Her job is hard to put a single label on. Not quite a designer, not quite a marketer. She sits at the intersection of both, with a compliance lens over everything. Packaging usually isn’t reinvented from scratch each time, it’s adapted: same product, different market, different language. But adaptation is where things get quietly dangerous. A word that’s neutral in English can sound like a promise in another language. A sentence that fits the layout in one length turns vague or overstated once it’s shortened to fit another. None of that is a translation error exactly. It’s the gap between what a sentence says and what it starts to imply once it’s squeezed into a different shape.
I didn’t find this by looking for a use case. I found it by asking what part of her week she’d quietly rather not do. That question took a while to actually answer. A task done long enough stops feeling like a problem and starts feeling like just the job.
The first version of what we built wasn’t it. It tried to flag too much, and everything looked urgent, which is its own kind of useless. It also missed the plainest thing: it needs to actually read the design like a person would, not just the text pulled out of it, since a claim can be technically fine and still sit next to an image that changes what it means. Trimming it down, and getting the reading part right, took longer than building the first version did.
What it does now is small on purpose. Upload the draft, and it reads the design, checks the claims against the relevant guidelines, and sends back a short report, three flags at most, ranked by how much they matter.
It doesn’t decide anything. It can’t. Too much of this is judgment: what a market will tolerate, what a claim implies next to a specific image, what’s worth a fight with a supplier and what isn’t. She still signs off on every version. What changed is what lands on her desk first. Most of the time now it’s close to clean, one small thing to check, instead of a long pass through every version by hand.
The agent itself is not finished. It doesn’t look particularly polished yet, and it probably won’t for a while, since it keeps getting nudged forward in small ways rather than built out all at once. And yet it has already become one of the most used agents we’ve built, and likely the most liked piece of AI work in that company so far. Not because it does the most. Because it removed something that had been quietly wearing people down for years.
