AI ROI is one comparison: the loaded cost of the hours a workflow consumes today versus the cost of licenses, implementation, and upkeep to automate it. If the first number isn't comfortably bigger than the second, don't buy yet. That's the entire discipline — the rest of this post is how to run it honestly.
Your CFO is right to demand this math. Most AI purchases skip it, which is why so many end up as unused seats on a renewal invoice. Ten minutes with a spreadsheet before the contract beats ten months of explaining afterward.
What's the basic AI ROI formula?
The formula is: (hours saved per week × weeks per year × loaded hourly cost) minus (annual license cost + implementation + maintenance). Positive and meaningful, you have a case. Marginal or negative, you have a hobby.
Three terms need honest inputs:
- Hours saved — not hours the task takes, but hours AI realistically removes. Drafting gets faster; deciding usually doesn't. Discount accordingly.
- Loaded cost — salary plus benefits, taxes, and overhead. A common rule of thumb is salary plus 25–40 percent. Using raw salary quietly understates your savings.
- Total cost — licenses are the visible line. Implementation, integration, and someone maintaining the thing are the invisible ones. Include them or the CFO will, later, less kindly.
What does the math look like in practice?
Here's a fully hypothetical example with deliberately round numbers — illustration, not a client result.
Say a team of five each spends six hours a week drafting emails, summarizing calls, and updating records — 30 team-hours weekly. At a hypothetical loaded cost of $50 an hour, that's $1,500 a week, or $75,000 across a 50-week year.
Now assume AI handles half of that work — a conservative figure for drafting-heavy tasks with human review. That's $37,500 a year in recovered capacity. If licenses, setup, and training land at, say, $15,000 in year one, the return is roughly 2.5x with payback inside five months. If the same tooling costs $60,000 to stand up, the math says wait — or pick a bigger workflow.
Run this with your own numbers — real headcount, real hours, real loaded costs. The point isn't ours; it's that the arithmetic takes ten minutes and settles most arguments. One refinement worth making: recovered hours only count if they get redirected to something valuable. Capacity that evaporates into longer coffee breaks is a rounding error, not a return — so name what the freed-up time will actually go toward before you count it as savings.
How do you calculate ROI for support automation?
Support gets its own formula because the unit isn't hours — it's resolutions: tickets per month × AI resolution rate × cost per ticket.
Cost per ticket is your loaded support cost divided by tickets handled. Then apply a resolution rate you'd defend in a meeting — password resets and order-status questions resolve reliably; judgment calls and angry customers don't, and shouldn't. Hypothetically: 2,000 tickets a month, $8 per ticket, and AI cleanly resolving 25 percent is $4,000 a month in avoided cost — $48,000 a year, before the deflected tickets that never got filed because the help center answer improved.
One honesty rule: only count resolutions where the customer didn't come back. A deflected ticket that reopens angrier cost you money, not saved it.
What hidden costs belong in the model?
Four costs sink more AI business cases than license fees ever do:
- Change management. Training time, workflow redesign, and the productivity dip while habits form. Weeks, not days.
- Review time. Humans checking AI output isn't overhead to eliminate — it's the guardrail. Budget it as a permanent line, thinner over time.
- Maintenance. Prompts drift, integrations break, processes change. Someone owns this or it decays.
- Data cleanup. If your CRM is a junk drawer, add the cost of fixing it — AI built on bad data automates the production of wrong answers at scale.
Put all four in the model. A business case that survives them is one you can actually defend.
When is the ROI just not there?
Sometimes the honest answer is: not yet. The math tends to fail when volume is low (a task done twice a month rarely justifies automation), when the process is undocumented (you'd be automating folklore), when the data underneath is dirty, or when the work is mostly judgment — which AI assists but doesn't reliably replace.
None of those are permanent. Document the process, clean the data, and rerun the numbers next quarter — the fix is usually cheaper than the failed rollout it prevents. We build these models with buyers before they license Claude, and we've told plenty of them to wait. If you want a second set of eyes on your spreadsheet — including a "not yet" if that's what it says — bring us your numbers.
Frequently Asked Questions
How do you calculate ROI for AI?
Multiply the hours a workflow consumes weekly by weeks per year and the loaded hourly cost of the people doing it, then subtract the full cost of AI: licenses, implementation, training, and ongoing maintenance. A strong case shows payback within months using conservative assumptions — not best-case ones.
What is a good ROI for an AI project?
A defensible first AI project typically models at least a 2x return in year one with payback inside six months, using conservative assumptions. If the case only works with aggressive estimates for hours saved or resolution rates, treat it as a signal to pick a higher-volume workflow instead.
What hidden costs do AI business cases miss?
The four most-missed costs are change management (training and the productivity dip while habits form), permanent human review time, ongoing maintenance of prompts and integrations, and data cleanup — dirty CRM or support data has to be fixed before automation multiplies its errors.