AI in compensation: What to automate
Automate the data work. Keep the pay decisions with people who can explain them.
Short answer: AI tools are useful for compensation data work: Cleaning payroll exports, suggesting job matches, running scenarios and flagging pay gaps for review. They should not make or quietly shape pay decisions. A match, a market rate or a raise recommendation from a tool is a draft. A person who can explain it decides. The reason is recorded.
A faster pay review
Picture a 210-person company in Lincoln that makes grain logistics software. Its one-person compensation function, a senior HR generalist, has just bought a compensation tool with AI features. In the first week it does what took her a month last year. It cleans the payroll export, suggests survey matches for every job, proposes a market rate for each and drafts a raise recommendation for every employee based on rating, position in range and budget.
She is delighted until the questions start. A manager asks why his data engineer's market rate dropped 8% from last year. The tool had matched the job to a reporting analyst role in the survey, because the job title included the word "data". A team lead asks why every raise in her team is within a few tenths of a percent of the others. The recommendations followed last year's pattern, which had followed the year before. And when the head of people runs a quick check by gender, the recommended raises show a small but consistent gap that nobody can trace to a reason.
Nothing in the tool was broken. It did what it was built to do: Produce plausible numbers fast. What it could not do was say why any number was right.
This is a composite of a common situation at companies this size, not a client. The details and figures are illustrative.
Why using AI in compensation is hard
AI in compensation sets two good goals against each other. The first is efficiency. Compensation work at a company of this size is often one person with spreadsheets. The data work alone can fill a pay review. Tools that clean data, suggest matches and model scenarios free that person for the judgment that matters. The second is explainability. Every pay number eventually has to be explained to a manager, an employee, an auditor or a court. A number that came from a model nobody can explain is a weak foundation for any of those conversations.
Adoption is already broad. In a survey of 1,001 US organizations in July 2026, 70% reported at least some automation in compensation, though only 1% described themselves as advanced. Among the tasks where AI was used, market pricing and benchmarking led at 53%, followed by salary increase recommendations at 50% and job matching and leveling at 49% [1]. Those are exactly the tasks where an error is hardest to see and most expensive to fix.
The data underneath is moving too. A compensation data provider reported in September 2026 that 48% of respondents said their market benchmarks no longer reflected the skills roles require and 61% were rewriting job descriptions because of AI [2]. These are the provider's own figures. A tool that matches on old descriptions or old surveys produces confident answers to the wrong question.
And the rules are arriving. In California, anti-discrimination rules for automated decision systems in employment have applied since October 2025 [3]. In the EU, AI used to make decisions affecting the terms of work relationships is classed as high-risk, with duties for employers who use it applying from 2 December 2027, including human oversight by people with the authority and training to exercise it [4] [5].
The risk is not that the tools are bad. Many are good at what they do. The risk is that their output looks finished. A table of matched jobs, market rates and proposed raises has the same appearance whether every row was checked or none was. At a small company, where one person owns the whole process, the pressure to accept a finished-looking table is strong. The errors it contains surface months later in a resignation, an equity review or a candidate's question about a posted range.
Two terms will help. Job matching is the step that links each internal job to the closest job in a salary survey. A raise recommendation is a proposed increase for a person, usually generated from rating, position in range and budget.
Five traps in using AI for compensation
"Let the tool match the jobs"
What it looks like: The tool suggests a survey match for every job based on title and description and the matches are accepted in bulk.
Why smart teams fall into it: Matching is slow, tedious work and the suggestions look reasonable. Most of them probably are.
What it costs: The ones that are not are where market pricing goes wrong. A match based on words in a title can put a data engineer next to a reporting analyst or a senior individual contributor next to a manager. Every number built on that match inherits the error: The range, the offer, the raise. Matching is a judgment about the content and level of a job, which is exactly what a title-based model cannot see. Use the suggestions as a first pass and have a person confirm each match against what the job actually does. Our guide on job matching for salary benchmarking covers where matches quietly go wrong.
"Take the recommended raise"
What it looks like: The tool proposes a raise for each employee and managers accept the proposals with few changes.
