How to Calculate AI ROI: A Framework With Formula and Examples

Most AI projects are approved on a vibe and abandoned on a vibe. Neither is a number. If you want AI to survive contact with a budget review, you need to calculate its return the same way you'd justify any other investment — and do it before you build, not after.

Here is the short version: AI ROI = (annual value created − annual cost of the system) ÷ total investment. The rest of this piece is about what actually belongs on each side of that equation, because that's where most business cases quietly fall apart.

The formula, stated plainly

ROI = (Annual benefit − Annual running cost) ÷ (Build + implementation cost) × 100%

Three inputs, and every one of them is routinely fudged:

• Annual benefit — the money the system makes or saves in a year, measured against a real baseline.
• Annual running cost — inference, infrastructure, licenses, monitoring, and the human time to keep it working.
• Build cost — the one-time cost to design, build, integrate, and roll it out.

Counting the benefit honestly

Benefit comes in three flavors, in rough order of how defensible they are:

1. Cost taken out. Hours saved × loaded hourly rate, or headcount you genuinely avoid adding. The cleanest to measure — but only count hours that convert into real capacity or real reductions, not "time saved" that just becomes slack.
2. Revenue moved. More conversions, less churn, faster cycle times that pull revenue forward. Attribute conservatively and, wherever possible, prove it with a holdout group.
3. Risk avoided. Fewer errors, better compliance, fewer incidents. Real, but soft — discount it heavily in the headline number.

The single most important discipline: establish a baseline before you start. "The agent resolves 42% of tickets" means nothing until you know what the resolution rate was without it. No baseline, no ROI — just a story.

Don't undercount the cost

The build is the cheap part. The line items teams forget:

• Inference at scale. A price per call that looks trivial in a demo becomes a monthly bill at production volume. Model it at real usage.
• The last mile. Evaluation, guardrails, monitoring, and human review are not optional extras — they are the difference between a demo and a system, and they carry ongoing cost.
• Maintenance and drift. Models and data change. Budget for the engineering time to keep quality from decaying.

A worked example

A 40-person support team wants an AI agent to deflect tier-1 tickets.

• Baseline: 20,000 tickets/month, average handle time 12 minutes, loaded cost $35/hour.
• Measured effect (from a scoped pilot): the agent fully resolves 30% of tickets and cuts handle time 20% on the rest.
• Annual benefit: deflected tickets save ~24,000 agent-hours/year; faster handling on the remainder saves ~11,000 more. At $35/hour that's roughly $1.2M/year — before discounting.
• Costs: ~$180K to build and integrate, ~$120K/year to run and maintain.

Apply a conservative 30% haircut to the benefit for over-attribution, and first-year ROI is still (840K − 120K) ÷ 180K ≈ 400%. Even halve the effect and it clears the bar. That's a project worth doing — and, just as importantly, you'll know whether it worked, because you set the baseline first.

Why most AI business cases still fail

Not because the math is hard, but because:

• There's no baseline, so "improvement" is unfalsifiable.
• The benefit is top-line and unattributed ("AI will boost sales") instead of tied to a measurable mechanism.
• The running cost is ignored, so a positive year-one ROI turns negative once inference scales.
• The pilot proves the model works but never proves it delivers, because success was defined as "it demos well."

The fix is to treat ROI as a design constraint, not a retrospective. Pick the use case with the clearest payoff, agree on the metric it should move, scope the first slice small enough to measure quickly, and instrument it from day one.

That's exactly how we run AI & data strategy engagements — we won't build something we can't tie to a number, and we'll tell you before you spend if the number isn't there.