How to calculate the ROI of AI in business

ROI (return on investment) is the relationship between what a project generates and what it costs. The formula is simple: ROI = (benefits − costs) ÷ costs. The hard part, in artificial intelligence projects, is filling in both sides with honest numbers.

This article shows how to gather the costs, how to measure the benefits, how to design a pilot that produces evidence and what the available surveys say about returns, with the caveats of each. It complements the guide on how to automate business processes with AI, which covers choosing and mapping the process.

Why the math is hard

Three reasons make AI ROI escape the usual calculations.

First, the cost is variable. Much of generative AI is billed by usage, and usage grows as the system succeeds.

Second, the benefit tends to show up as time, not money. A team that answers emails faster only generates a financial gain if the freed-up time is used for something of value or avoids a hire.

Third, there is no comparison. The AI risk management framework from NIST, the US standards institute, points out as a challenge the lack of a baseline for comparison with human performance in systems that augment or replace human activities. The same document notes that metrics can be oversimplified or gamed (NIST AI RMF 1.0, 2023).

The cost side: what goes into the calculation

Separate implementation costs (one-off) from recurring costs (every month).

Implementation

• Process mapping and definition of metrics.
• Integration with existing systems (ERP, CRM, email, databases).
• Preparation of data and reference documents.
• Testing, quality evaluation and pilot.
• Legal and data protection review, including an impact assessment when there is high risk to personal data.
• Team training and changes to procedures.

Recurring

• Per-user tool licenses.
• API consumption, usually billed per token, the unit of text the model processes (see tokens in AI models).
• Infrastructure, storage and monitoring.
• Human review time on what the AI produces.
• Maintenance: adjusting instructions, switching model versions, fixing integrations.
• Governance: audits, documentation updates, incident handling.

Per-token pricing has details that change the math. In Anthropic’s pricing table consulted on October 8, 2026, for example, the price is given per million tokens, output tokens cost more than input tokens, batch processing has a 50% discount and reusing repeated parts of the request (prompt caching) costs a fraction of the normal price. The same page states that tool definitions add tokens to every call and that models with the newest tokenizer produce about 30% more tokens than previous ones for the same text (Anthropic, Pricing). Other providers have their own rules. The general lesson: measure actual consumption in the pilot instead of estimating from text length alone.

Costs can also grow out of control. OWASP, a reference community in software security, lists unbounded consumption among the risks of LLM applications: excessive use, accidental or malicious, that exploits usage-based billing and causes financial losses. Recommended defenses include quotas, request limits and monitoring (OWASP LLM10:2025). In the McKinsey survey published in August 2026, about one in five respondents said AI operating costs, including tokens, limited usage in their organization (McKinsey, The state of AI in 2026).

Finally, there are non-monetary costs. NIST recommends examining and documenting both the expected benefits and the potential costs, including non-monetary ones that result from system errors. It also recommends considering non-AI alternatives. Sometimes a simple rule or an integration solves the same problem for less.

The benefit side: what to measure

AI benefits usually fall into four groups:

• Time: minutes per case, hours per week, time to completion.
• Quality: error rate, rework, complaints, returns.
• Capacity: volume handled with the same team, cleared queues, extended hours.
• Revenue: conversion, average order value, customer retention.

Each metric needs a baseline: the value measured before the change, in the same process, with the same method. Without it, any “gain” figure is a guess.

Watch out for the difference between individual gains and gains for the company. In the same McKinsey survey, 80% of respondents said AI improved their individual productivity. Only 37% attributed any impact on EBIT, earnings before interest and taxes, to AI. A study using administrative data from Denmark found rapid chatbot adoption and workers reporting productivity gains. But the effects measured on wages and recorded hours were precise zeros: the authors rule out effects larger than 2% two years after ChatGPT’s launch (Humlum and Vestergaard, NBER, March 2026 revision). In other words, feeling more productive and generating measurable results are different things.

For productivity metrics in software development teams, see how to measure productivity with AI.

A worked example

A hypothetical example from the editorial team, with numbers invented solely to illustrate the method.

A company receives 4,000 customer emails a month. Today, each takes 6 minutes to read, classify and route: 400 hours a month. With an hourly cost of R$ 60 (salary and payroll charges), the process costs R$ 24,000 a month.

With automation, the AI classifies and routes 70% of the emails, and a person checks each one in 1 minute. The other 30% remain manual, at 6 minutes. Time falls to 2,800 + 7,200 = 10,000 minutes, about 167 hours, or R$ 10,000 a month. Gross savings: R$ 14,000 a month.

