To calculate AI automation ROI, compare the fully-loaded cost of the manual process (hours × loaded hourly rate + error costs + opportunity costs) against the automation's build cost, running cost, and residual human-review time. A well-chosen automation typically reaches payback in three to six months; if projected payback exceeds twelve months, pick a different process.
Why most ROI estimates are wrong in both directions
Vendors overstate ROI by counting every saved minute as if it converts to payroll savings. Skeptics understate it by counting only labor, ignoring the error costs and missed revenue that often dwarf it. The honest calculation has three benefit categories and a cost side people routinely forget.
The cost of the manual process (the baseline)
- Direct labor: hours per week × fully-loaded hourly cost (salary + taxes + benefits + overhead — typically 1.25–1.4× base pay).
- Error costs: rework, billing disputes, refunds-by-apology, compliance exposure. Manual data entry error rates of 1–4% are typical; each error has a cleanup cost.
- Latency costs: what slow processing loses — leads that cool, invoices that age, customers who churn waiting for answers. Often the largest and least-counted line.
- Opportunity cost: what your best people would produce if they weren't doing this.
The cost of the automation (the honest denominator)
- Build: design, integration, testing — one-time.
- Running: model API usage, platform subscriptions, hosting — monthly, usually modest.
- Human residual: someone still reviews exceptions and monitors quality. Count it.
- Maintenance: processes change; budget for periodic tuning (this is what support retainers cover).
Worked example: automating invoice processing
A distribution company processes 2,500 supplier invoices a month. Two AP clerks spend a combined 50 hours a week keying, matching, and chasing — at a loaded cost of $42/hour.
| Line | Manual (annual) | Automated (annual) |
|---|---|---|
| Labor | $109,200 (2,600 hrs) | $16,400 (390 hrs exception review) |
| Error cleanup (2.1% → 0.3%) | $18,900 | $2,700 |
| Early-payment discounts missed | $11,000 | $1,500 |
| Build cost (one-time, amortized yr 1) | — | $28,000 |
| Running + maintenance | — | $9,600 |
| Year-1 total | $139,100 | $58,200 |
Year-one saving: roughly $81,000, with payback in month four. Year two, without the build cost, the same system returns over $110,000 annually. These are representative numbers from our engagements, rounded for clarity — your inputs will differ, which is exactly why the audit comes before the proposal.
The thresholds that matter
- Payback under 6 months: automate now — this is the tier the first project should come from.
- Payback 6–12 months: automate after the quick wins have built trust and freed budget.
- Payback over 12 months: usually the wrong process, or the right process at the wrong scale. Revisit when volume grows.
What doesn't show up in the spreadsheet
Consistently, the benefits clients rank highest a year in are the ones the model can't price: the peak season that didn't require panic hiring, the audit that took an afternoon because every document had a trail, the employee who stopped job-hunting when the soul-crushing part of the role disappeared. Count the spreadsheet benefits to justify the project; expect the uncounted ones to justify the strategy.
