Put the New Senior in the Review Seat, Not in a Course

What Breaks First, the Ramp or the Review?

The review breaks first, and it breaks quietly. A senior who cannot follow how a piece of work was produced still has to sign off on it, and the cheapest way through that sign-off is to read the output, find it plausible, and approve.

Nothing looks wrong on the day. The team ships faster for a month, and the most experienced check on the work has stopped operating.

Somebody is already drafting the other response: a vendor course in week two, a prompt workshop in week five, a certificate by week ten.

The plan is not wrong so much as late. Every decision the senior was hired to make lands before week ten: which of the team's three approaches goes to the client, whether the reconciliation is trustworthy enough to close on. Then there is what the model got confidently wrong about a market the senior has worked in for twelve years.

Judgment does not wait for the syllabus.

The mirror case runs the other way. A manager who cannot tell what an AI-heavy new hire knows can still structure that hire's first ninety days around the steps instead of the output, because the manager holds the review.

Reverse the seats and the review is the thing at risk.

Why Does the Course Aim at the Cheap Half of a Senior's Gap?

A course teaches the part of the gap that costs the least. Tool skill transfers quickly, and the measured gains from it land mostly on people with less experience. What does not transfer quickly is the thing the senior already has and the team does not: enough scar tissue in the domain to know when a fluent answer is wrong.

Training aims at the cheap half.

In 2020 and 2021, one software firm rolled out a conversational assistant to 5,179 of its customer support agents in stages. Access raised issues resolved per hour by 14% on average ¹.

The average hid the shape. Novice and low-skilled agents improved 34%, and experienced, highly skilled ones barely moved ¹.

The assistant in that study was trained on the firm's own transcripts, and everyone using it did the same job in the same company. Read it as a direction rather than a rate you can expect.

Support work is a long way from what a senior does, so the harder question is whether the same split survives in expert work.

Three company-run randomized trials covered 4,867 developers at Microsoft, Accenture and an anonymous Fortune 100 firm. Developers using an AI coding assistant completed 26.08% more tasks, and the less experienced among them both adopted the tool more and gained more from it ².

That 26.08% carries a standard error of 10.3%, so the honest range runs from roughly 6% to 46%.

A quarter planned around a promised percentage has nothing to stand on inside a range that wide. What both measurements leave standing is smaller and more useful: the gain lands where the experience is thinnest, and that is not where the new senior sits.

Nobody should be running this quarter against a ramp clock anyway.

The standard eight-month time-to-productivity figure comes from a March 2012 online survey of 500 US HR professionals, sponsored by a moving company. In it, 58% of respondents said their company does not measure new-hire productivity at all ³.

The more useful frame is that output ramps in weeks now and judgment still takes months.

Hand Them a Review Before You Hand Them a Tool

Give the senior a live review assignment in week one, before anybody books the training. Pick work where the question the review has to answer is a domain question: is this the right comparable, does this reconciliation actually tie, would a regulator read the clause the way the draft assumes.

Tool fluency is not what answers any of those. Twelve years in the field is.

Then change what the team submits. Alongside the artifact, require the reasoning: what was asked for, which claim the deliverable rests on, what was checked and against what.

A senior reading that reasoning can review the work without reading the tool.

The failure this catches is specific. In a field experiment run with Boston Consulting Group consultants, one task was deliberately chosen to sit outside the capability of the 2023 model they were given.

Consultants using GPT-4 were 19 percentage points less likely to reach the right answer, averaged across the two AI conditions. The control group got it 84.5% of the time, against 60% and 70% for the two groups working with the tool ⁴.

The consultants using the tool could not tell which side of the line the task sat on, and telling is what a domain reviewer is for.

One task and one sample, so read it as the shape of the failure and not its rate. The shape is a wrong answer on work that looks like the work the tool handles well.

Recognizing that shape takes hours on the tool, which is why a manager who has never used AI cannot review AI work even with the domain knowledge intact.

Pair Them Both Ways, and Let Them Name the Gap

Pair the senior with someone junior on the team for two weeks, and run the exchange both ways on purpose. The junior shows the workflow: where the context comes from, what gets pasted, which step is automated and which is checked. The senior brings a case from their own file and works it out loud.

Neither person is the student for the whole hour.

Direction is the whole design. A one-way arrangement puts a junior in an evaluative position over their own manager, and the honest feedback stops on day three. A two-way exchange is two people comparing notes on the same problem, which is a thing seniors do without embarrassment.

Name the gap in week one, in the senior's own words, to the team.

A senior who has said out loud that they are behind on the tooling can ask a direct question in a review meeting and get a real answer.

Then judge the two weeks on the work itself.

No self-report is worth much here, the senior's included.

One small randomized trial put that kind of self-report next to a clock. It gave early-2025 AI tools to 16 experienced open-source developers, on repositories they had worked on for about five years, across 246 real tasks.

They expected the tools to make them 24% faster.

Afterwards they said the tools had made them 20% faster. The measurement came back 19% slower ⁵.

The trial is 16 people in one setting, so 19% is not a number to expect, and nothing in it reaches a junior working in unfamiliar code.

What it does suggest is which question a check-in should ask. The developers in that trial were wrong about their own speed after doing the work, and wrong in the direction that flattered the tool.

So ask the team what changed in the work, not how the senior is doing.

Write a Ninety-Day Condition That Never Mentions a Tool

Write the ninety-day condition about what the senior can review, not about what they can prompt. A usable version reads: by day ninety, the senior can take any deliverable this team produces and say which claim it rests on. They can name what would have to be false for the recommendation to flip, and send it back with a stated reason.

None of that mentions a tool.

The plan is the manager's to write. Settle what actually goes in a 30-60-90 day plan, and who should write it before the first Monday.

A plan the hire writes alone becomes a list of things they already know how to do.

Borrow the vocabulary rather than inventing it. The AI Fluency Framework comes from Rick Dakan, a professor at Ringling College of Art and Design, and Joseph Feller, a professor at University College Cork. It names four competencies: Delegation, Description, Discernment and Diligence ⁶.

Under Diligence sits Deployment Diligence, which the framework defines as taking responsibility for verifying and vouching for AI-assisted output, including fact-checking, testing for accuracy and validating claims ⁶.

Use it as shared language.

The framework carries no levels, no scoring anchors and no validation study, so a ninety-day condition written against it cannot be a number. A senior and a junior arguing about whether an output was discerned are at least arguing about the same thing.

What If the Senior Just Approves Faster?

A senior who approves faster has failed the seat, and the quarter has bought nothing. A review seat only works if the person in it is willing to refuse. A senior who is anxious about looking behind their own team may approve faster rather than slower, precisely because they were handed the seat in week one.

One more thing, and it belongs to the next req rather than this one. Did any round in the loop ask the senior how they review work somebody else produced with a model?

If none did, the surprise at onboarding was a selection result wearing a training costume.

The exercise that shows a manager can review AI work is the missing round, and it is cheaper to add than the quarter you are now running.

So rewrite the plan somebody drafted in week one. The course can stay in week five if the budget is already spent.

The review seat, the two-way pairing and the condition about reviewing start on Monday.

By week five you will know whether the course is still the thing you need.

See the benchmarks >