Why smart teams fall into it: The proposals follow the merit guide and the budget, so they look consistent. Changing them takes time managers do not have.
What it costs: A recommendation trained on past decisions repeats them, including the patterns nobody would defend. If last year's raises were slightly lower for one group, this year's recommendations will be too, with an air of objectivity that makes the gap harder to question. Managers who accept proposals without engaging also lose the ability to explain them. In the EU, decisions affecting the terms of work made with AI will need human oversight by people with the competence and authority to override the system [5]. A recommendation is a draft. The decision belongs to a manager who can say why. The result should be checked by group before letters go out.
"Market price from a model"
What it looks like: The tool produces a market rate for each job, sometimes from its own data blend, sometimes from sources it does not disclose.
Why smart teams fall into it: It is fast, it covers every job and it comes with a confidence score.
What it costs: A market rate the company cannot trace is a market rate it cannot defend. Under US federal equal pay guidance, market value counts as a reason for a pay difference only if the employer can show it actually assessed the market [6]. "The tool said so" is not that showing unless the company knows which data, which match and which date the number rests on. Use tools to assemble and age data from sources you know. Keep the source of every market rate on record. Where a tool blends several sources into one number, ask the vendor what the blend contains and how often it changes. A rate that moves when the vendor changes its method, rather than when the market moves, should be treated with caution.
"Paste the payroll into a general chatbot"
What it looks like: To save time, someone uploads a payroll export with names, salaries and ratings into a general-purpose AI assistant and asks it to find anomalies.
Why smart teams fall into it: General assistants are good at spotting patterns in tables and the analysis is genuinely useful.
What it costs: Pay data is among the most sensitive personal data a company holds. Sending it to a tool the company has not approved, under terms nobody has read, can breach the company's own commitments to employees and, for staff in some places, data protection law. The analysis can be done with names removed, in a tool the company has approved for personal data, with the same result. The question is where the data goes, not whether AI is involved. Most of the useful pattern-finding in pay data does not need names at all. Job, level, location, rating, pay and a coded identifier are enough to find the gaps worth a closer look.
"Let it write the explanation"
What it looks like: Managers use an AI assistant to draft the message that explains a raise, a range or a pay decision to an employee.
Why smart teams fall into it: Pay conversations are hard to write and the drafts sound calm and professional.
What it costs: A fluent explanation is not the same as a true one. A draft written without the real reason will supply a plausible one, often the budget or the market, which may not be why the number came out as it did. Employees compare notes and a generic or inaccurate explanation destroys more trust than a short honest one. In the EU, workers will have the right to ask for information about pay levels and the criteria used to set them [7]. The answers have to be right. Our guide on explaining a small raise to a strong performer covers what a good explanation contains.
What sound use of AI in compensation looks like
Sound use treats AI as an assistant for the data and modeling work and keeps every pay decision with a person who can explain it. At 50 to 500 people, the approaches that hold up share these qualities.
The line is written down. One page says which tasks tools may do on their own, which they may draft for a person to confirm and which they may not touch. Data cleaning and scenario modeling sit on one side. Levels, final matches, pay decisions and explanations sit on the other.
Every number has a traceable source. For each market rate and each recommendation, the company can say which data, which match and which rules produced it.
Matches are confirmed by a person. Tool suggestions are a first pass. Someone who knows the job confirms the match against its content and level. Matches that change from one year to the next are flagged, because a changed match moves every number built on it.
Recommendations are checked by group. Before decisions are final, proposed raises and offers are reviewed by gender, team and position in range, whatever produced them.
People can override and do. Managers and the compensation lead are expected to change recommendations where they disagree and to record why. A process where nobody ever overrides the tool is not oversight.
Personal data stays in approved tools. Pay data goes only into systems the company has approved for it, with names removed where they are not needed.
The tool's changes are noticed. When a vendor updates its model or its data, someone checks whether matches and rates moved and why. A tool update can change a pay structure without anyone deciding to.
Scenario modeling is used before decisions, not after. Tools are at their best when they show what a budget, a range change or a different merit guide would do to every person before anyone decides. That is where automation adds judgment rather than replacing it.
Explanations come from the reason. Managers can use help with wording, but the content of any pay explanation comes from the actual reason for the decision.
