Home

AI and Payroll: What Payroll Failures Cost, and What Is Known About AI Doing It

Sources checked: November 9, 2026. AI-assisted draft, pending editor review.

Payroll is where "usually right" is not good enough. The same inputs must give the same pay every time, and a mistake lands in someone's paycheck. Generative AI writes fluent answers, but it can also write a confident wrong one. That makes it a poor calculator and a useful assistant around the calculation. The record shows why.

The Wage and Hour Division's own data show how recoveries have moved. Fiscal year 2025 was the highest since 2019, although fewer employees were paid back than in 2019 to 2021.

The first three rows of the first table above are not AI failures. They show what payroll errors cost. We searched for a verified case of a generative AI tool causing a payroll error at scale and did not find one. The closest item is a November 2025 blog post relaying an anonymous account of an AI payroll tool halting a paycheck over an "irregular" timecard, which we could not verify. On the vendor side, UC Today reported on May 13, 2026 that UKG launched an AI payroll product that flags errors while payroll professionals keep approval control. A 2019 Censuswide survey commissioned by payroll software vendor MHR, of 251 HR and payroll managers at UK private employers with 1,000 or more staff, found 76% admitted paying employees incorrectly or late at least once in the past year, and 52% still used spreadsheets in payroll. It is a vendor-commissioned survey, but it shows the baseline is not perfect even without AI. So the advice below rests on how payroll works and on the AI cases that exist (MyCity, Air Canada), not on a documented AI payroll failure.

The Kronos and Phoenix cases both ended in employers improvising. Write a one-page plan now: who can approve an off-cycle payment, how you would pay from the last good payroll, how you would tell employees, and where the last export is stored. Test it once a year.

Payroll data is among the most sensitive a company holds: names, Social Security numbers and bank details. IBM's 2025 report put the average breach cost at $4.44 million globally and $10.22 million in the U.S. The FTC's business guide advises knowing what personal information you keep, keeping only what you need, protecting it, disposing of what you no longer need and planning for incidents. Read AI Tools and Employee Data before any pay data goes into an AI tool.