How to Measure the ROI of AI at Work (Formula + Worked Example)

TL;DR: AI ROI is the same as any other ROI, (value created − total cost) ÷ total cost, but both sides get faked constantly. Value must be measured against a pre-AI baseline, and cost must include training and human review time, not just licenses. Track leading indicators (adoption, acceptance rate) weekly to predict success, and lagging indicators (hours saved, cycle time, quality) monthly to prove it. Value saved time at loaded hourly cost with a redeployment discount, saved minutes only count if they became other output. A well-run assistant pilot typically shows measurable savings in 30-60 days and breaks even inside a quarter.

“What’s the ROI?” is the question that decides whether your AI initiative gets budget next quarter. It’s also the question most AI rollouts can’t answer, because nobody set up measurement before switching the tools on.

This playbook gives you a measurement setup you can run in a spreadsheet: what to measure, the formula, a worked example with real numbers, and the mistakes that make ROI claims collapse in front of a CFO. It pairs with our AI Adoption Roadmap, measurement is Phase 3 and 4 of that process.

Start with the baseline (or nothing else works)

Every credible ROI number is a difference: after minus before. If you don’t measure “before,” you’re estimating both sides of the subtraction and the number means nothing.

Baseline for 1-2 weeks before the AI tool arrives. For each target use-case, capture:

Self-reported time logs feel unscientific. They’re fine, you’re comparing the same people’s self-reports before and after, so the bias mostly cancels. The unforgivable error isn’t imprecision; it’s having no “before” at all.

The formula

$$ \text{AI ROI} = \frac{\text{Value created} - \text{Total cost}}{\text{Total cost}} \times 100% $$

Per period (monthly or quarterly). Both terms need honest accounting:

Value created

1. Time savings, the workhorse: hours saved vs. baseline × loaded hourly cost × redeployment rate
   • Loaded hourly cost ≈ (salary × 1.25-1.4 for benefits/overhead) ÷ 2,080 hours. A $70k employee ≈ $42-47/hour loaded.
   • Redeployment rate, the honesty multiplier. Saved time only becomes value if it turns into more throughput, other work, or reduced overtime/contractor spend. If people just absorb the slack, value ≈ 0. Use 50-70% unless you can show where the hours went; use 100% only when saved hours directly reduce paid hours (contractors, overtime) or increase billable/produced output.
2. Quality gains, fewer errors × cost per error; fewer revision rounds × time per round. Only claim these if the baseline captured an error rate.
3. Throughput/revenue gains, more proposals sent, faster response times that lift close or retention rates. Powerful but claim conservatively: tie to a measured rate change, not a story.
4. Avoided costs, a hire postponed, a contractor reduced, a tool cancelled. Count only when the avoidance actually happened.

Total cost

1. Licenses/usage, seats (typically $20-40/user/month for ChatGPT, Claude, Copilot, or Gemini business tiers) or API usage
2. Implementation, setup, integration, any engineering (amortize builds over 12-24 months)
3. Training time, participants’ hours in training and check-ins, at loaded cost
4. Review time, the one everyone forgets: human minutes spent checking and correcting AI output, every single use. If a task took 30 minutes and now takes 5 minutes of generation plus 10 of review, the saving is 15 minutes, not 25.
5. Ownership overhead, the fraction of someone’s time running the program

Leading vs. lagging metrics

Lagging metrics prove ROI; leading metrics predict it early enough to fix problems. Track both:

The early-warning patterns to know:

• Adoption decaying by week 4-5 predicts an ROI of roughly zero regardless of how good week-1 numbers looked. Novelty is not value.
• Acceptance rate below ~60% means people spend so much time fixing outputs that net savings will disappear. That’s a prompt/training/use-case-fit problem, fix it or kill the use-case before scaling.
• High adoption + high acceptance almost always converts into lagging-metric gains a month later. If you see this, start preparing the scale case.

Worked example: support-ticket drafting, 8-person team

A 40-person company pilots AI-assisted first drafts for customer-support replies. Eight agents, 60-day pilot, using an off-the-shelf assistant.

