TL;DR. An AI tool raised your output, and the instinct is to ask for a matching raise. The evidence says that argument fails and shows which one works. Controlled studies measure real gains on tasks: 15% more customer issues resolved per hour, 26.08% more completed coding tasks, a writing task finished 40% faster. But a large study linked to administrative records, covering about 25,000 workers in Denmark, finds no effect of chatbot use on earnings and rules out effects larger than 2%; 97.7% of adopters say the tools have not changed what they earn. Rent-sharing research helps explain why: a 10% rise in a firm’s value added per worker has typically lifted wages 0.5% to 1.5%, and in a large US study workers captured about 30 cents of each dollar of new surplus, with roughly twice that share going to incumbents who were costly to replace. Ordinary annual wage growth runs at 3.2% to 3.4% in current US data. The CEO move: stop selling speed and present a business case with a number, built on the part of the gain that would leave with you. The student move: keep building and documenting the judgment, verification and workflow design that make the gain yours rather than the licence’s.
This piece takes the gap between individual AI gains and company results down to one pay conversation. Freelancers face the same question as a pricing problem, covered in the repricing playbook for raising rates. What you bring to either conversation is the subject of career capital in the AI era.
How large are the measured AI output gains, really?
“AI multiplied my output” is a claim about a number, so start with what has been measured under controlled conditions.
Table 1. Measured productivity effects of AI assistance (figures as printed in each source)
| Study | Setting and sample | Measured effect | Who gained most, or where it failed |
|---|---|---|---|
| Brynjolfsson, Li and Raymond (Quarterly Journal of Economics, 2025) | Customer support at one software firm, 5,172 agents | Issues resolved per hour up 15% on average | Less skilled and less experienced agents: an approximate increase of 30%; minimal impact on the more experienced |
| Cui and co-authors (working paper, February 2025) | Three field experiments, 4,867 developers at Microsoft, Accenture and a Fortune 100 company | 26.08% more completed tasks (standard error 10.3%) | Less experienced developers |
| Peng and co-authors (2023) | 95 programmers recruited on a freelancing platform, one JavaScript task | Task completed 55.8% faster (95% confidence interval 21-89%): 71.17 against 160.89 minutes | Developers with less programming experience |
| Noy and Zhang (Science, 2023) | 453 college-educated professionals, incentivized writing tasks | Time taken down 40%, output quality up 18% | Inequality between workers decreased |
| Dell’Acqua and co-authors (Harvard Business School working paper, 2023) | 758 knowledge workers, experiment run with Boston Consulting Group, 18 tasks | 12.2% more tasks completed, 25.1% more quickly | On a task chosen to lie outside the tool’s capabilities, AI users were 19% less likely to produce correct solutions |
Sources: the named papers. The 2023 working paper version of the first study printed 14% from 5,179 agents and 34% for novice and low-skilled workers; Table 1 uses the journal abstract and the November 2024 revision.
Three things matter for a pay conversation. Every gain is measured on a task or one output metric, not a whole job: even 55.8% faster means 2.26 times the speed on one programming exercise. The gains concentrate among less experienced workers, so the tool narrows the distance between you and the next person, the opposite of scarcity. And on the wrong task, the tool made people less accurate.
Does any of that gain show up in pay?
Most raise arguments assume the answer is yes. The best evidence available so far says: not by default.
Anders Humlum and Emilie Vestergaard linked large surveys of AI chatbot use in 11 exposed occupations in Denmark, among them software developers, marketing professionals and accountants, to administrative records of earnings and hours. The latest round covers about 25,000 workers across 7,000 workplaces. Their March 2026 revision reports “precise null effects on earnings and recorded hours”, ruling out effects larger than 2% two years after the launch of ChatGPT. The null holds for daily users, for workers who report saving more than an hour a day, and in firms that encourage the tools. It holds from centrally bargained teachers to individually negotiating software developers and marketers, in occupations where most workers negotiate with their employers every year.
