The ROI of AI consulting, worked out with federal wage data
A return is hours removed times the true cost of an hour, set against the first-year price including running costs. Three worked examples with federal wage data, and the inputs to measure yourself.
Contents
Why most ROI claims cannot be checked
Every AI proposal carries a return on investment, and few can be verified. The hours saved come from the vendor, the hourly cost is often a salary without benefits, and the investment leaves out the cost of running the system after it is built. A buyer can rebuild the calculation from public data in an afternoon, and should, because the inputs that decide it are specific to the company.
The formula
Three inputs determine the annual saving:
- Hours. How many hours a year the process consumes, measured from timesheets, queue counts or a two-week log.
- Share removed. The proportion of those hours the system takes over. This is the uncertain input, and the one a first phase should measure.
- Loaded hourly cost. Wages plus benefits. Wages make up 70% of private employers' compensation costs and benefits 30%, according to the Bureau of Labor Statistics' June 2026 data1, so a median wage divided by 0.70 approximates the loaded cost.
Annual saving = hours × share removed × loaded hourly cost. Payback in months = first-year cost ÷ (annual saving ÷ 12). First-year cost is the assessment, the build and twelve months of running costs, not the build alone.
Loaded hourly costs from federal data
| Role | Median wage, May 2025 | Approximate loaded cost |
|---|---|---|
| Customer service representative | $21.532 | $30.76 |
| Bookkeeping, accounting or auditing clerk | $24.363 | $34.80 |
| Management analyst | $48.974 | $69.96 |
| Data scientist | $57.805 | $82.57 |
Loaded costs are the median wage divided by 0.701, and exclude space, equipment and management time, so they understate the full cost of an hour.
Three worked examples
The examples below use published prices and federal wage data. The hours and the share removed are assumptions chosen to show the arithmetic, not findings, and a company should replace them with its own measurements.
| Example | Hours a year | Share removed (assumed) | Loaded cost | Annual saving |
|---|---|---|---|---|
| A customer-service queue | 6,000 | 30% | $30.76 | $55,368 |
| Bookkeeping and invoice matching | 3,000 | 40% | $34.80 | $41,760 |
| Analysts' weekly reporting | 1,000 | 50% | $69.96 | $34,980 |
Set against first-year costs built from published prices:
- A fixed-price project with managed service costs $33,500 to $71,500 in year one, before tools and usage: a $3,500 to $8,500 audit, a $12,000 to $45,000 sprint and twelve months of managed service from $1,5006. Against the customer-service saving of $55,368, payback falls between about 7 and 16 months.
- A company-owned platform with ongoing engineering costs from $55,800 in year one for hosting and developer support, plus 5.5% on AI credits and the architect engagement7. It pays back in the first year only if the saving across the processes it runs exceeds that figure, which is why it suits companies with several queues rather than one.
- The same bookkeeping project bought by the hour at the GSA median AI rate of $18789 costs $37,400 for 200 hours, before running costs, against a saving of $41,760.
The pattern is consistent. A single process of a few thousand hours supports a fixed-price project; a platform with a monthly engineering allowance needs several processes to clear its fixed cost; and every route is decided by the share removed, which no published source can supply.
A plan that makes the number real
- Log the hours for two weeks on the process you would automate first, and annualise them.
- Use loaded costs, not salaries, for the people doing the work.
- Buy a fixed-price first phase whose deliverable is the measured share removed on that one process.
- Recalculate with the measured share before committing to the next phase.
- Count the running cost every year, and stop or change what no longer clears it.
At the end, a company holds a measured saving per process and a first-year cost per route, and the decision follows from comparing the two. The hourly rates guide gives the benchmarks for checking labour costs in a proposal, and the main guide compares providers on how much of their price they publish.
Questions buyers ask
How do you calculate the ROI of AI consulting?
Annual saving equals the hours a process consumes, times the share AI removes, times the fully loaded hourly cost of the people doing it. Compare that saving with the first-year cost, which is the assessment, the build and twelve months of running costs. Payback in months is the first-year cost divided by the monthly saving.
What is a fully loaded hourly labour cost?
It is wages plus benefits. US Bureau of Labor Statistics data for June 2026 shows wages are 70% of private employers' compensation costs, so dividing a median hourly wage by 0.70 gives an approximate loaded cost before overheads such as space and equipment.
What share of a process can AI remove?
There is no reliable general figure, which is why the share should be measured in a fixed-price first phase on a single process. Vendor estimates of hours saved are illustrations, not results, until they are measured in your own operation.
Sources
- Employer costs for employee compensation, June 2026. U.S. Bureau of Labor Statistics, 9 Sep 2026. Accessed 8 Nov 2026.
- Occupational Outlook Handbook: Customer service representatives. U.S. Bureau of Labor Statistics. Accessed 8 Nov 2026.
- Occupational Outlook Handbook: Bookkeeping, accounting, and auditing clerks. U.S. Bureau of Labor Statistics. Accessed 8 Nov 2026.
- Occupational Outlook Handbook: Management analysts. U.S. Bureau of Labor Statistics. Accessed 8 Nov 2026.
- Occupational Outlook Handbook: Data scientists. U.S. Bureau of Labor Statistics. Accessed 8 Nov 2026.
- What an AI consultant costs. Main & Machine. Accessed 8 Nov 2026.
- Pricing. OpenLazarus. Accessed 8 Nov 2026.
- CALC+ ceiling rates: labor categories matching “artificial intelligence” (MAS contracts). U.S. General Services Administration. Accessed 8 Nov 2026.
- CALC+ ceiling rates: labor categories matching “machine learning” (MAS contracts). U.S. General Services Administration. Accessed 8 Nov 2026.
Updated 8 Nov 2026: First published.