Why Does an AI-Skills Hire Work Like the Rest of the Team After Six Months?

Why did six months erase what you hired for?

It didn't erase anything. It taught. A newcomer's job in the first months is to work out what this team treats as finished work, and the answer they found is the standard you already had. In a longitudinal study of business school graduates surveyed at four and ten months in new jobs, institutionalized onboarding (the collective, sequenced, run-by-insiders kind) was negatively related to both attempted and actual role innovation ¹.

The same study found those tactics positively related to job satisfaction, organizational commitment and organizational identification ¹. That is the trade, and most of the time it is the one you want: a ramp that produces conformity is a ramp that produces someone who stays, knows the rules and does not have to be told twice. It also means an individual capability that contradicts local practice has nowhere to land.

What "local practice" means here is narrower than culture. It is which file a reviewer opens before approving, what has to be attached to a deliverable before it ships, and whose objection can hold something up. Your hire read all three inside a quarter. Had they kept working their own way, they would have been the only person on the team producing artifacts nobody read, on time nobody had budgeted.

Six months is simply when it becomes visible to you. The signal arrived much earlier, which is why the useful version of this problem is designing the first ninety days rather than diagnosing the second quarter.

What in the workflow taught them to stop?

Three places, and none of them is a conversation. What a reviewer actually opens, what has to exist before the deliverable, and whose objection can stop something shipping. If none of the three asks for framing, a traced source or a check, then the person doing those things is spending unbilled time on work nobody reads. Add the social cost of being visibly different and the quiet is fully explained.

That social cost is measurable. In Slack's Fall 2024 Workforce Index, a survey of 17,372 desk workers across 15 countries fielded in August 2024, 48% said they would be uncomfortable admitting to their manager that they had used AI for at least one common work task. The reasons ran to feeling like using AI is cheating (47%), fear of being seen as less competent (46%) and fear of being seen as lazy (46%). Company policy was the least-cited reason at 21%, and 45% of those surveyed said they had no explicit permission to use AI at all ². In the same survey, workers comfortable saying they had used AI were 67% more likely to have used it for work than workers who were not ².

So nobody told your hire to stop. They watched which submissions got waved through, which got questions, and which drew a remark about spending a long time on the setup. Six months of that is a training signal delivered daily by the people whose approval the work needs.

The three places, concretely:

• What a reviewer opens. If approval reads the deliverable and nothing else, the framing note, the source list and the check are invisible by construction. They were done or they were not, and the record cannot tell you.
• What has to exist before the deliverable. A plan, a criteria list, an outline. If none is expected, the fastest legitimate route to done is to ask for the finished thing and tidy the edges.
• Whose objection stops something. If the only people who can block a ship never look at how a claim was settled, then checking has no consequence attached to it, and unconsequential work is the first thing a busy person drops.

The middle one is the one teams skip, because an intermediate artifact looks like overhead right up until someone has to review a decision without it. It is also the first thing a capable AI user abandons, since it is the step nobody asked for.

Was the team right to hold the line?

Partly, and that part deserves taking seriously. Teams that resist AI-shaped work usually have a reason on file. In research by BetterUp Labs and the Stanford Social Media Lab, a September 2025 survey of 1,004 full-time US desk workers, 40% said they had received AI-generated work in the past month that looked polished but was unhelpful or off the mark, at an average of about one hour and 51 minutes spent on each instance ³.

The largest share of that traffic was lateral: 40% of it moved between peers, 19% from employees up to their managers and 16% from managers down to their teams ³. A team that has absorbed a few of those builds an informal rule against work that reads as generated, and the rule does real work. It is also indiscriminate. It cannot separate a draft that was generated and then checked from one that was generated and shipped, because the finished artifact looks the same either way.

That distinction is the whole problem, and it is the same one that shows up when every take-home submission comes back polished: fluency is not the variable. What separates the two is whether anything was traced, refused or tested, and none of that survives into a finished document unless something asks for it.

So the move is not to ask the team to relax its standard. It is to give the standard something to read.

Change what a review step asks for

Add one required field to the artifact your team already reviews, and give the reviewer a question they can already answer. Not a policy, not a training module, not a channel. The unit that carries a working practice is the thing someone has to open before approving, because that is the only place a new person's habits get corrected in either direction.

Where the field goes, by function:

• Software engineering. In the pull request template: which generated change you kept, and what you ran against it that the existing tests did not already cover.
• Financial analysis. On the memo cover page: the figure the recommendation turns on, and the filing page it was traced to.
• Marketing. In the brief header: which statistic in this deck was opened at source, and what its sample actually was.
• Legal operations. In the redline summary: which cited authority was read in full rather than retrieved, and where the playbook and the signed precedent disagree.
• Data and analytics. In the top cell of the notebook: the result the assistant explained fluently, and the check that would have contradicted it.
• Revenue cycle and claims. In the appeal packet: which denial reason was verified against the payer's own policy text, and which claim you are conceding.

A required field decays into a box people fill with nothing unless someone can send it back, and sending it back takes domain judgment rather than AI expertise. That is the useful news for a manager who does not use these tools themselves. "Which claim does this rest on, and where was it checked" is answerable by anyone who knows the field, which is the same move behind helping managers who don't use AI judge AI-assisted work.

Transfer runs through artifacts rather than through proximity. Across 5,179 customer support agents, access to a conversational assistant raised issues resolved per hour by 14% on average and by 34% among novice and low-skilled workers, with the authors offering suggestive evidence that the model was disseminating the best practices of the more able workers ⁴. The thing doing the spreading sat in everyone's workflow. A capable colleague two desks away is not a mechanism, and neither is a lunch session. See how Olive measures this

What to build before the next hire like this

Write down the decision this capability is supposed to change, the artifact where it will show, and the person who reviews that artifact. Three lines. If you cannot fill in the second and third before the req opens, the role is a hope rather than a plan, and the same six months will run again with a different person inside them.

Start with the person you already have, who is the cheapest source of the answer. Ask what they did in month one that they no longer do, and what made them stop. Do not frame it as a performance conversation, because that closes the only information channel you have. They watched your process more carefully than anyone who has been there three years, and they can usually name the review step that made the extra work pointless in one sentence.

Then change the ramp, not just the role. The same study that found institutionalized onboarding suppressing role innovation also found self-appraised performance associated with more individualized socialization ¹. Read practically: if the point of a hire is to change practice, putting them through the standard collective ramp works against the reason you hired them. Give the next one a first assignment where the new practice is the deliverable, with a named reviewer and a date.

Before the req opens, map what the team can already do with AI, and decide whether any of this belongs in the competency model at all, since an AI competency written before the deliverables changed rewards vocabulary rather than practice. If the plan is simply to hire again, hire for AI skills or train the team you have is the prior question. A second capability hire into an unchanged review path produces a second dormant one.

Rereading six months of their output will not settle this. A quiet stretch records what your process asked for, and this person's ceiling left no trace in it. The evidence worth having is what someone does with an assistant on real occupational work: what they framed before generating, which source they demanded, what they kept for themselves, what they refused, and what they tested against something outside the conversation.

See the benchmarks >