Picture a tailor shop. A few months ago the owner handed alteration quotes over to an agent: someone sends a photo of a jacket and a request, the agent reads the fabric type and the job from past quotes and drafts a price. It got good fast. Now she barely reads past the total before she taps approve.
Here's the uncomfortable question behind that convenience: if it broke tonight, could she still price a jacket the way she used to, from nothing?
That's not a hypothetical worry. It's the actual finding, twice over, in two pieces of research this year, and it points at something specific and useful: which part of the work is worth protecting, and which part was never yours to keep.
What a colonoscopy study has to do with your quotes
In late 2025, researchers published a study in The Lancet Gastroenterology and Hepatology that followed endoscopy specialists using an AI system that flags precancerous growths during colonoscopies. Some days the AI ran, some days it didn't, on the same doctors.
Before the AI arrived, these specialists caught the growth in question about 28% of the time. After months of working alongside the AI, on the days it was switched off, that rate dropped to roughly 22%. Not because they'd gotten worse doctors. Because the muscle they used to build that 28% had quietly stopped getting exercised.
The study's own co-author said more research is needed before anyone treats this as settled science, and that caveat matters (more on that below). But the mechanism it points at is not exotic. It's the same reason you forget a phone number once your phone remembers it for you.
The part that erodes isn't the doing
A second study, this one a randomized trial of 52 software engineers doing basic coding tasks, found something more precise. Half the engineers used an AI assistant, half didn't. Afterwards, everyone took a quiz on the reasoning behind the code they'd written.
The AI-assisted group scored 50%. The unassisted group scored 67%. But here's the useful detail: the AI group hadn't written worse code. They'd written fine code. What they couldn't do afterward was explain why a particular line worked, or spot the error when something went wrong. They'd kept the output and lost the reasoning.
That's the actual shape of the risk for a small business, and it's narrower than "AI makes you dumb." Your agent drafting the tailor shop's alteration quote isn't costing anyone the ability to type a price. It's quietly costing the habit of asking why that price, for that fabric, for that job. And that's exactly the judgment needed on the one job in fifty the agent prices wrong, the tricky silk repair it's never seen before.
Workers already feel this, even before the studies land
A global survey of 2,500 workers by the software firm GoTo, reported by HR Dive, found 39% saying their reliance on AI had weakened their own skills, and half admitting they'd come to depend on it too heavily. Among Gen Z workers specifically, 46% said the same. That's a feeling, not a lab result, and feelings aren't proof. But a workforce that already suspects this about itself is worth listening to before the peer-reviewed version catches up.
Honest limits
Neither study is about someone who checks in on a task now and then. The colonoscopy doctors and the software engineers were doing the work daily, at volume, the way a radiologist or a full-time developer does. If you approve a dozen quotes a week, you are not on the same curve as someone reading a scan every hour of every shift. The engineering study is also a preprint, not yet peer-reviewed, so treat its exact numbers as a strong signal, not a settled fact.
And the skill that erodes is specific: judgment on the reasoning behind a decision, not every skill you have. The tailor's actual hands, the ones that can tell how a silk will fray before they've made a single cut, aren't at risk because a quoting agent exists. The risk sits precisely where the agent's decision-making and hers used to be the same job: reading the fabric and the customer and landing on a fair number.
The one habit that keeps it yours
You don't need to stop using the agent, and you don't need to read every output line by line either, that's a different question with its own answer. What the research points to here is smaller and cheaper: once a month, pick one quote or one reply and price it yourself, from nothing, before you look at what the agent would have said. Then compare.
Most of the time you'll agree with it, and that's fine, that's the system working. Once in a while you'll catch a reason it got right for the wrong logic, or a case it would have gotten quietly wrong. That's the whole point. You're not checking its homework. You're keeping your own hand in, on purpose, so the judgment is still there the day the agent hits a case nobody trained it for.
Clara F.