Payscale Puts AI Upskilling Against a 3.5% Raise Budget

AI-generated editorial illustration.

On October 15, Payscale published a blunt split in its AI workforce research: 74% of surveyed employers planned to invest in AI upskilling over the next year, while 56% of surveyed workers believed they should earn more after acquiring AI skills. Payscale had surveyed those groups separately in August. The numbers do not say that the same people will take a course or get a raise. They do set up the next conversation in a compensation meeting: after the company pays for training, what work has changed enough to pay more for?

Another Payscale survey supplies the budget backdrop. Among 1,266 organizations that submitted salary plans in May and June, the average planned U.S. salary increase budget for 2027 was 3.5% of payroll. That is a survey average, not a legal cap or a promise made by any particular employer. It is also a different sample from the AI research. For a company planning a conventional merit cycle near that average, though, the gap is concrete. A training program has its own invoice. The new salary requirement, if there is one, will reach payroll every year.

The report preview records two answers from the employer panel: 61% said they were rewriting job descriptions because of AI; 49% said salary structures had not kept pace. Payscale sells compensation software and recommends more frequent benchmarking. A buyer still has to decide whether the new skill changes the job’s scope, its market price, its result or only the tool used to finish familiar work.

At the annual review, a manager can approve a course, assign more complex work and nominate a higher rating. Finance may have authorized one pool of raises for the entire workforce. Employees can see a new external opening with an AI requirement and a pay range while their own job title remains unchanged. The organization must explain why one capability earns a premium at the front door and another earns only a certificate inside.

A planned budget meets a new skills claim

Payscale’s 2026-27 salary budget survey reports that U.S. employers in its sample gave average base-pay increases of 3.4% in 2026 and planned 3.5% for 2027. The median is the same in both years. These are annual increase plans gathered from Payscale clients and contacts, not a measure of every worker’s wage change. The survey also says 63% expected next year’s budget to stay about the same, 30% expected a larger budget and 8% expected a smaller one. Rounding accounts for the total above 100%.

The tenth of a point between years looks minor until a manager asks for exceptions. A firm with a $10 million payroll would have a $350,000 increase pool if it chose 3.5%. This is illustrative arithmetic, not a reported company plan. Ten specialists each receiving a $10,000 base-pay adjustment would use $100,000 of that pool in the first year, before benefits or later compounding. Finance would have $250,000 left for everyone else unless it opened another line. Calling the adjustment upskilling does not remove the cost.

Payscale’s survey distinguishes several kinds of increase. Merit recognizes performance. A promotion can move a person into a higher-level role. A market adjustment responds to outside offers or a changed benchmark. An across-the-board cost-of-living increase protects purchasing power. One employee may qualify for more than one, but combining all of them into a single percentage makes the compensation story harder to explain.

The pool also has a timing problem. Annual budgets are usually set before managers know exactly which trained workers will assume a materially different scope. Payscale’s newer AI Workforce Impact preview reports that 74% of employer respondents intend to invest in AI upskilling during the next 12 months. An intention to fund instruction is not an approved payroll adjustment for every participant. If a budget treats all those future outcomes as routine merit, the money for promotions and scarce specialties may already be committed when the outcomes arrive.

Ruth Thomas, Payscale’s chief compensation strategist, described the mismatch as AI capability demand moving faster than employers’ ability to price it. She has an interest in a solution that uses Payscale’s data. Yet the calendar itself poses a question that software cannot answer for finance: which share of the 2027 pool should be held for jobs whose scope has not yet changed? A reserve protects the later decision, but workers expecting routine merit will see a smaller pool now. An additional appropriation protects merit, but raises the cost of the AI program.

The plan is still a plan. WorldatWork’s fall salary budget survey was open through October 30 to compare what employers actually spent in 2026 with their spring projections for 2027. It had not published a fall result as of this article’s date. The existence of a second survey window is a reminder to label the Payscale 3.5% as a projection rather than as a settled 2027 outcome. A single company’s authorized budget could land above or below it.

The employer survey covers 500 HR, compensation and talent respondents at organizations with at least 100 employees. It is an online opt-in panel. The separate worker survey has 1,000 full-time employees and active job seekers at similarly sized organizations. Neither sample describes startup pay, all U.S. employers or a matched employer-worker pair. These answers show a split in expectations; they cannot predict a particular employee’s raise request.

