Pipeline
Move the Band When the Tasks Moved, Not When the Posting Did
Adding an AI requirement to a posting does not by itself change what the job is worth. The published AI wage premiums, 62% from PwC and 28% from Lightcast, are gaps between two piles of job adverts, not measurements of anyone's pay. Neither controls for education, experience or location. Move the salary band when the task mix moved: more judgment held inside the role, more drafted output to check, more consequence riding on a decision nobody downstream repeats. Post a good-faith range either way.
The takeThe premium figures get read as a price when they are closer to a census of who writes AI language into postings. Where the comparison is left uncontrolled, technology employers write it most and technology already paid the most, so part of that gap predates anyone learning a tool. A band that follows the sentence you added is a band set by your own copywriting. A band that follows the work still makes sense to whoever reads the requisition two years from now, which is the only audience a pay audit brings into the room.
Where Olive fits
Open a role and see what the work shows
Olive is priced per attempt rather than per seat, and one attempt returns six evidenced findings about how a candidate works with an AI assistant, each anchored to a moment in the session. Ten attempts a month are free, which makes it cheap to find out whether the AI line on the posting describes work anybody in the pipeline actually does.
Assess your shortlistThe average is weighted toward technology, and your requisition may not be
AI language concentrates in a technology-heavy part of the market, and in a raw comparison like Lightcast's, which nets out no industry, that concentration pulls the average up before anyone is paid anything. How much it concentrates is something Olive counted in house rather than read off a study, on live postings, with three hundred postings on each arm still to be read by hand.
That count, taken on 2026-08-11, puts 79.7% of the 34,743 postings carrying an explicit AI-usage requirement in Technology.
A premium averaged over postings like that partly reports which occupations wrote the words.
A count taken the same day brackets the requirement's share of the 998,166 postings dated July 2026 between 1.4% and 8.1%, using two instruments that disagree, and an unfinished reading is published as a range rather than a point estimate.
Concentration does not explain the whole gap, and the counterweight sits inside one of the same sources. Lightcast reports that in 2024 more than half of the postings requesting AI skills were outside IT and computer science 2.
Marketing, finance and operations adverts are asking too.
So the question is whether your requisition is a technology posting or one of the rest. A marketing manager posting with a new line about drafting campaign copy in a model has little in common with a staff engineer taking over an agent pipeline, and the technology-heavy average prices neither of them.
Where the AI language sits on the posting changes what it does. Placed in the title, the word arrives carrying a pay expectation, which is part of why putting AI in the title changes who applies rather than what the role does.
Which changes to the work actually move a band?
Changes in what the role owns. Three belong in the requisition: judgment that used to sit with a senior reviewer and now sits with this hire, drafted output arriving in volume that somebody has to check before it ships, and a decision nobody downstream repeats.
Each is a level argument, and a band follows a level.
Write the before and after as tasks, never as tools. A line reading "experience with Claude and Copilot" prices nothing, because it names what is installed. A line reading "reviews and signs off on model-drafted reconciliations that a second analyst used to prepare" is a scope change, and comp can read it.
Check what the requirement is asking for before pricing it, because the common version asks for something that cannot exist yet. A demand for three years of AI experience prices scarcity the posting invented, which is the argument in how many years of AI experience you can ask for.
The mirror case runs the other way. When AI absorbed part of a job rather than adding to it, the move is releveling a job after AI absorbed part of the work.
The answer there runs against the instinct: the assistant takes the production half first, and production was usually the junior half.
Post a good-faith range before the band argument ends
Where a disclosure law applies, the range is due when the posting is, not when the offer is. New York's Labor Law 194-b, in the text last revised in September 2023, requires an advertised job to carry the compensation or a range of compensation 3. The statute defines that range as the minimum and maximum the employer in good faith believes to be accurate at the time of the posting 3.
Other states and cities run their own versions with different scope and penalties, so confirm the one you post under with employment counsel. None of this is legal advice.
A published range is a commitment made before the first application arrives, which catches teams expecting to settle the band at offer.
One working paper measured what Colorado's January 2021 disclosure law did to postings. The share of postings carrying salary information rose about 30 percentage points, from roughly 35% before the law to nearly 70% in the months after 4.
Among employers that posted salaries both before and after, average posted salaries rose about 3.6% 4. The authors compute that inside the same employer, occupation and county, so it is not different jobs becoming visible.
Publishing a number moves the number, with no AI line involved.
Compliance never came close to complete. The same paper reports substantial non-compliance, with postings for high-paying occupations less likely to comply, and it is a working paper measuring one state under one law 4.
What does a band raised only for new hires cost at the next review?
The difference, paid twice. A band moved for incoming candidates and left alone for the people already doing the work builds compression that surfaces at the next review cycle, when someone hired in March sits above a colleague three years into the job. The second payment is the correction.
A pay audit reads the internal comparison, not the market one.
Where one finalist genuinely warrants more, the instrument is an exception written on that offer rather than a move to the whole band. An exception that survives a later audit has a fixed set of contents, set out in paying above the band for demonstrated AI capability.