Asked directly, 97.7% of adopters said AI chatbots had not affected their labor earnings. Most users (85%) reported reallocating the saved time to other job tasks. Earlier 2025 versions of the paper put a number on the leak: time savings averaged 2.8% of work hours, and the authors estimated from workers’ own reports that only 3-7% of those savings translated into higher earnings.
The first study in Table 1 adds a detail. Its authors state that their data “do not allow us to observe changes in wages”, but describe the pay system: bonuses tied to targets such as chats per hour, estimated at 20 to 40 percent of take-home pay. Where pay is contractually tied to an output metric, part of a gain flows through automatically. Most salaried knowledge work has no such link, so nothing moves unless someone negotiates.
How much of a gain do employers usually share?
Economists study this as rent sharing: when a firm becomes more productive or profitable, how much reaches its workers?
A review by David Card, Ana Rute Cardoso, Joerg Heining and Patrick Kline concludes that studies controlling for who works where typically find that “a 10% increase in value-added per worker leads to somewhere between a 0.5% and 1.5% increase in wages”, an elasticity of 0.05-0.15. Patrick Kline, Neviana Petkova, Heidi Williams and Owen Zidar followed US firms whose valuable patent applications were allowed and found that workers captured roughly 30 cents of every dollar of the resulting surplus. The share was about twice as high for workers already at the firm: stayers captured roughly 61 cents. The authors could detect no response of entry wages.
Table 2. From output gain to pay: what each layer of evidence measures (CEOtudent analysis of the named studies)
| Layer | Source | What moved | What reached pay |
|---|---|---|---|
| Task output with AI | Table 1 studies | 12.2% to 26.08% more output; tasks done 25.1% to 55.8% faster, as each study defines it | Not measured |
| Earnings of AI users | Humlum and Vestergaard, March 2026 | Time savings and new tasks for chatbot adopters in 11 occupations, Denmark | No effect; effects larger than 2% ruled out; 97.7% of adopters report no earnings impact |
| Workers’ self-reports | Same study, 2025 versions | Time savings averaging 2.8% of work hours | 3-7% of the time savings (authors’ estimate), which implies 0.08% to 0.20% of earnings |
| Firm productivity | Card, Cardoso, Heining and Kline review | A 10% rise in value added per worker | 0.5% to 1.5% in wages |
| Firm surplus from a valuable patent | Kline, Petkova, Williams and Zidar | One dollar of new surplus per worker | About 30 cents on average; 61 cents for stayers; no detectable response of entry wages |
| Illustrative arithmetic | Rows 1 and 4 combined | A 15% to 26.08% gain, if it were firm-wide value added | 0.75% to 3.91%, not a forecast |
CEOtudent analysis; derived cells are recomputed from the printed figures. The rows are not comparable enough to yield one pass-through rate: they mix tasks with whole jobs, experiments with administrative records, Denmark with the United States, and individual with firm-level gains. What they share is order of magnitude: output gains in tens of percent, pay effects in low single digits or less.
The patent study also explains who gets the larger share. Its authors interpret the earnings response as “the capture of economic rents by senior workers, who are most costly for innovative firms to replace”, and note that firms share with incumbents to raise the odds of keeping them. On this evidence, pay follows replacement cost more than output. A gain that any licence holder can reproduce raises nobody’s replacement cost.
Whose gain is it: yours or the tool’s?
That makes attribution the central question. One experiment shows how the case can go against you: in the working paper version of the writing study, 68% of treated participants reported submitting the chatbot’s initial output without editing it, and the authors found no evidence that human editing improved that output. In that setting the gain belonged to the tool.
