"AI will save you hours a week" is a claim, not a calculation. Before you can decide whether an AI implementation was worth it — or whether a proposed one is worth funding — you need a number built from your own costs and your own outcomes, not a vendor's case study from a completely different company.
This article walks through the actual formula, what to include on both sides of it, a worked example with real numbers, and the mistakes that quietly inflate almost every AI ROI calculation you'll see floated in a meeting.
The Formula, and Why It's Not the Whole Story
That formula is correct — and almost useless on its own, because the entire calculation lives or dies on what you decide counts as "benefit" and "cost." Two people can run the exact same formula on the exact same AI rollout and get wildly different answers, simply because one of them forgot to count the training time, or the other quietly excluded the token costs from unpredictable usage spikes.
The real skill in AI ROI isn't the division — it's building an honest, complete list of both sides before you ever touch the formula.
Total Cost: The Side People Undercount
AI cost is rarely just "the subscription." A complete cost picture usually includes five categories:
| Cost category | What it includes |
|---|---|
| Direct tool cost | Subscription fees, API/token usage, per-seat licensing |
| Implementation | Integration work, connecting data sources, custom prompts or workflows built |
| Training & adoption | Time spent learning the tool, onboarding materials, internal champions' time |
| Oversight & correction | Time spent reviewing AI output, fixing errors, maintaining quality checks |
| Ongoing maintenance | Updating prompts as needs change, monitoring usage, renewing or renegotiating contracts |
The "oversight and correction" row is the one almost everyone forgets — the time your team spends checking and fixing AI output rarely gets logged anywhere, which makes it invisible in a spreadsheet even though it's very real on a calendar.
Total Benefit: What Actually Counts
Benefits generally fall into three buckets, and they're not equally easy to measure:
- Time saved — the most common benefit claimed, and the easiest to measure: hours previously spent on a task, multiplied by an hourly cost figure for the people who used to do it.
- Quality or error reduction — harder to measure directly, but often larger in value: fewer mistakes reaching customers, fewer redone reports, fewer compliance issues.
- Revenue or capacity gained — the hardest to attribute cleanly, but sometimes the biggest number: more proposals sent out, more leads followed up on, more capacity to take on work without hiring.
Only count a benefit if you can point to where the number came from — a timesheet, a before/after measurement, a documented error rate. "It feels faster" is a hypothesis worth testing, not a benefit worth counting yet.
A Worked Example
Take a small team that implements an AI tool to draft first-pass responses to routine customer support tickets, with a human still reviewing and sending each one.
| Cost item | Monthly cost |
|---|---|
| Tool subscription (5 seats) | $250 |
| Implementation, amortized over 12 months | $150 |
| Training time (one-time, amortized over 12 months) | $60 |
| Review time — 10 min/day per agent to check drafts | $420 |
| Total monthly cost | $880 |
| Benefit item | Monthly value |
|---|---|
| Time saved drafting replies — 40 min/day per agent, 5 agents | $1,750 |
| Fewer late replies (measured drop in response-time complaints) | $300 |
| Total monthly benefit | $2,050 |
That's a strong result — but notice what made it credible: the review time was counted as a real cost, not ignored, and the time-saved figure came from an actual measured before/after difference in how long drafting used to take, not a guess.
Don't Stop at the Percentage — Find the Break-Even Point
A single ROI percentage hides an important detail: how long it took to become positive. A workflow with upfront implementation costs might show a low or negative return in month one, and a strong one from month three onward. The break-even point — when cumulative benefit finally overtakes cumulative cost — is often more useful for decision-making than the steady-state monthly ROI.
Mistakes That Inflate (or Deflate) the Number
- Using industry-average time savings instead of your own measurement. A vendor's case study reflects their customer's process, not yours.
- Forgetting the adoption curve. The first month of any new workflow is usually slower than the tool's eventual steady state — measuring ROI too early makes a genuinely good tool look weak.
- Counting "time saved" that isn't reallocated to anything. If ten minutes saved per day just becomes ten minutes of idle time rather than more output, the financial benefit is much smaller than the raw time figure suggests.
- Ignoring token or usage-based cost variability. Flat subscription estimates can understate real costs once usage scales up during busy periods.
- Leaving out the cost of errors that did happen. If the AI produced a handful of mistakes that needed damage control, that cost belongs in the calculation too — not just the successes.
A Simple Calculation Checklist
- Listed all five cost categories, not just the subscription price
- Measured time savings against an actual before/after baseline, not an estimate
- Included review and correction time as a real cost
- Calculated both a steady-state ROI and a break-even timeline
- Re-measured after 60–90 days once the adoption curve has flattened
An honest ROI number is worth more than an impressive one. The honest one is the only kind you can actually make a decision with.
Frequently Asked Questions
There's no universal threshold — it depends on your industry, the size of the investment, and what else that budget could have funded. What matters more than hitting a specific percentage is that the number was built from real, measured inputs rather than assumptions, so you can trust it when comparing this investment against alternatives.
Wait at least 60 to 90 days before treating any ROI figure as representative. The first few weeks typically reflect the adoption curve — people are still learning the tool and adjusting the workflow — rather than its steady-state performance.
Look for a measurable proxy: the cost of the errors the tool is reducing (rework hours, refunds issued, compliance penalties avoided) before and after implementation. If no clean proxy exists, it's more honest to note the quality improvement as a qualitative benefit alongside the ROI number than to force it into a shaky dollar figure.
No — their salary is a sunk cost you're paying regardless of whether they use the tool. What belongs in the calculation is the cost of the extra time they spend on AI-related tasks (reviewing, correcting, learning it) and the value of what they do with the time it frees up.
Related Reading
This article is part of a series. These go deeper on ideas introduced above: