How to Measure AI ROI: A Formula With Real Project Numbers

Most articles on AI ROI quote surveys, which cannot tell you whether your own project will pay off. For that you need a formula, honest inputs and a way to test the result after launch. The difference between projects that pay back in months and projects that never pay back usually sits in two inputs that teams skip: how many people actually use the system and whether the time it saves turns into money.

The ROI of an AI project equals annual benefit minus annual cost, divided by total cost, where benefit counts only the time savings that reduce spending or add revenue and cost includes model usage, hosting, maintenance and human review.

This article gives the formula, works through it with real figures from a delivered legal AI project, tests its sensitivity and explains how to measure ROI after launch. It does not cover tax treatment of AI spending, and it is not investment advice.

Key takeaways

• First-year ROI = (annual benefit − first-year cost) ÷ first-year cost. Payback in months = build cost ÷ monthly net benefit.
• Annual benefit = units per year × hours saved per unit × loaded hourly cost × adoption rate × realization rate.
• The realization rate is the share of saved time that turns into money, through avoided hiring, extra billable work or lower overtime. Skipping it is the most common way ROI gets overstated.
• In a delivered legal project, model costs were about $3–4 per case and attorney throughput rose about 1.5 times. With reasonable assumptions, that gives a payback of under four months.
• The same project falls to a negative first-year ROI when adoption, time saved and realization are all low, which is why a pilot should measure them before a full build.
• Measure a baseline before building, compare against it after launch and recalculate at 30, 90 and 180 days.

The AI ROI formula

Benefit

Annual benefit = units per year × hours saved per unit × loaded hourly cost × adoption rate × realization rate + other gains

• Units per year. Cases, invoices, tickets, calls or documents processed through the workflow.
• Hours saved per unit. Measured on real work during a pilot, not taken from a vendor demo.
• Loaded hourly cost. Salary plus taxes, benefits and overhead, usually 1.25 to 1.4 times the base hourly pay.
• Adoption rate. The share of units that actually go through the AI system.
• Realization rate. The share of saved hours that become money. If a team saves 20 hours a week and fills them with work that was already done, the realization rate is close to zero. If the saved hours absorb growth that would otherwise need a new hire, it is close to one.
• Other gains. Fewer errors, faster turnaround that wins business, or revenue from work that was not possible before. Count these only when you can put a number on them.

Cost

First-year cost = build cost + model usage + hosting + maintenance + human review + change management

• Build cost. Discovery, proof of concept, development, integration and testing.
• Model usage. Tokens or minutes paid to model providers, or the cost of your own GPUs.
• Hosting. Servers, databases, storage and monitoring.
• Maintenance. Prompt updates, model upgrades, bug fixes and integration changes, typically 15–25% of the build cost a year.
• Human review. Time people spend checking AI outputs. Count it as a cost unless it is already included in the hours saved.
• Change management. Training, documentation and time for the team to adjust.

Results

• First-year ROI = (annual benefit − first-year cost) ÷ first-year cost.
• Payback period in months = build cost ÷ (annual benefit − annual running cost) × 12.
• Second-year ROI uses running costs only, since the build is already paid for.

Worked example: AI for legal case files

Brandon J. Broderick, a US personal injury law firm, prepares legal opinions for insurance compensation from case bundles of about 250 pages of medical records, police reports and notes. .wrk built a tool in which five specialized AI assistants read the bundle and produce a structured draft opinion for attorney review. Two figures come from the delivered project: model costs of about $3–4 per case, and attorney throughput about 1.5 times higher. Read the BJB legal document processing case study.

The other inputs below are assumptions for a firm of similar size. Replace them with your own numbers.

Annual benefit: 1,200 cases × 2 hours × $120 × 80% adoption × 100% realization = $230,400.

First-year cost: $60,000 build + $4,800 model usage (1,200 × $4) + $3,600 hosting + $12,000 maintenance = $80,400.

First-year ROI: ($230,400 − $80,400) ÷ $80,400 = 187%.

Payback: annual running cost is $20,400, so the monthly net benefit is ($230,400 − $20,400) ÷ 12 = $17,500, and the $60,000 build pays back in about 3.4 months.

Second-year ROI: with only $20,400 of running costs, the benefit is about 11 times the cost.

Model usage is the smallest line in the calculation. At $4 per case, it is under 2% of the value of the time saved per case. This holds for most document-heavy AI work: the build, the integration and the people decide the economics, and model prices matter much less than they appear to. Our breakdown of what it costs to run an LLM on a 250-page case file compares model costs per case in detail.

How sensitive the result is

The same project produces very different returns depending on three inputs. The table changes them one at a time, then all together, with the build and running costs unchanged.

Two conclusions follow. First, the project is still profitable when any single input disappoints, which makes it a reasonable investment. Second, when all three disappoint together, it loses money for years. A pilot that measures adoption, hours saved and realization on real work removes most of that risk before the full build. Our guide to running an AI proof of concept explains how to set those measurements up.

Using public figures as inputs

Published results from large companies are useful as sanity checks for your own inputs, as long as you remember they come from companies with large volumes and large teams.

• Customer service. Klarna reported in February 2024 that its AI assistant handled 2.3 million conversations in its first month, two-thirds of all customer service chats, doing the work of 700 full-time agents. Resolution time fell to under 2 minutes, down from 11, and the company estimated a $40 million profit improvement for 2024. At 2.3 million conversations a month, that works out to roughly $1.45 of benefit per conversation. Klarna later said customers should always be able to reach a human agent, which shows that quality and customer trust belong in the calculation too.
• Internal support. IBM's AskHR case study reports 94% containment for common employee questions and a 40% reduction in the HR team's operating costs over four years. The timeframe is the useful input: large returns came from years of expanding the system, not from the first release.
• Across companies. In McKinsey's 2026 State of AI survey, 37% of respondents attributed at least some EBIT impact to AI, and about 6% were high performers that attribute at least 5% of EBIT to it. Most companies using AI have not yet seen it in their profit figures, which is a reason to calculate carefully, not a reason to wait.

For more company examples and the results behind them, see our list of AI agent use cases for B2B companies.

How to measure AI ROI after launch

1. Record a baseline before building. Measure time per unit, volume, error rate and cost for at least four weeks. Without a baseline, every later number is an estimate.
2. Define the unit and the owner. Agree on what counts as a unit, such as a case, ticket or invoice, and who reports the numbers.
3. Log usage in the system. Record which units went through the AI and how long review took, so adoption and hours saved come from data, not surveys.
4. Track realization separately. Note where saved time went: extra volume handled, hiring avoided, overtime reduced or new work started.
5. Compare against a control where possible. If some teams or regions use the system and others do not yet, compare them over the same period.
6. Recalculate at 30, 90 and 180 days. Early numbers include learning time. Adoption and hours saved usually improve over the first three months, and the 90-day figures are a fair basis for decisions.
7. Include quality. Track how often reviewers change AI outputs and whether error rates or customer complaints move. A system that saves time but increases rework does not pay off.

Costs that ROI calculations often miss

• Review time. Every output a person checks costs time. Include it in the hours saved, or count it as a cost.
• Integration upkeep. Changes in connected systems, such as a CRM update or a new document format, need development work.
• Model changes. Providers retire model versions, and each switch needs testing against a fixed set of examples.
• Exception handling. Cases the AI cannot handle still need a person, and they are often the slowest cases.
• Governance. Data protection reviews, access controls and audit logging take time to set up and maintain.

Our article on AI development cost breaks down the build side of the calculation in more detail.