Table 3. The “Whose gain is it?” attribution test (CEOtudent editorial framework)
| Test | Question to answer honestly | It is the tool’s gain if | It is your gain if | Evidence to bring |
|---|---|---|---|---|
| 1. Licence | Would a competent colleague with the same licence and two weeks of practice match the result? | Yes | Your result sits clearly above the team’s average with the same tool | Your numbers against the team baseline |
| 2. Edit | Does the work ship largely as generated? | Mostly unedited | Your structure, cuts and decisions change the outcome | Before and after examples |
| 3. Frontier | Do you know where the tool fails, and catch it? | Errors reach reviewers or clients | You stop them first | A log of corrected outputs and what each error would have cost |
| 4. Workflow | Who designed the prompts, checks and templates? | You use someone else’s | Colleagues use yours | Reuse count, internal documentation |
| 5. Scope | What did the saved time become? | More of the same tasks | Named responsibilities that did not exist a year ago | A dated list of what was added |
| 6. Replacement | What would it cost to replace you and rebuild this? | A licence and a short ramp-up | Months of lost capacity and context | A sober estimate, used privately to set your walk-away point |
CEOtudent editorial framework, informed by the studies above; a preparation tool, not a finding of those studies.
Count the tests where the honest answer is “yours”. With zero to two, do not lead with AI: ask for the benchmark increase below and spend two quarters changing the answers. With three or four, you have a raise case in the middle of the worksheet range. With five or six, you are arguing that the job has changed level. The same inventory from the employer’s side is in how to audit your job for AI replaceability.
What are the official pay benchmarks right now?
A raise request is judged against what pay is doing anyway. Three official US series answer that; many national statistics offices and central banks publish equivalents.
Table 4. Official US pay and productivity benchmarks, latest editions as of October 2026
| Benchmark | Latest figure | Source and edition |
|---|---|---|
| Wages and salaries, all civilian workers, 12 months | +3.2% | BLS Employment Cost Index, June 2026, released July 31, 2026 |
| Wages and salaries, private industry, 12 months | +3.1%, and -0.4% after inflation | Same release |
| Median wage growth, people who stayed in their job | 3.4% | Federal Reserve Bank of Atlanta Wage Growth Tracker, September 2026, three-month average |
| Median wage growth, people who changed jobs | 4.7% | Same |
| Median wage growth, overall | 3.9% | Same |
| Labor productivity, nonfarm business, year over year | +2.2% | BLS Productivity and Costs, second quarter 2026, revised, released September 3, 2026 |
| Real hourly compensation, last four quarters | -0.1% | Same release |
| Labor share of output | 52.8%, the lowest in a series that begins in 1947 | Same release |
Sources: US Bureau of Labor Statistics and Federal Reserve Bank of Atlanta, figures as published. The next Employment Cost Index release is scheduled for October 30, 2026.
Read the table as a floor. An offer of 3% to 3.5% is not a reward for anything; it is the market moving, and in private industry a 3.1% nominal rise was a 0.4% cut after inflation. The last three rows show the default: productivity up, real pay flat, labor’s share at a record low. Your contribution starts above the floor.
For the level of pay rather than its change, the public US source is the BLS Occupational Employment and Wage Statistics program. Its May 2025 estimates, released May 15, 2026, cover about 830 occupations for the nation, the states and about 530 areas, with wages at the 10th, 25th, 50th, 75th and 90th percentiles, excluding nonproduction bonuses and benefits. The bands are wide: in the illustrative table BLS uses to explain percentiles, the 75th percentile sits 20% above the median. Moving up a band is a larger conversation than an annual raise.
In the European Union, Directive (EU) 2023/970 gives workers the right to request and receive in writing their individual pay level and the average pay levels, broken down by sex, for categories of workers doing the same work or work of equal value, with a reply due within two months. It bars employers from asking applicants about pay history. Member States had to transpose it by 7 June 2026; how it applies to you depends on national law. Commercial salary websites are self-reported: use them as a cross-check, never as the main exhibit.
How do you turn an output gain into a number?
The worksheet converts an output gain into an ask in seven steps. Every figure in the example column is illustrative, not data. The sharing rates in step 4 come from the rent-sharing studies above; they are anchors, not an entitlement.