Hold the band and the market may move without asking.
So far it has not shown up in young hires' starting base pay. Stanford's Digital Economy Lab read ADP payroll records and found no obvious relationship between an occupation's AI exposure and real starting pay for newly hired young workers. The lab concluded that the adjustment ran through whether young people were hired at all 5.
The limit is named in that paper. Base pay only, excluding bonuses, commissions and equity, which are largest in precisely the highest-paid and most exposed occupations, so an adjustment could be running where nobody is measuring 5.
The posting needs two edits, not one. Delete the AI line if no task behind it moved, since a line no task depends on is the kind every candidate now clears in the screen. Keep it if the work moved, and write the moved tasks into the requisition where comp will read them.
The posting goes live carrying a number either way. What decides whether the number holds is the next review cycle, when the people already doing the job read the range you published. They will ask which of them it describes.
Common questions
Does adding an AI requirement to a posting oblige you to raise the salary band?
No. The disclosure duty in New York's statute is a range held in good faith at the time of posting, and no published premium figure measures what a single role is worth. What obliges a change is a change in the work: judgment moving into the role, drafted output arriving in volume that has to be checked, a decision nobody downstream repeats. Where none of those happened, the band holds, and under a disclosure law the posting still needs a good-faith range in it. Where they did happen, the case to comp is a level argument, and the band follows the level rather than the sentence.
Can you quote the 62% AI wage premium to comp?
Quote the 62% as direction, never as a price. PwC's 2026 Barometer computes that figure by comparing advertised salaries on postings requiring at least one AI skill against postings that do not. The same page says the estimates do not control for education, experience or location. Its own spread runs from 118% in consumer markets to 16% in government and the public sector. Lightcast, applying its own rules to its own database, puts a similar comparison at 28%. A comp partner who has read either methodology will say all of that first.
Do you have to post a range while the band is still being argued?
In a jurisdiction with a disclosure law, yes: the posting carries a range while the band is still being argued, and it has to be a range held in good faith. New York's Labor Law 194-b defines the required range as the minimum and maximum the employer in good faith believes to be accurate at the time of the posting, so a placeholder nobody intends to pay is not compliance. The practical consequence is a deadline: the band question closes before the requisition goes live rather than at offer. Scope and penalties differ by state and city, so confirm yours with employment counsel.
What if candidates start asking for more because the posting mentions AI?
Expect candidates to ask for more when the posting mentions AI, and answer with the role rather than the market figure. A candidate quoting a 62% premium is quoting a difference between two piles of job adverts, which describes neither the role nor them. If the work genuinely changed, the level changed, and the conversation is about scope. If it did not, hold the range you published and compete on what the job owns: the decisions the person makes without asking anyone, and how fast the work reaches somebody who cares about it.
How do you tell a real task change from a reworded job description?
Put the old requisition beside the new one and read only the verbs. A rewrite adds tool names, adjectives, and a line about comfort with emerging technology. A real change moves a task: something the role now signs off on, something it now produces at ten times the old volume, something a second reviewer used to catch and no longer does. If the verbs are identical and only the vocabulary moved, the job is the same job and the band is the same band.
References
- 1. Two futures for jobs in an AI era — 2026 Global AI Jobs Barometer (global findings) ✓ pwc.com Supports the 62% advertised-salary premium for postings requiring at least one AI skill, PwC's statement on the same page that its estimates do not control for education, experience or location, and the sector spread from 118% in consumer markets to 16% in government and the public sector. Cited as an uncontrolled comparison of adverts, never as a price for a role.
- 2. Beyond the Buzz: Developing the AI Skills Employers Actually Need ✓ lightcast.io Supports the 28% higher advertised salaries on postings mentioning AI skills, roughly $18,000 a year, and the claim that in 2024 more than half of postings requesting AI skills were outside IT and computer science. Cited as a raw posting-level comparison, not a controlled wage estimate.
- 3. New York Consolidated Laws, Labor Law Section 194-B, Mandatory disclosure of compensation or range of compensation ✓ nysenate.gov Supports the duty to state the compensation or a range of compensation in an advertised job, and the section's definition of range of compensation as the minimum and maximum annual salary or hourly range the employer in good faith believes to be accurate at the time of the posting. The page shows a most recent revision of 2023-09-22. New York only; not legal advice.
- 4. The Impact of Pay Transparency in Job Postings on the Labor Market ✓ darnold199.github.io Supports Colorado's January 2021 law raising the share of postings with salary information by about 30 percentage points, from roughly 35% before to nearly 70% in the following months; the 3.6% within-job rise in average posted salaries among employers posting both before and after, estimated with employer-by-occupation-by-county fixed effects; and the authors' finding of substantial remaining non-compliance, with postings for high-paying occupations less likely to comply.
- 5. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence ✓ digitaleconomy.stanford.edu Supports the finding that real starting pay for newly hired young workers shows no obvious relationship with an occupation's AI exposure, so the measured adjustment ran through employment rather than base compensation, and the paper's own stated limit that the pay measure is base salary excluding bonuses, overtime, commissions, equity and tips.
5 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.