How to Calculate AI ROI: A Framework You Can Actually Use

The core formula looks like the one you would apply to any investment: ROI = (value gained minus total cost) divided by total cost, shown as a percentage. AI is not hard because of the arithmetic. It is hard because the "value gained" portion sits in four separate buckets (time saved, revenue added, risk avoided, and speed gained). Most firms track only the first bucket, then ask why leadership is not convinced.

This guide gives you the formula for all four pillars, a worked example for each one so you can see the arithmetic in action, and a straight answer about why so many companies still struggle to show that their AI spending paid off.

Why this is genuinely hard, not just neglected

A July 2025 study from MIT Media Lab's Project NANDA examined this question directly. The researchers reviewed more than 300 publicly disclosed AI initiatives and interviewed representatives from 52 organizations. Its headline finding is that 95% of organizations are getting zero return from generative AI (The GenAI Divide: State of AI in Business 2025). One caution about this source: the full report is hosted by a third party rather than on an MIT domain. This does not mean 95% of projects failed outright. It means that in 95% of cases, no one could point to a figure and say "this is what it did."

In a separate effort, Gartner polled 822 business leaders in finance, HR, marketing, sales, customer service, supply chain, procurement, and legal/risk/compliance between September and November 2023. Those who had already implemented generative AI reported an average 15.8% revenue increase, 15.2% cost savings, and 22.6% productivity improvement (Gartner, July 2024). Gartner itself adds the caveat: "benefits are very company, use case, role and workforce specific." These are self-reported averages across projects that differ a great deal from one another. They are not a figure you can slot into your own business case. Use them as a rough idea of scale, not as a target.

The lesson from both is the same. Build your own baseline and your own way of measuring before you build the AI project, not afterwards.

The four places ROI actually lives

1. Efficiency: time and cost saved on existing work

What to measure: labor hours saved, error and rework rates, operational cost per unit of output.

The formula: (hours saved per week x 52 x fully loaded hourly cost) minus (AI tool cost plus implementation cost)

Illustrative example (hypothetical numbers, not a real deployment): suppose a support team of 20 agents currently spends 15 hours a week per person on ticket categorization, and an automation tool trims that to 4 hours a week per person.

• Hours saved: 11 x 20 = 220 hours a week
• Annual hours saved: 220 x 52 = 11,440 hours
• At a $50/hour fully loaded cost: 11,440 x $50 = $572,000 in labor time freed up
• Tool cost: $500/month = $6,000 a year, plus a $15,000 one-time implementation
• Net first-year value at these inputs: $572,000 minus $21,000 = $551,000

Alter any input (the hours saved, the loaded rate, the tool cost) and the result shifts. Test it with your own numbers before you rely on it. If you have not run a workflow automation project before, that guide is a sensible place to look for where the fastest efficiency gains usually show up. Manual data entry is another common target for this pillar; see our breakdown of the cost of manual data entry for how to size that baseline.

2. Revenue: what AI adds rather than saves

What to measure: conversion rate changes from personalization, income from a new AI-powered feature, movement in average order value from a recommendation engine.

The formula: (new or incremental revenue) minus (AI solution cost plus program cost)

Illustrative example: picture an online store that switches on AI product recommendations. Average order value rises from $85 to $92 across 5,000 monthly orders.

• Additional revenue per month: $7 x 5,000 = $35,000
• Annualized: $420,000
• Platform cost: $50,000 a year
• Net value at these inputs: $370,000

This is the pillar a sales pitch most often pads. "AI increased conversion" is easy to claim and nearly impossible to separate from everything else that moved at the same time (a price change, a seasonal spike, a competitor stumbling). Run an A/B test against a control group to isolate the effect before you credit the AI feature with the gain.

3. Risk avoided: what did not go wrong

What to measure: compliance violations and fines avoided, fraud losses prevented, customer churn reduced, security incidents avoided.

The formula: (estimated cost of the risk, before) minus (estimated cost of the risk, after) minus (AI cost)

Illustrative example: take a company with $2M in annual fraud losses. It deploys AI-assisted fraud detection and losses fall by an estimated 40%.

