Most AI ROI calculations are theater. They count hours saved per task without netting against the costs (license, training, internal time, output review) — and then compound speculative annual savings into a number that looks great on a slide and means nothing. This is a more honest methodology.
ROI = (Net benefit / Total cost) over a defined period. "Net benefit" means realized value after accounting for the cost of using the tool — not the gross time "saved."
Period matters. For AI tooling, the right frame is usually 12-month rolling, with a separate first-90-days view that often shows negative ROI (which is fine — you are buying the foundation).
| Category | Often missed? | Example |
|---|---|---|
| Direct license / platform | No | Claude Team \$30/seat/month |
| Implementation | Sometimes | Treetop engagement, internal build hours, vendor onboarding |
| Training | Often | Onboarding sessions, training material, time spent learning |
| Internal admin | Almost always | AI lead time, weekly office hours, prompt library maintenance |
| Review & QA | Almost always | Human-review time for AI outputs, error-correction time |
| Tool sprawl | Often | When AI tool # 4 duplicates something tool # 2 already did |
| Opportunity cost | Sometimes | What the AI lead could have been doing instead |
The two most-missed cost categories are internal admin time and human review time. Both are real costs that should be in the denominator. Excluding them produces fake ROI numbers that fall apart on second review.
Hours saved per task × tasks per period × load-cost-per-hour. Only count hours that visibly translate into either more output or actual time off. Saving the CEO 2 hrs/week that just gets absorbed back into more email does not count as ROI.
If shorter quote turnaround = higher win rate, the additional deals are real ROI. Track it as: incremental deals attributable to faster cycle × average deal contribution margin.
More content shipped, more proposals sent, more tickets resolved per FTE. Count the additional output only if it produced incremental revenue or eliminated a hiring need.
If the AI rollout means you do not hire the next coordinator/analyst/SDR, that is real value. Use loaded cost of the avoided hire, time-weighted to when the hire would have happened.
Real, but rare in measurable form at <$50M scale. Skip unless you have hard data.
20-person B2B services firm. 6-month period. Three workflows rolled out: proposal drafting, content production, and meeting summarization.
Net benefit: \$84,960 / Costs: \$49,800 = ~170% return over 6 months. Annualized (cautiously, since the first 90 days are atypical) closer to 250-300%.