A premium and a raise measure different things

In June, PwC reported a 62% average wage premium for jobs requiring AI skills. The headline traveled because it is large. It is not the raise a worker receives after learning to use a new tool. PwC compared jobs that require AI skills with other jobs in a global job-ad dataset, after its modeling choices. The jobs differ in occupation, seniority, industry and other traits. The statistic describes a labor-market association, not a contractual entitlement or the marginal value of one course.

Payscale’s new preview makes the same point at a smaller scale. It says agentic AI and machine-learning capabilities may carry 9% higher median pay for a Business Process Analyst II, while those competencies show no extra median premium for a Machine Learning Engineer, where they are already ordinary requirements. The report does not publish enough detail to transfer either differential to an employer’s own pay grid. The comparison does show why a single companywide “AI premium” would be a poor rule. A skill can be scarce in one job and already included in another job’s price.

That creates two decisions. The first is what the external market pays for a role with a defined scope, location and level. The second is whether an incumbent’s work now matches that role. A new title on a course certificate answers neither. The manager needs evidence of changed responsibilities or output; compensation needs a comparable job; finance needs a cost it can support after the first year.

There is a respectable argument against paying a separate premium for every new AI requirement. When most people in a job family need the same tools, the capability becomes part of the normal role. An employer that gives protected learning time, maintains an honest salary band and lets workers progress can treat basic fluency as a baseline. Paying every user an add-on would shift money away from employees whose scope or results changed most.

But “baseline” needs a test. It should describe work expected of comparable employees, with comparable access to training. Applied to one person’s expanded responsibility, the word becomes a way to postpone a pay decision.

An external market number can also mislead in the other direction. PwC’s 62% figure covers jobs that require AI skills, not identical jobs before and after a short course. Payscale’s 9% analyst example is closer to a role-level differential but still comes from the vendor’s own data. Neither statistic tells an employer what its own employee produced or what it would cost to replace that person. The useful comparison is an actual role profile against an outside range, followed by a check of internal employees already doing the work.

Payscale says 40% of employer respondents struggled to find candidates with sufficient AI fluency. The answer measures reported difficulty; it cannot establish what every vacancy pays. The preview’s separate Lightcast analysis reports an 8.7-fold increase in postings within a custom AI sector since 2022. Postings are not hires. That sector can contain software specialists and managers using AI in ordinary work. Salary bands need a defined capability, not the word AI in an advertisement.

For the employee, the task list is easier to see than the survey. Someone asked to supervise an automated workflow may now be responsible for exceptions, corrections and an answer to the customer. Another person may use a summarization tool to complete the same job faster. Both can honestly say their work uses AI. The pay decision should describe the difference in risk, decision rights or scarce skill, rather than the brand of the tool.

Training creates a second pay decision

August’s employer answers show where the approval chain breaks. While 61% said they were rewriting job descriptions because of AI, 48% said their market benchmarks no longer reflected the capabilities those jobs now required; 49% said salary structures had not kept pace. These answers concern employers’ perception of their own systems. They do not establish that half of posted salaries are wrong. They show that the job description may move before the authorized range.

For a manager, the easiest move is often to keep the old title. It avoids a headcount request and gives the worker a chance to learn. Later, when that worker owns a model evaluation, redesigns a service process or takes responsibility for a decision, the job may no longer be the one compensation priced. A recognition award can be fair for a temporary assignment. It becomes less convincing if the higher scope becomes permanent.

Payscale surveyed employees separately. 56% said they should receive higher pay for developing AI skills. That is an expectation, not evidence that they had received an offer or that a newly trained worker had improved measured output. The report also asks employers about strategy, but its public preview lists premium, baseline and undecided answers that do not form a clean exclusive partition. Treating those percentages as a single market share chart would make the evidence look tidier than it is.

Training costs arrive in different accounts. Learning may pay for a course or software license. A team gives up paid hours while someone practices. Payroll sees the effect later, if the work changes and the employee seeks a durable adjustment. When a skill leads to production work, the expected salary cost belongs in the original business case. Otherwise, a training plan can appear cheaper than the work it creates.