Table 5. Raise-case worksheet with illustrative numbers (CEOtudent editorial framework)
| Step | What you do | ILLUSTRATIVE example |
|---|---|---|
| 1. Measure the gross gain | Same unit, same quality bar, six to twelve months before and after | 20 to 28 deliverables a month: +40% |
| 2. Subtract the licence baseline | Use the team’s average gain with the same tool | Team +15%, so your margin is 1.40 / 1.15 – 1 = 21.7% |
| 3. Price the margin | Value the extra capacity at your own pay rate, a conservative proxy | 21.7% of a 60,000 salary = 13,043 a year |
| 4. Apply a sharing range | Low 5% and middle 15% (the review’s range), high 30% (the patent study’s average) | 652 / 1,957 / 3,913, or 1.1% / 3.3% / 6.5% of pay |
| 5. Add the market floor | The 12-month wage benchmark, here +3.2% | 4.3% / 6.5% / 9.7% |
| 6. Set three numbers before the meeting | Opening ask, target, and the point below which you switch to the fallback | Open at 9.7% (65,833), target 6.5% (63,877), fallback below 4.3% (62,572) |
| 7. Check the level | Compare the result with the official percentile band for the occupation and area | If the target still sits below the median, raise the level question too |
CEOtudent editorial framework. Salary, output counts and team baseline are invented for illustration; the 3.2% floor is the BLS figure in Table 4. General information, not individual advice.
The lesson: a 40% output gain supports an ask of roughly 6% to 10% in this example, not 40%. Asking for the full 40% would mean 24,000, which is 1.84 times the entire money value of your margin over colleagues. A manager does not need the research to sense that the request claims more than the gain.
What should you actually say?
The scripts assume a scheduled meeting and a one-page summary: the metric, the team baseline, your margin, the three things that produce it, and the number.
1. The opening ask
“I would like to talk about my pay and level. Over the last twelve months my output on [metric] went from [X] to [Y] at the same quality bar. The team average with the same tools moved by [Z], so the difference is not the licence. It comes from three things I built: [verification step], [workflow the team now uses], [new responsibility]. Based on that and on current wage benchmarks, I am asking for [opening number]. The numbers are on this page. What would you need to see to support this, and how are pay decisions for this role made?”
2. The “it was the tool, not you” objection
“Part of it is the tool, and I have taken that part out. Everyone on the team has the same licence, and the team’s gain is [Z]. Mine is [Y]. The gap is what I am asking you to price: the errors I catch before they ship, the workflow the team now runs on, and the work I took on with the time I saved. If I have that gap wrong, which number would you use instead?”
3. The “no budget” objection
“I understand the budget is fixed for this cycle. Can we separate the decision from the timing? If you agree the case is sound, I would like that confirmed in writing, with a review date and the criteria that will settle it. In the meantime, which of these can be decided without a salary budget: scope, title, a one-time bonus, or a learning budget?”
4. The non-salary fallback
“If base pay cannot move now, here is what would make this fair, in order. First, a title and scope that match the work I already do: [specific]. Second, a one-time bonus tied to [result]. Third, [amount or days] of training in [skill], which raises what I can deliver here. And a salary review dated no later than [month].”
5. The written follow-up, sent the same day
“Thank you for today. To confirm: I asked for [number], based on [metric, X to Y, team baseline Z] and the responsibilities I have added: [list]. You said [summary]. We agreed that [next step] will happen by [date] and that the decision will rest on [criteria]. The one-page summary is attached. Please correct anything I have recorded inaccurately.”
Three choices sit behind the wording. The opening asks how pay decisions are made, because ambiguity suppresses asking: in a field experiment with 2,422 job-seekers, 9.4% of applicants negotiated when the wage was not explicitly advertised as negotiable and 23.1% when it was. The tool objection is conceded before it is contested, which only works if step 2 of the worksheet is real. The fallback puts title and scope first because a level travels with you; the path is described in the specialist to orchestrator transition framework. Rehearse all five aloud using practice sparring with AI.
Who asks, and what happens when they do?
