on their first day. Which folder. Which crop for which placement. The naming pattern with two examples. The fields in the portal and where each value comes from. What "done" looks like. It took her about twenty minutes and it was the first time that process had ever existed outside her head.
Run it with a human watching. The first pass is supervised. An agent drives the browser, she watches, she corrects. The corrections go back into the written steps. By the third run she was checking the output instead of producing it.
Keep the brief, not the run. The written process is the asset. Models change every few months. A clear set of steps for your retailer upload survives all of them, and it means the next producer on the team can hand it off on day one instead of inheriting a Friday night ritual.
Why this goes first, not last
Three reasons I now push teams to start here rather than treat it as an afterthought.
First, nobody's identity is attached to it. You don't have to convince a designer that the machine has taste. You just have to show them they'll get their Thursday evening back. Adoption on avoided work is close to instant, and that early win buys you patience when you move into the craft.
Second, it teaches the right skill. Writing steps clearly enough for an agent to follow is the same muscle as writing a brief clearly enough for a model to produce good work. Teams that get good at the boring handoffs get good at the creative ones faster.
Third, it's where the hidden hours are. When we report an average 90% reduction in production time across our clients, people picture image generation. A real part of that number lives in the unglamorous layer between "approved" and "live," the layer that used to require a person clicking through a portal one asset at a time.
What not to do
Don't let one enthusiast build a private pile of automations nobody else can see. That's how you end up with a process that breaks the week they go on holiday. The written steps live somewhere shared, with an owner, and anyone on the team can run them.
Don't skip the supervised runs. Agents that click will occasionally click the wrong thing, and a portal upload to the wrong SKU is a bad afternoon. Watch the first few. Then spot check.
And don't treat this as a side project for ops. This is the operating model. If AI only touches the work your team enjoys, you've made two people faster at the fun part. If it touches the work they avoid, you've changed how the whole team spends its week.
If you want to find your team's version of that Thursday evening, the half-day Audit Workshop is built to surface exactly this: where the hours actually go, and which of them should go first. The 8-week Operating Model is where we build the handoffs into how the team runs.