This does not mean every learner should be paid immediately. A company can make AI literacy a baseline expectation, just as it once made spreadsheet use or digital collaboration standard in many roles. It must provide time and access to learn that baseline. A worker asked to apply a scarce capability to a different job has a stronger case. Training and pay meet when the job changes, even if the HR system has not yet recorded it.

Illustrative scenario, not a reported employer case or meeting. A business analyst completes company-funded training, then begins checking an automated billing process. For the first month, the analyst reviews sample outputs with a supervisor. Six months later, the analyst decides which errors require manual correction, changes the process for recurring failures and signs off on a weekly quality report. The job title has not moved. The decision rights have.

The employer has at least three defensible options: update the role and range, offer a temporary assignment payment while the process is tested, or return the analyst to the original duties. Leaving the expanded responsibility in place while treating the course as the only reward is also a decision, but it should appear in the budget as a retention risk. The worker should be told what evidence would lead to a new level and when that decision will be made. Otherwise the outside posting becomes the only visible price for the work.

AI-generated editorial illustration. The shared pool has several claims before a raise is approved.

Which jobs changed enough to reprice

Artificial Intelligence Director ranks ninth by growth in job pricings in Payscale’s Q3 labor market report. The report puts technology industry wage growth at 6.4%, down from 6.9% in the previous quarter. A job pricing means an employer sought a benchmark; it does not record an accepted offer or a filled role. Payscale’s Peer dataset contains 10.4 million employee salaries across more than 45,000 jobs and 4,500 organizations in 60 countries. That is a sizable commercial dataset, not a random national census.

The title of the role matters. A director who owns an AI program may need a different range because authority, budget and leadership scope have changed. An analyst learning to use AI may need a market check for a narrower skill differential. A technician or customer-service worker may face new tools and more work without a new title. A compensation committee should avoid importing a director benchmark into that worker’s review, or using the worker’s old range to price the director.

There is an outside check on the wider wage climate. The U.S. Bureau of Labor Statistics reported on October 14 that real average hourly earnings for private nonfarm employees fell 0.5% from August to September and 0.7% from September 2025. In September, average nominal hourly earnings rose 0.1% while the consumer price index rose 0.6%. This is an economy-wide average adjusted for prices. It does not identify what happened to people who learned AI or to an individual employer’s pay pool.

ADP’s September payroll report follows a different unit: changes in individual workers’ pay. Its median base-pay growth was 3.2% year over year overall, 3.0% for job stayers and 4.8% for job changers, drawn from more than 14.7 million year-over-year pay-change observations each month. These medians cannot be subtracted from the BLS mean to calculate an AI premium. They do suggest why an internally trained employee may look at the external market when the internal raise route is constrained.

Four routes through the compensation meeting

Before assigning a percentage, a compensation committee has to choose a type of pay action. This matrix is an editorial proposal. It describes neither a Payscale product feature nor an observed employer policy.

Work after AI training	Evidence to request	Pay route and budget owner	Fairness check
Same job, new everyday tool	Updated task list, access to training, normal performance evidence	Baseline capability in the existing band; manager and learning budget fund instruction	Were all comparable workers given time and access to learn?
More responsibility inside the same level	Documented decision rights, sustained workload and market comparison	In-band merit or targeted adjustment; manager and compensation share the case	Are peers with the same scope treated alike?
New level or scarce specialty	Approved role profile, changed accountability, credible external benchmark	Promotion or market premium; business sponsor and finance approve recurring payroll	Will an outside hire be paid differently for the same work?
Measured result beyond the role’s standing scope	Pre-agreed outcome, baseline and attribution, including failed or corrected work	Time-limited bonus or project award; operating budget owns the result	Does the measure reward quality rather than raw AI activity?

A one-time bonus and a base-pay increase have different future costs. A bonus can fit a six-month pilot with a measured result. When the new responsibility becomes permanent, a string of discretionary awards leaves the employee without a stable range. Promotion requires an approved role at a higher level. A market premium requires comparable jobs, rather than a headline about average AI pay.

