AI ROI for Small Business: Formula, Example and Review Costs
Disclosure: We may earn a commission when you purchase through links on this site, at no extra cost to you. How we make money.
Quick answer: Measure AI value against a comparable manual baseline, then subtract drafting, review, correction and exception-handling effort. Include subscription and setup costs. Hours saved are capacity, not automatically profit. A fast output that fails your quality or risk checks should not pass just because its spreadsheet return looks positive.
Updated September 14, 2026. This is an AI-assisted planning method with fictional numbers, not a customer case study, financial forecast or independently validated benchmark.
Choose one task and define acceptable output
Start with a bounded task such as preparing an internal weekly summary from approved inputs. Specify what a usable result must contain, who reviews it and what cannot be sent automatically. Avoid mixing different jobs into one average: a routine summary and a difficult exception are not comparable units.
Write the acceptance criteria before the pilot. For a summary, those might include accurate figures, traceable sources, no invented explanations and a clearly marked missing-data section. For customer-facing material, identify the additional approval steps. Do not remove review merely to make the time-saving calculation look better.
The formulas
Net minutes saved per task = manual baseline minutes − AI-assisted minutes. AI-assisted minutes include preparation, prompting, checking, corrections and exception handling, not just generation time.
Monthly capacity value = net minutes saved × monthly task count ÷ 60 × hourly value assumption.
Net monthly value = monthly capacity value + separately evidenced incremental margin − recurring costs − allocated setup costs.
Planning ROI = net monthly value ÷ total monthly costs × 100. If the cost denominator is zero, report ROI as not applicable rather than an infinite percentage. State the period and cost treatment beside the result. This is a planning convention, not an accounting standard.
A worked example that includes correction time
| Input | Assumption |
|---|---|
| Manual baseline | 20 minutes per task |
| AI drafting and preparation | 5 minutes per task |
| Review | 8 minutes per task |
| Correction and exception effort | 5 minutes per task |
| Net saving | 2 minutes per task; 80 minutes per month |
| Hourly value assumption | $45; monthly capacity value $60 |
| Recurring tool costs | $30 per month |
| Setup allocation | $120 spread over 3 months: $40 per month |
| Net monthly value during allocation | $60 − $30 − $40 = −$10 |
| Planning ROI during allocation | −$10 ÷ $70 × 100 = approximately −14.3% |
Counting only five minutes of drafting would imply a 15-minute saving and $450 of monthly capacity value. Including review and corrections changes the decision. After the setup allocation ends, this example produces $30 of positive monthly capacity value, but only if the assumptions hold and the recovered time is useful.
Capacity is not the same as cash
If the same person works the same paid hours after adopting a tool, a faster task may improve available capacity without reducing payroll. To claim a cash saving, identify the actual expense avoided. To claim additional profit, show additional contribution margin that is plausibly attributable to the change, not all revenue booked during the pilot.
Do not count the same recovered hour once as labor savings and again as revenue capacity. Keep realized cash effects and estimated time value in separate columns. Use a lower hourly assumption and a no-revenue-uplift scenario to see whether the decision depends on optimistic inputs.
Record enough detail to explain a bad week
For each case, log task type, baseline, assisted time components, pass/fail result, rework reason and any escalation. Record failures instead of dropping them from the average. Note changes to the workflow or model so you do not combine incompatible versions without explanation.
A pilot decision should consider the distribution, not only the mean. Check whether a small number of difficult cases consume most of the savings. Decide what to do when source data is missing, a reviewer is unavailable or the output includes an unsupported claim. The free AI support workflow demo illustrates a bounded review and escalation path; it does not establish your own ROI.
Use a quality gate before a financial gate
A positive estimate does not authorize a risky use case. If the process cannot protect appropriate data, obtain required review or recover from errors, stop or narrow the pilot. The NIST Generative AI Profile is a broader risk-management reference. Linking it here does not mean this worksheet or your process is certified or compliant.
Keep, revise or stop
- Keep: quality requirements pass, the effort is repeatable and the conservative value case is positive.
- Revise: the task has promise but a specific correction pattern consumes the benefit. Narrow the scope and repeat a comparable test.
- Stop: material risks remain uncontrolled or the workflow costs more effort than it saves without another documented benefit.
The $39 Small-Business AI Runbook includes a quality log for baseline, draft, review and correction time, plus a rollout plan and risk-aware use-case scorecard. It gives you a place to record this evidence, not a guarantee that AI will pay for itself.