"What's the ROI?" is the question that kills — or funds — most enterprise AI projects. Too many teams answer it with a vibe ("it'll save time!") and lose the budget to someone who brought numbers. ROI on AI isn't mystical; it's value minus cost, honestly counted. Here's a framework you can apply to any AI initiative, and present to a CFO without flinching.
The formula (don't overcomplicate it)
ROI (%) = (Annual value − Annual cost) / Annual cost × 100
Payback (months) = One-time build cost / Monthly net savingsEverything else is just filling in those numbers honestly. The discipline is in not fooling yourself on either side.
Want the math done for you? Plug your numbers into the free AI ROI Calculator — team size, hours, cost, expected savings — and it returns annual net benefit, ROI %, and payback period.
Step 1 — Quantify the value (be conservative)
AI value usually comes from one of three places. Pick the one that's real for your use case and measure it:
- Time saved — people-hours × loaded hourly cost. (Loaded cost = salary + overhead, not just salary.) This is the most common and easiest to defend.
- Revenue lift — more conversions, faster sales cycles, upsell. Harder to attribute; use a conservative estimate and label it as such.
- Quality/risk — fewer errors, faster resolution, avoided incidents. Quantify the cost of the thing you're preventing.
Apply an adoption discount. A tool that saves 5 hours/week only delivers that if people actually use it. Multiply by a realistic adoption rate (50–80%), not 100%.
Step 2 — Count ALL the costs (this is where estimates lie)
The build cost is the obvious one; the ongoing ones are what people forget:
- Build — engineering time, design, integration (one-time).
- Running — LLM/API tokens (see token budgeting), infra, vector DB, monitoring (monthly).
- Maintenance — prompt updates, evals, model migrations, support (ongoing — budget 15–25% of build annually).
- Change management — training, rollout, the productivity dip while people learn.
Under-counting run + maintenance is the #1 reason "positive ROI" projects quietly go negative.
Step 3 — Compute payback and ROI
With honest value and cost:
- Payback period = build cost ÷ (monthly value − monthly running cost). Under 12 months is usually an easy yes; over 24 needs a strong strategic reason.
- Year-1 ROI = (annual value − annual running cost − build cost) ÷ total cost × 100.
- Run a 3-year view too — build cost amortizes, so year 2–3 ROI is usually much stronger. Show both.
Step 4 — Present it to executives
- Lead with the number, then the assumptions. "~18-month payback, ~140% three-year ROI, assuming 70% adoption."
- Show your conservative assumptions explicitly — it builds trust and pre-empts "did you inflate this?"
- Give a range, not false precision: base / conservative / optimistic.
- Tie it to a metric they own — hours reclaimed, tickets deflected, revenue influenced.
Common ways ROI estimates lie
- Assuming 100% adoption.
- Counting gross time saved but ignoring that saved time isn't always recaptured as value.
- Forgetting run + maintenance costs.
- Ignoring the productivity dip during rollout.
- Using list salary instead of loaded cost (understates value) — or wildly optimistic revenue lift (overstates it).
FAQ
What's a "good" AI ROI? Context-dependent, but a payback under ~12 months or a clearly positive 3-year ROI is a strong case. Strategic bets (capability, learning) can justify longer.
How do I estimate value before I've built anything? Run a small pilot or time-and-motion study on the target task, measure the baseline, then project conservatively with an adoption discount.
Should I include "soft" benefits like morale? Mention them, but don't rest the case on them. Lead with hard, defensible numbers; list soft benefits as upside.
How do I keep running costs from eroding ROI? Budget and meter tokens (token budgeting), and control spend at an AI gateway.
AI ROI is just honest value minus honest cost. Quantify value conservatively (with an adoption discount), count all the costs (especially run + maintenance), compute payback and a 3-year view, and present a range with your assumptions on the table. That's the difference between getting funded and getting cut.
Run your numbers in the AI ROI Calculator. More: token budgeting and scaling AI from POC to production.
Building the business case for an AI initiative and want it pressure-tested? Let's talk.