This matrix also prevents an easy evasion. A company may describe expanded responsibility as “AI fluency” and keep it in the baseline row to save money. The worker may describe every use of a tool as a scarce skill to seek a premium. Both sides need the task list, decision rights, duration and outside benchmark. If those records are absent, an argument about pay will turn into an argument about labels.

Finance, the manager and compensation each hold a different part of the answer. Finance can account for the recurring cost of base pay and the one-time cost of a course. The manager knows whether the trained person now resolves harder cases or merely finishes familiar tasks more quickly. Compensation can find peers inside and outside the company and test whether the pay action creates compression. No one owner can infer all three from a certificate or a vendor’s premium statistic.

There is a missing incentive too. A manager whose team absorbs the new work may benefit from faster service while the salary adjustment is charged to a central pool. If the business unit gets the output and the rest of the company funds the premium, requests will multiply. If the unit must fund every adjustment, it may delay acknowledging changed jobs. A short written rule can connect a request to the budget that benefits: the sponsor names the changed work, finance states whether it is a recurring commitment, and compensation checks peers before an offer or raise is approved.

The arithmetic can be made explicit. In the $10 million payroll illustration, finance could decide that the $350,000 planned increase pool covers normal merit, while promotions and documented market corrections require separate approval. Or it could keep one pool and publish how much is reserved for each purpose. Either design has a cost. What cannot work indefinitely is promising a broad training program while managers assume every newly skilled worker will be paid from an unspecified future source.

The worker can see two prices

PwC’s September global workforce survey found that 64% of nearly 50,000 respondents had used AI at work in the prior year, and daily generative-AI use rose from 14% to 22%. PwC divides respondents into a 14% group with high AI advantage and high-demand capabilities and a 56% core workforce. The groups are defined by survey answers and do not correspond to Payscale’s U.S. pay-budget sample. The same survey says only 34% could pay their bills and have money left over. Pay is not an abstract recognition issue for most workers.

Payscale’s own salary survey reports that a quarter of responding organizations believed they lost talent in 2026 because increases were insufficient. That is a reported employer perception, not a measured resignation rate. ADP’s gap between the 3.0% median base-pay gain for stayers and 4.8% for changers adds a different observation: moving employers and staying put were associated with different pay changes in its September payroll panel. It does not establish that AI training caused any one person to move.

An internal posting can make the difference visible. The worker who completed the course may see the same AI capability listed in a newly opened job with a higher range. That range may legitimately include more scope, management responsibility or a different labor market. If it does, the company should name the distinction. If the duties are the same, the problem is pay compression, and a course certificate will not explain it away.

Pay transparency also forces a second check: who got access to the course and the work that produces a premium? The employee who already works near an AI project can build a portfolio. Another worker with the same potential may remain in a job with no tool access, no protected learning time and no chance to demonstrate the capability. A premium based on outcomes can still reproduce unequal opportunity if the employer never looks at how people reached those outcomes.

The worker needs an answer that can survive a conversation without market jargon. “Your 2027 merit increase is 3%, and your new process responsibility is under review for a level change in March” is more specific than “AI skills are valuable here.” It separates a decided raise from a pending role review. An employer can decline a premium, but should say why: the skill is already priced into the role, the job has not changed, or the budget rejected an otherwise justified adjustment. Each reason creates a different choice for the employee.

November will bring a fuller test

Payscale expects its complete AI Workforce Impact Report in November. The public preview gives sample sizes and headline answers, but no matched employer-worker panel or reliable raise-per-skill estimate. It does not show whether planned upskilling became completed training, roles were reclassified or premiums entered payroll. Buyers of compensation data need those denominators before turning survey percentages into policy.

The next 2027 budget meeting can make a smaller decision now. Finance can ask compensation to list which AI-related jobs have actually changed scope, which employees already do that work, and how internal pay compares with a genuinely comparable outside role. Learning can then put the cost of instruction next to the cost of a possible recurring raise. Managers can say what work they will assign after the course, and when the result will be reviewed.

In the illustrative analyst case, the next review may bring an external posting with the new skill in its requirements and a higher range. The manager has a completed course and six months of changed decision rights. Finance has an approved pool. The unresolved line is the price of keeping that work inside the company.