Baseline (measured over 2 weeks before the pilot):

• 8 agents × 460 tickets/week ≈ 1,840 tickets/month requiring a written reply
• Average 11 minutes per reply → 337 hours/month
• Loaded cost $38/hour → current cost of the task ≈ $12,800/month

Pilot results (final 4 weeks, when usage had stabilized):

• Agents used AI drafts on 70% of tickets (1,288/month); the rest were too unusual
• On AI-drafted tickets: 3 min generation + 4 min review/edit = 7 min vs. 11 baseline → 4 min saved each
• Acceptance rate 78%; customer-satisfaction scores unchanged (quality held)

Value per month:

• Hours saved: 1,288 tickets × 4 min ÷ 60 = 86 hours
• Gross value: 86 × $38 = $3,268
• Redeployment: the team measurably absorbed a ~15% ticket-volume increase without the planned part-time hire → use 70% → $2,290/month (conservative; the avoided hire alone is arguably worth more)

Cost per month:

• Licenses: 8 × $30 = $240
• Training: 8 people × 3 hours in month one, amortized over 6 months ≈ $150/month
• Program ownership: ~2 hours/week of a team lead ≈ $330/month
• (Review time is already inside the 7-minute figure, don’t double-count it)
• Total: ≈ $720/month

ROI = (2,290 − 720) ÷ 720 ≈ 218% per month, with breakeven in the first month. Note what made this number defensible: a measured baseline, review time counted inside the task time, a redeployment rate justified by an observed fact (absorbed volume growth, avoided hire), and stabilized week 5-8 data rather than launch-week enthusiasm.

Also note what it isn’t: a claim that AI “saved $39k a year.” It’s $27k/year on this one use-case at current volume, smaller, but it survives a CFO’s questions. Three or four use-cases like this, compounding, are how the real money shows up.

The six measurement mistakes that fake (or hide) ROI

1. No baseline. After-only numbers are unfalsifiable, and everyone in the room knows it.
2. Ignoring review time. Counting generation speed while a human spends 10 minutes fixing each output flips real losses into paper gains. Always measure end-to-end task time, review included.
3. Valuing saved minutes at full salary with zero redeployment. “15 people × 20 minutes/day × salary = $180k!”, only if those minutes became output. Apply the redeployment discount and be able to say where the hours went.
4. Cherry-picked users. Your two prompt-wizards aren’t the average. Report the pilot median, not the champion, and include the non-users in the denominator.
5. Extrapolating week two. Usage almost always dips after novelty fades. Use stabilized data, the back half of a 60-day pilot, for any annualized claim.
6. Measuring only cost, never quality. A tool that saves 20% of time while doubling error rates has negative ROI that pure time-tracking will never show. Carry at least one quality metric through the whole exercise.

There’s a seventh, subtler one: measuring so heavily that measurement eats the savings. A shared spreadsheet, weekly 15-minute check-ins, and numbers your systems already produce are enough. If your measurement program needs its own program manager, it’s too big for the pilot stage.

Reporting it: the one-page format

When you present, one page beats a deck:

1. Baseline: what the task cost before (hours, $, quality)
2. Result: what it costs now, with the measurement window stated
3. ROI: the formula with your actual numbers visible, show the redeployment rate and let it be challenged
4. Decision asked: scale / adjust / kill, and what scaling costs and returns
5. Honesty footnote: what you didn’t count and why

Showing your conservative assumptions is a credibility weapon. A 218% ROI with visible discounts beats a 900% ROI nobody believes, and it makes your next budget request pre-trusted.

FAQ

What is the formula for AI ROI?

ROI = (value created − total cost) ÷ total cost, per month or quarter. Value = hours saved versus a pre-AI baseline × loaded hourly cost × redeployment rate, plus defensible quality or revenue gains. Total cost = licenses + implementation + training time + review time + program ownership.

How long before AI shows measurable ROI?

Assistant-based use-cases show measurable time savings within 30-60 days of a baselined pilot and typically break even inside the first quarter, because license costs are low. Custom or integrated AI carries build costs and usually takes 6-12 months to pay back.

What metrics should we track for AI adoption?

Weekly leading indicators: active users on the use-case, outputs generated per user, and acceptance rate (share of outputs used with minor or no edits). Monthly lagging indicators: hours saved vs. baseline, cycle time, error/revision rates, and cost per unit of work.

How do you value saved time in an AI ROI calculation?

At loaded hourly cost, roughly salary × 1.25-1.4 ÷ 2,080, multiplied by a redeployment rate. Saved hours count as value only insofar as they became other work, higher throughput, or reduced paid hours (overtime, contractors, avoided hires). 50-70% is an honest default rate.

Why do most AI ROI calculations fall apart under scrutiny?

The recurring five: no baseline, review time excluded from cost, saved minutes valued at full salary with no redeployment story, cherry-picked power users presented as the average, and launch-week usage extrapolated to a year.

Not sure where your company stands? Take the free AI-Readiness Assessment, it scores your measurement readiness along with strategy, data, people, and governance, and tells you what to fix first.