The calls only you can make
The tasks you hand over. Data cleaning and modeling are easy to hand to tools. Matching, pricing and recommendations need a decision about how much review each gets.
The tools you approve. A specialist compensation tool, a general assistant under a company agreement or neither. Each has different data terms and different transparency.
What managers see. Whether managers see the tool's recommendation or only the guide and their budget. Seeing the recommendation saves time and anchors their judgment.
Who owns the review. Someone has to own the check of matches, rates and outcomes by group. At this size it is often one person. Name them and give them time for it in the pay review calendar.
How you choose a vendor. Whether a tool must show the sources behind each market rate and each match before you buy it. Transparency of method is easier to require at the start than to win later.
What you tell employees. Whether and how you tell people that tools were used in pay decisions. In some places this will be required. Everywhere it affects trust.
The data and the law
- Adoption. In a survey of 1,001 US organizations in mid-July 2026, 70% reported at least some automation in compensation and 1% described themselves as advanced. AI was used for market pricing and benchmarking by 53%, salary increase recommendations by 50% and job matching and leveling by 49% [1].
- Benchmarks and descriptions. A compensation data provider's September 2026 preview found 48% of respondents saying benchmarks did not reflect the skills roles require and 61% rewriting job descriptions because of AI [2]. These are the provider's own figures.
- California. Anti-discrimination rules covering automated decision systems in employment decisions have been in effect since 1 October 2025 [3].
- European Union, AI Act. AI systems used to make decisions affecting the terms of work relationships, promotion and termination, or to monitor and evaluate performance, are classed as high-risk (Annex III, point 4) [4]. Employers using them will need to assign human oversight to people with the necessary competence, training and authority and to inform workers' representatives and affected workers before use (Article 26) [5]. These duties apply from 2 December 2027 [8].
- United States, equal pay. Market value counts as a reason for a pay difference only if the employer proves it assessed the market [6].
- European Union, pay transparency. Workers can ask in writing for their own pay level and average pay levels by sex for workers doing the same work or work of equal value (Article 7) [7].
This is general information, not legal advice. Rules and their national versions differ. Check the rules where you employ people.
Our view
We think AI belongs in compensation work and does not belong in compensation decisions. The data work is where tools save the most time and do the least harm: Cleaning, assembling, modeling and flagging. The decisions are where errors are hardest to see and where the company will be asked to explain itself.
So we would write down where the line sits, keep a traceable source for every number, have a person confirm every match and every pay decision, check outcomes by group whatever produced them and keep pay data in approved tools. Used that way, AI makes a small compensation function faster without making it harder to defend. The time it saves is best spent on the judgment it cannot do.
Where we help
Our pay policy work sets the line for tools in your compensation process: Which tasks they may do, what a person confirms, how matches and rates are traced and how outcomes are checked by group before decisions are final. The starting point is compensation policy, quoted within one business day. If a conversation would help first, Book a free call.
Sources
- Mercer, US employers stick with moderate pay increases for 2027 (31 August 2026; QuickPulse US Compensation Planning Survey, 1,001 organizations, mid-July 2026), checked 2026-10-30
- Payscale, AI Workforce Impact Report preview, press release (24 September 2026; vendor survey), checked 2026-10-30
- California Civil Rights Department, Civil Rights Council Secures Approval for Regulations to Protect Against Employment Discrimination Related to Artificial Intelligence (30 June 2025; in effect 1 October 2025), checked 2026-10-24
- EU AI Act, Annex III point 4, as published with the Digital Omnibus amendments by the AI Act Explorer, checked 2026-10-24
- EU AI Act, Article 26 (obligations of deployers of high-risk AI systems), same source, checked 2026-10-24
- US EEOC, Compliance Manual Section 10: Compensation Discrimination, 10-IV.F.2.g (market factors), checked 2026-10-30
- Directive (EU) 2023/970 on pay transparency, Article 7 (EUR-Lex), checked 2026-10-28
- EU AI Act, Article 113 as amended by Regulation (EU) 2026/1744 (Annex III high-risk rules from 2 December 2027), same source, checked 2026-10-24