Table 6. What studies of asking for pay report (figures as printed)
| Source | Sample | Finding |
|---|---|---|
| Pew Research Center, 2023 | 5,188 employed US adults, asked about their last hiring | 32% of men and 28% of women asked for more than the first offer. Of those who asked, 28% got what they asked for, 38% got more than offered but less than asked, 35% only the first offer |
| Marks and Harold, 2011 | 149 newly hired employees | Those who chose to negotiate raised their starting salaries by an average of $5000 |
| Artz, Goodall and Oswald, 2018 | About 4,600 workers in 840 Australian workplaces | 75% of men and 66% of women had asked for a pay rise; the mean rate of having obtained better pay by negotiating, without changing roles, was 0.16 |
| Exley, Niederle and Vesterlund, 2020 | Laboratory experiment with a penalty for failed negotiations | When negotiating was compulsory, 51% of negotiations ended in gains for workers and 29% in losses; forcing women to negotiate raised their frequency of losses from 9% to 33% |
Sources: the named publications. The first two rows concern starting pay at hiring, not raises in an existing job.
Asking tends to pay: in the Pew survey 66% of those who asked got more than the first offer. But not every ask is free. Most Australian workers had asked at some point, and far fewer reported success inside the same role. In the laboratory, women who could choose when to negotiate almost always gained, while those forced into every negotiation took more losses and no extra gains.
When should you not ask, and when is changing employer the better move?
Hold the ask when the attribution test returns two or fewer, when you have no before-and-after record, or when the employer is cutting costs (then ask for scope and a dated review). Where pay follows a fixed scale, the target is the grade, not the rate.
Changing employer is the usual alternative, and it looks better than it is. In September 2026 the Atlanta Fed tracker put median wage growth at 4.7% for people who changed jobs and 3.4% for those who stayed, a gap of 1.3 points, or 780 a year on the worksheet’s illustrative 60,000. The gap is cyclical: it reached 2.7 points in November 2022 (8.2% against 5.5%) and was negative in August 2009 (1.7% against 2.8%). The tracker also counts as a switcher anyone whose occupation, industry, employer or job duties changed, so part of that premium is earned by changing role without leaving.
What the evidence supports is moving for a different role. In the Danish study, occupational switching was the only administrative outcome that moved with chatbot adoption, and adopters who switched occupations saw earnings grow 12 percentage points faster between November 2022 and December 2024 than other Danish workers employed in both months. That is an association, not a proven effect. It suggests a rule: stay and negotiate when your value is tied to this employer’s context; move when your level is wrong, a dated review has passed without change, and the market for the larger role is hiring. For reading that market, see which roles recovered in the 2023-2026 layoff data.
What are the CEO move and the student move?
The CEO move is to own the business case and the number: one page, three numbers fixed before the meeting, a dated decision. A refusal is information: either the case needs two more quarters of evidence or the employer does not price this kind of contribution.
The student move is to keep the gain yours. A tool-provided gain decays into the baseline once colleagues catch up; the judgment around it does not. Keep a log of errors caught and decisions made, do some demanding work without the tool so the checking skill stays sharp, learn the next layer (evaluation, workflow design, domain depth), and teach your workflow, because reuse by others is the strongest evidence in Table 3.
What does this evidence not mean?
It does not mean AI gains are worthless to the person who produces them; they buy scope and mobility before they buy pay. The Danish results cover one country and the first two years after ChatGPT’s launch. Table 4 describes the United States. Rent-sharing estimates are averages across firms, not rules for individuals, and the worksheet’s sharing rates are anchors. Most evidence on asking concerns starting pay. Nothing here is legal or financial advice; pay rules differ by country and contract.
Frequently asked questions
Should my raise match the percentage my output increased?
No. In the worksheet example a 40% output gain supports an ask of roughly 6% to 10%: the market floor of 3.2% to 3.4% plus a share of the money value of your margin. Rent-sharing research finds that a 10% firm-level productivity gain has typically moved wages 0.5% to 1.5%.
Is the gain mine or the tool’s?
Subtract what colleagues with the same licence achieve. What remains, traceable to your checking, workflow design and added responsibilities, is yours. Table 3 gives six tests.
What if my manager says anyone could do this with the tool?
Concede the tool’s share, show the team baseline, and ask the manager to price the gap. If there is no gap, the manager is right, and the work is to create one.
