Releveling a Job After AI Absorbed Part of the Work

Does AI absorbing part of a job lower the level?

Usually not. What an assistant absorbs first is production, and production was the junior half of most knowledge roles. What is left leans toward review, judgment and consequence, which is more senior work performed in fewer hours. The level should follow what remains, and the hours are the wrong input to it.

The measurements point the same way from two directions. Across a staggered rollout to 5,179 customer support agents, access to a generative assistant raised issues resolved per hour by 14% on average, 34% for novice and low-skilled agents, and close to nothing for experienced ones ¹. Pooling three company-run randomized trials covering 4,867 developers, an AI coding assistant raised completed tasks by 26.08%, with a standard error of 10.3%, and less experienced developers both adopted it more and gained more ².

Read those as what they are. Both are narrow settings, one firm and one occupation in the first case, code completion rather than agents in the second, and the honest range around 26.08% runs from roughly 6% to 46%. What survives the caveats is the pattern: the assistant is closest to a substitute for the least experienced work in the role, and closest to useless on the part an expert was doing anyway.

That is the part most plans get backwards. If the tool does the most for the least experienced worker, then the work it replaces is the work you used to hire juniors to do, and the work it leaves is the work you used to hire seniors to check. Cutting the level cuts the wrong half. Before touching the level at all, it is worth being explicit about how many people the plan actually needs, because a releveling decision made inside a headcount argument is usually a headcount decision wearing a level's clothes.

Start from what is left, not from the old title

List the tasks that remain once the assistant drafts the first pass, then price the wrong call on each one. Two things set the level: how much of the remaining work ends in a judgment, and what it costs when that judgment is wrong. The old title carries neither number, which is why releveling from it goes wrong in a predictable direction.

The procedure is short enough to run in an afternoon:

1. Write the remaining task list. Not duties. Tasks small enough that somebody can say whether a person or an assistant does the first pass. 2. Mark each remaining task with the cost of being wrong. Reversible in an hour, reversible in a week, or not reversible: a payment made, a claim published, an offer sent, a diagnosis coded. 3. Count how many of the not-reversible ones have nobody above them. That count is the level, more reliably than years or headcount ever were. 4. Check who catches the miss. If the answer is the person you are about to hire, you are hiring a reviewer, and reviewers are not junior.

A worked contrast makes the split obvious. An underwriting role where the assistant drafts the summary and the person decides the risk has one not-reversible task with nobody above it, so it went up a level. A reporting role where the assistant drafts the deck and a manager still signs every number has none, so it went down, and the posting should say so plainly rather than keeping a scope that implies otherwise.

The same arithmetic answers the question one level down, which is whether to take the junior who is fast with AI over the senior who is not. If the remaining tasks are mostly not reversible, speed on the drafting half buys very little.

What does the AI salary premium actually measure?

Advertised salaries in postings that mention AI skills, compared with postings that do not. Lightcast puts that gap at 28%, roughly $18,000 more a year, and reports that in 2024 over half of postings requesting AI skills sat outside IT and computer science ³. It is a raw comparison between two groups of postings, and it is not a premium anybody owes a particular hire.

Three things sit between that figure and your range. AI-mentioning postings skew senior, urban and toward higher-paying industries, and the published analysis does not adjust for any of that. It is a posting-level number, so it measures what employers advertise rather than what anyone is paid. And Lightcast sells skills data, which is a reason to read the direction rather than the decimal.

The deeper problem is that the premium answers a different question than the one on your requisition. It describes what the market pays to attract people whose specialism is AI. Your question is what to pay for a role that used to be half production and is now mostly review, in an occupation where nobody's title says AI at all. Those come apart completely: the specialist premium is a scarcity price in a narrow market, while your role is an ordinary job with a changed shape.

Which means the premium is a poor input and a decent sanity check. If the posting is asking somebody to own judgment that used to sit a level above them, and the range is unchanged or lower, the market number is at least a signal that the pricing is running the wrong way. Use it to argue against a cut, not to justify a raise.

Decide: cut the scope or raise the range

Two honest exits, and the posting has to take one. If the remaining work is high-consequence review, raise the level and the range and write a smaller, sharper scope to go with them. If it is genuinely lower-stakes assembly, keep the range and cut the scope in the posting so the two agree. The third option, an unchanged scope at a reduced range, is the one that produces declined offers and short tenures.

What each exit looks like on the page:

• Raise the level. The scope shrinks and gets specific. Fewer responsibilities, each of them a call the person owns outright, with the production work described as context. Say who reviews them, since a senior hire wants to know whether there is anyone to argue with.
• Cut the scope. Say which decisions stay with the manager and which pieces the hire owns. A junior posting that admits its ceiling attracts people who want the ceiling, and it stops attracting people who will leave in seven months when they find it.

There is a wider pattern worth putting in front of an executive team before either exit. In administrative payroll records, employment of 22-25 year olds in the two most AI-exposed quintiles fell about 11% between November 2022 and June 2026 while the same age group in the three least-exposed quintiles grew about 10%, with the declines concentrated in occupations where AI substitutes for human tasks rather than complementing them ⁴. The authors are explicit that this is descriptive rather than causal, and it is occupation-level rather than a claim about any employer.

Still, it names the decision. A team using the assistant to replace the junior half is choosing the first pattern; one using it to extend what people already do is choosing the second, and only one of those leaves anywhere for a first hire to stand. Whether entry-level hiring still pays is the argument underneath the releveling question, and it is worth having before the requisition rather than after two failed searches. If the role is also remote, what the posting has to say about review becomes part of the same paragraph.

See the benchmarks >