Item	Hypothetical value
Implementation (mapping, integration, pilot, training, legal review)	R$ 90,000, one-off
API tokens	R$ 200 a month
Licenses and infrastructure	R$ 1,500 a month
Monitoring and maintenance (20 h × R$ 100)	R$ 2,000 a month
Total recurring cost	R$ 3,700 a month

Over 12 months, benefits total R$ 168,000 and costs R$ 90,000 + R$ 44,400 = R$ 134,400. First-year ROI is (168,000 − 134,400) ÷ 134,400, about 25%. The payback, the time to recover the implementation cost, is R$ 90,000 ÷ R$ 10,300 in monthly net gain, close to 9 months.

Now change a single assumption: instead of 70%, the AI handles 50% of cases. Savings fall to R$ 10,000 a month, the net gain to R$ 6,300, and first-year ROI turns negative, around −11%. Payback exceeds 14 months.

The example shows three things. The actual automation rate matters more than the token price. The math only works if the 233 freed-up hours turn into something of value. And any estimate made before the pilot should be presented as a range, not a single number.

How to design a pilot that produces evidence

A pilot is a limited deployment, designed to test assumptions before investing at scale. For it to answer the ROI question:

1. Define the main metric and the baseline before you start. Choose one or two, not ten.
2. Have a comparison group. Ideally, part of the cases or teams stays on the old process during the same period. That way you separate the AI’s effect from seasonal variation.
3. Measure real costs. Record token consumption, human review time and maintenance hours, not just the license.
4. Count the errors. Measure how many wrong results slipped through and how much each cost to fix. See how to evaluate AI responses.
5. Set the stopping criterion in advance. For example: “if the automation rate stays below X% or the error rate above Y%, we will not scale up”.

The study by Brynjolfsson, Li and Raymond is a good design model. They took advantage of the staggered rollout of an AI assistant to more than 5,000 customer support agents and compared those who had access with those who did not yet. The average result was about 15% more issues resolved per hour, with larger gains for the less experienced (Brynjolfsson, Li and Raymond, revised 2024 version). The figure is specific to that context. The method, comparing groups over time, works for any company.

What surveys say about returns

There are many figures circulating about AI returns. Most come from surveys of executives, in which the company itself reports the result. They are useful for spotting trends, but are not equivalent to audited measurements.

• McKinsey, August 2026. Online survey conducted between May and June 2026 with 1,719 participants from 97 countries. Nearly nine in ten report regular use of AI in at least one function. The 37% who attribute some EBIT impact to AI are practically the same share as in 2025. The “high performer” group, which attributes at least 5% of EBIT to AI, remains at about 6%. This group is twice as likely to have defined processes for measuring the impact of initiatives (McKinsey, The state of AI in 2026).
• BCG, September 2025. Survey of 1,250 executives who make decisions about AI. According to the report, 5% of companies achieve AI value at scale and 60% achieve no material value, with minimal revenue and cost gains despite substantial investment. The report itself states that the data are self-reported, that many figures reflect expected rather than realized impact and that responses may carry perception bias. Among the challenges cited by respondents is the lack of clear metrics and ROI measurement (BCG, The Widening AI Value Gap).
• Stanford AI Index 2026. The economy chapter records organizational adoption of 88% in 2025 and use of AI agents still “in the single digits” in almost every function. It also summarizes studies with productivity gains of 14% to 15% in customer support, smaller in tasks that require deeper reasoning (AI Index 2026, Economy).

The overall picture, in the editorial team’s reading, is consistent: adoption is broad, gains on specific tasks show up in controlled studies, and the translation of this into company profit is rare and hard to attribute. Survey figures do not replace measurement in your own process. For the documented causes of projects that fail to deliver, see why AI projects fail.

Common pitfalls

• Counting only the license. Integration, human review and maintenance usually weigh more.
• Adding up hours as if they were money. Freed-up hours with no destination do not show up in the results.
• Measuring without a baseline, or with a baseline from another period or another team.
• Ignoring the cost of errors, including rework, complaints and legal risk.
• Extrapolating from the pilot. Pilot cases tend to be the easiest; real volume brings more exceptions.
• Using the market average as a target. A survey figure says nothing about your process.

Key takeaways

• Add up all costs, implementation and recurring, including tokens, human review, maintenance and governance, and cap consumption to avoid surprises.
• Measure the baseline before making changes and treat hours saved as money only when they are reallocated or avoid an expense.
• Run a pilot with a comparison group, a stopping criterion and error counting, and present estimates as ranges.
• Read return surveys as self-reported, dated data: as of 2026, they show broad adoption and impact on profit still concentrated in a few companies.