Is it better to ask or to change jobs?
In September 2026 US job switchers were getting 1.3 points more median wage growth than stayers, and that premium has been negative in weak labor markets. Moving pays most when it comes with a larger role.
Does asking work?
Often. In a 2023 Pew survey, 66% of US workers who asked for more than the first offer at hiring received more. Timing matters: in a laboratory experiment, being forced into every negotiation increased losses.
Sources
- Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889-942 (November 2024 revision and NBER Working Paper 31161 of 2023 consulted).
- Cui, K. Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S. and Salz, T. (2025). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Working paper, February 2025.
- Peng, S., Kalliamvakou, E., Cihon, P. and Demirer, M. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv 2302.06590.
- Noy, S. and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192 (abstract), and the MIT working paper of March 2, 2023.
- Dell’Acqua, F. and co-authors (2023). Navigating the Jagged Technological Frontier. Harvard Business School Working Paper 24-013 (abstract).
- Humlum, A. and Vestergaard, E. (2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777, revised March 2026 (authors’ version of March 13, 2026). Earlier versions, titled Large Language Models, Small Labor Market Effects: Becker Friedman Institute Working Paper 2025-56, April and September 2025.
- Card, D., Cardoso, A. R., Heining, J. and Kline, P. (2018). Firms and Labor Market Inequality: Evidence and Some Theory. Journal of Labor Economics, 36(S1), S13-S70 (NBER Working Paper 22850 consulted).
- Kline, P., Petkova, N., Williams, H. and Zidar, O. (2019). Who Profits from Patents? Rent-Sharing at Innovative Firms. The Quarterly Journal of Economics, 134(3), 1343-1404 (NBER Working Paper 25245 consulted).
- US Bureau of Labor Statistics. Employment Cost Index, June 2026. News release USDL-26-1270, July 31, 2026.
- Federal Reserve Bank of Atlanta. Wage Growth Tracker, data through September 2026, updated October 8, 2026.
- US Bureau of Labor Statistics. Productivity and Costs, Second Quarter 2026, Revised. News release USDL 26-1434, September 3, 2026.
- US Bureau of Labor Statistics. Occupational Employment and Wages, May 2025. News release USDL-26-0725, May 15, 2026, with the program’s Frequently Asked Questions and Percentile Wages pages.
- Directive (EU) 2023/970 of the European Parliament and of the Council of 10 May 2023 on pay transparency and enforcement mechanisms. Official Journal of the European Union, L 132, 17 May 2023.
- Parker, K. (2023). When negotiating starting salaries, most U.S. women and men don’t ask for higher pay. Pew Research Center, April 5, 2023.
- Marks, M. and Harold, C. (2011). Who asks and who receives in salary negotiation. Journal of Organizational Behavior, 32(3), 371-394 (abstract).
- Leibbrandt, A. and List, J. A. (2015). Do Women Avoid Salary Negotiations? Evidence from a Large-Scale Natural Field Experiment. Management Science, 61(9), 2016-2024 (NBER Working Paper 18511 consulted).
- Artz, B., Goodall, A. H. and Oswald, A. J. (2018). Do Women Ask? Industrial Relations, 57(4), 611-636 (IZA Discussion Paper 10183 consulted).
- Exley, C. L., Niederle, M. and Vesterlund, L. (2020). Knowing When to Ask: The Cost of Leaning In. Journal of Political Economy, 128(3), 816-854 (NBER Working Paper 22961 consulted).
Tables 1, 4 and 6 report figures as printed in the named sources. Table 2 is a CEOtudent analysis of those sources; Tables 3 and 5 are CEOtudent editorial frameworks, and the worksheet’s salary, output and team figures are illustrative, not data. Deliberately not used: wage-premium figures for AI skills from commercial job-advertisement reports, which could not be opened in the original; numbers from commercial self-reported pay websites; the quality figure often quoted for the consulting experiment, absent from the abstract consulted; and any lifetime value of negotiating, not found in a primary text. CEOtudent ran no survey for this article.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