• Fraud prevented: $800,000
• Detection system cost: $120,000 a year
• Net value at these inputs: $680,000

This pillar is the hardest to defend in front of a skeptical board, because you are claiming credit for something that never took place. Document the baseline (the fraud rate, the violation count) before the AI system goes live. Otherwise the "prevented" figure is only an assertion. For a closer look at this pillar in a regulated environment, see secure AI deployment for regulated industries.

4. Agility: value from moving faster

What to measure: faster product development cycles, faster response to a market shift, decision-making speed.

The formula: (value of the time advantage) minus (AI investment cost)

Illustrative example: a software company ships a feature 2 months ahead of where its previous release cadence would have put it, and that earlier entry helps close a $200,000 contract that a slower competitor would otherwise have won.

• Value attributed to the earlier ship date: $200,000 (a judgment call, not a hard number, since "would we have lost this deal anyway" is unknowable)
• Development tooling cost: $75,000
• Net value at these inputs: $125,000

This is the softest of the four pillars, and the reason is the counterfactual. What would have happened without the faster cycle is a guess. Treat any agility figure as a range you have confidence in, not a number to present as if it were precise.

Running all four together

Do not stack illustrative examples like the ones above and present a single combined ROI figure to a board. Each pillar's inputs arrive from a different part of the business, and each carries its own amount of uncertainty. Folding them into one number hides which pieces are solid and which are a guess. Instead, show each pillar's actual measured number separately, give each its own confidence level, and let efficiency carry the most weight in the total. It has the least uncertainty, since hours and hourly cost are both directly observable.

Your ROI Worksheet: Fill In Your Own Numbers

There is no universal figure here on purpose. Both the MIT and Gartner data above are averages

across projects that differ a great deal from one another, not a number you can slot into your own

business case. Use this worksheet to run your own numbers through each pillar's formula before you

build the business case.

Before you fill in a single cell: record your baseline for that pillar first. If you cannot state

the "before" number, you will not be able to prove the "after" number either.

Do not add the four rows together into one combined figure. Each pillar's inputs come from a

different part of the business and carries its own amount of uncertainty. Report each pillar's

measured number separately, with its own confidence level, and give efficiency the most weight in

the total since hours and hourly cost are both directly observable.

A 30-day measurement plan

Week 1: baseline. Record current performance on the pillar you are targeting before any AI system touches it. If you cannot state the "before" number, you will not be able to prove the "after" number either.

Week 2: tracking. Define the exact KPI for that pillar and put a tool in place to measure it continuously, not just once at the end.

Week 3: projection. Use your pillar's formula to project a range, not a single number. Write your assumptions next to the projection so anyone reviewing it can see what would break it.

Week 4: the business case. Present the projection with its assumptions and the baseline it is measured against. A number with no baseline and no stated assumptions is not a business case. It is a guess with a dollar sign in front of it.

Tools for each pillar

• Efficiency: time-tracking apps (Toggl, Harvest), process mining software (Celonis, UiPath)
• Revenue: analytics platforms (Google Analytics, Mixpanel), your CRM (Salesforce, HubSpot), A/B testing tools (Optimizely, VWO) to isolate the AI-specific lift
• Risk avoided: the reporting built into your existing compliance and fraud detection systems, customer success tools (Gainsight, ChurnZero) for churn tracking
• Agility: project management tools (Jira, Asana) for cycle-time tracking

The bottom line

The formula is simple. What is hard, and what both MIT and Gartner's own numbers point to, is that the answer differs for every company and every project, and it only exists if you measured a baseline before you started. Begin with one project, one pillar, one measured baseline. Prove that number before you build the next one. For the operational mistakes that keep a project from ever reaching the measurement stage, see why AI projects fail, and how to fix it.

Keep reading

For the complete strategic picture, read the CEO's guide to AI transformation. You might also find value in choosing between Zapier, Make, and custom code. Related: calculating the ROI of deploying an AI SDR agent.

Ready to take the next step? Book a free strategy call or explore our services.

y call](/contact/) or explore our services.

Sources

• Gartner press release, 29 July 2024 (primary source, checked 2026-09-14)
• MIT NANDA, The GenAI Divide: State of AI in Business 2025 (primary source, checked 2026-09-14)