AI vs human employee cost, worked out from official wage data

Rarely one for one. Task exposure scores measure how much of a role's task time AI can speed up, not how much of the job disappears. Measured effects so far are small: a Danish study using administrative records rules out effects on earnings or hours larger than 2% (Humlum and Vestergaard, NBER 33777, 2026), and the strongest field result for customer support is a 15% gain for people using AI, not AI working alone (Brynjolfsson, Li and Raymond, QJE 2025). This calculator prices the share of tasks you expect AI to handle and charges human review against it.

Annual wage times the employer cost multiplier, divided by productive hours. For a customer service representative in the United States at the BLS median: $44,770 x 1.312 / 1,872 hours = $31.38 per productive hour. The 1.312 is BLS Employer Costs for Employee Compensation (June 2026) for office and administrative support, and 1,872 is 2,080 paid hours minus 8 holidays, 11 vacation days and 7 sick days (BLS, March 2026).

They are statutory minimums only: employer National Insurance and the auto enrolment pension in the UK (about 1.15), CPP, CPP2 and EI in Canada (about 1.08), and the 12% Superannuation Guarantee in Australia (1.12). The United States figure (1.23 to 1.31) comes from a survey of what employers actually pay, including health insurance and paid leave. Health care is publicly funded in those three countries, but voluntary benefits still exist, so enter your own multiplier if you know it.

Official statistics only. United States: BLS Occupational Employment and Wage Statistics, May 2025, released 15 May 2026. UK: ONS Annual Survey of Hours and Earnings 2025 provisional, released 23 October 2025. Canada: ESDC Job Bank 2025 wage file, based on the Labour Force Survey for 2023 to 2024. Australia: ABS Employee Earnings and Hours, May 2025, which publishes means only at this occupation level. Each row shows its occupation code and links to its source.

The default is Anthropic's observed exposure (March 2026): of the tasks in a role that a language model could speed up, the share actually seen in work related Claude usage, with automated use at full weight and assisted use at half weight. The second bar is the theoretical ceiling from Eloundou and colleagues (Science, 2024), rated by human annotators. They often disagree, sometimes by a lot, which is a reason to measure your own tasks before planning around either.

Because the work does not vanish. Anthropic measured (model graded) that success on automated tasks falls from about 60% for tasks under an hour to about 45% for tasks of five hours or more, and firms that adopted AI created new oversight tasks (Humlum and Vestergaard). No primary source publishes review minutes per AI task for these roles, so the default of 2 minutes on every AI handled task is an example to replace with your own measurement.

Usually far less than the review time. Model cost is tokens per task times the provider's price per million tokens, from my AI model leaderboard. At 4,000 input and 600 output tokens per task, Claude Sonnet 5.5 costs about $0.014 per task and Claude Haiku 5.5 about $0.0007 (prices checked 10 Oct 2026). Agent workflows with tool calls and long context use more tokens, which is why tokens per task is an input and sits in the sensitivity chart.

Only if something you pay for changes: fewer hires, less overtime, lower agency spend, or a person moved to work that earns more. Otherwise freed hours become slack or new work. The calculator shows freed hours as a full time equivalent so you can see the scale, but the saving is only real once that decision is made.