How to Calculate AI ROI (With the Actual Formulas)
AI ROI is (annual value created − annual cost) ÷ annual cost. Value comes from three measurable sources: recovered time (hours saved × loaded hourly cost), cost avoidance (retired tools and prevented spend), and revenue lift (faster cycles and better decisions). The denominator is a fully-loaded cost — licenses, build, run, governance, and change management — which is the number most business cases lowball.
Key takeaways
- AI ROI = (annual value − annual cost) ÷ annual cost. Everything else is estimating those two numbers honestly.
- Value has three sources: recovered time, cost avoidance, and revenue lift — each with its own formula.
- Recovered time is usually the biggest and most defensible: hours saved × fully-loaded hourly cost.
- The cost denominator is where cases break — include build, run/inference, governance, and change management, not just licenses.
- The credibility killers: double-counting, ignoring adoption rates, mixing one-time and recurring, and having no baseline.
Most AI business cases are vibes with a spreadsheet stapled on. A big number, a confident slide, and no way to check the math when the returns don’t show up. That’s a shame, because AI ROI is genuinely calculable — the formulas aren’t complicated, and using the real ones is the difference between a case that survives a CFO and one that evaporates on contact.
Here are the actual formulas, the cost people lowball, and the mistakes that make good numbers fall apart.
How do you calculate AI ROI?
The base formula is the same one you’d use for any investment:
AI ROI = (annual value created − annual cost) ÷ annual cost
The arithmetic is trivial. The work is estimating the two inputs honestly — and most cases go wrong by inflating the value and understating the cost. Let’s do both properly.
The three sources of value
AI value comes from three measurable places. Estimate each separately so you don’t double-count, and apply a realistic adoption rate to all of them, because value only comes from people who actually use the thing.
- Recovered time. Hours saved per person per week × people × working weeks × fully-loaded hourly cost. This is the recovery of the Context Tax, and it’s usually the largest and most defensible component.
- Cost avoidance. Retired overlapping tools, reduced vendor spend, and hires you didn’t have to make because the work got absorbed. Concrete, and easy to verify after the fact.
- Revenue lift. Faster cycle times, better decisions, higher conversion. The hardest to attribute cleanly, so weight it conservatively unless you can tie it to a measured change.
Getting the cost right
The denominator is where business cases quietly break. Teams put the license fee in and call it the cost, then wonder why the realized ROI is a fraction of the projection. A honest annual cost includes the license plus the build or implementation, the run cost (inference, compute, storage), the governance and security overhead, integration and maintenance, and the change management it takes to actually drive adoption. Governance in particular is easy to forget and expensive to bolt on late — which is an argument for platforms where it’s already built in.
A worked example
Take a 500-person company recovering 3 of the 5 hours a week each person loses to knowledge friction, at a $110,000 loaded cost, with realistic adoption. Here’s the shape of the case.
| Recovered hours/year (500 × 3 × 46) | 69,000 |
|---|---|
| Loaded hourly cost ($110,000 ÷ 2,080) | $52.88 |
| Gross recovered-time value | ≈ $3.65M |
| × realistic adoption (say 80%) | ≈ $2.92M |
| Fully-loaded annual cost (all-in) | $750,000 |
| ROI = (2.92M − 0.75M) ÷ 0.75M | ≈ 289% |
Notice what the conservative choices did: applying an 80% adoption rate and a fully-loaded cost cut the headline number roughly in half — and it’s still a strong return. That’s the point. A defensible 289% beats an indefensible 600% the moment someone checks your work.
The mistakes that break AI ROI cases
Four errors turn a credible case into a fantasy. Double-counting the same benefit under recovered time and cost avoidance. Assuming full adoption when real usage is 60–80% — value scales with the people who actually use it, and a tool that never gets adopted is the most common way AI never reaches production anyway. Mixing one-time and recurring gains as if a single win repeats forever. And having no baseline, so there’s nothing to measure the improvement against. Fix those four and your number gets smaller and far more durable — which is exactly what the ROI Calculator is built to enforce.
Frequently asked questions
How do you calculate AI ROI?
What is the formula for recovered-time value from AI?
What costs should be included in an AI ROI calculation?
Why do AI ROI calculations often overstate returns?
Run the numbers on your own organization.
The ROI Calculator applies these formulas to your inputs — recovered time, recoverable share, and cost — so your CFO can argue with the math, not the vibes.