MMOOZZGG.com

AI investment

AI ROI: how to calculate it and why many companies see none

5 min read2 sources

AI ROI is not guaranteed: in a 2026 McKinsey survey, 37% of organizations see some AI contribution to operating profit (EBIT), and about 6% credit AI with 5% or more of EBIT. You calculate payback before the start with a formula and check it after launch against a metric written into the contract.

On this page · 5 sections

What the numbers say

McKinsey surveyed 1,719 participants from 97 countries in May-June 2026. 44% of organizations scale AI across the company, 37% see at least some AI contribution to EBIT, and about 6% get 5% or more of EBIT from AI.

Between 37% and 6% lies a large group of companies: AI adds to their profit, but less than 5% of EBIT.

The MIT NANDA report of 2025 is harsher: 95% of organizations got no measurable return from generative AI. Only 5% of custom AI tools reach real use.

The conclusion for a company: AI does not pay back by itself. A specific solution on a specific process pays back, if its effect was calculated before the start and checked after.

At the same time, 60% of organizations expect their AI investment to grow next year. Every company that raises its budget will face the payback question.

Why many get no return

There is no return when AI is introduced without an answer to the question of which number it must change. The pilot runs, the launch report exists, but no money shows in the accounts.

1. No metric. Without one, the result is judged by the fact of launch, not by money. Why having no metric ruins a project is covered separately.

2. The pilot never reaches real work. According to MIT NANDA, only 5% of custom AI tools reach real use; the rest stay demos.

3. A calculation on the full salary. AI takes part of an employee’s day, and a calculation that assumes people are fully replaced gives a number no report will confirm.

4. Scale without selection. 44% of organizations scale AI, but about 6% see a noticeable EBIT contribution: the number of launched solutions does not equal return.

5. No owner. If nobody accepts corrections to the agent, it keeps making the same mistakes, and savings move months away.

The payback formula

Savings per month = employees on routine work x salary x 1.3 x the share of work AI takes, minus support. The factor 1.3 covers taxes and contributions on top of salary.

The share of work does not mean replacing a person; it is part of their day: answers to typical questions, moving requests, reconciliations, reports. You estimate it for each process separately, and almost the whole result depends on it.

Payback period = one-time cost of the project / savings per month. Savings in the first year = savings per month x 12 minus the one-time cost.

Besides support, subtract model fees per request and the time of your own people who handle disputed cases and accept corrections.

ROI, the return on each unit of money invested, comes from the same numbers: (savings for the year minus investment) / investment. A manager usually finds a period in months clearer, so we show both.

The calculation before the start uses only savings: you can check them in the accounts. Growth, such as more requests handled, comes on top of the calculation and shows after launch.

The McKinsey survey measures return as a share of EBIT: 37% of organizations see some contribution. For one process it is easier to measure money per month and to build the EBIT share from the sum of processes.

The metric before the start

Choose the metric before the contract, together with the process owner: what it was, what it must become and how to measure it. For example, the time to answer a client or the share of requests the agent processed without a person.

1. Take the starting number for the last month from your CRM or accounting system (such as 1C).

2. Write down the target number and the date you expect it by.

3. Agree who takes the number after launch, and from where.

4. Put the metric into the contract with the vendor.

A launch alone says nothing about money: according to MIT NANDA, 95% of organizations got no measurable return from generative AI. A metric moves the talk from we launched to we saved this much.

After launch, put the real share of work into the same formula. If the number is below the plan, the reason is visible: people do not pass requests to AI, the rules are out of date or the process changed.

A sample calculation

Take the payback calculator settings: employees on routine work 5, salary $950, share of work for AI 60%, salary with taxes x1.3.

Support of $480 per month is already subtracted. Savings per month: $3,220.

For an AI department at $4,999, this gives $33,600 of savings in the first year and a payback period of 2 months.

The numbers change with the inputs: with a smaller share of routine work, the period grows. Enter your own data in the calculator on the page about calculating the effect.

Where there will be no effect: less routine than one employee’s worth, a process nobody has described, no records collected. In such cases an honest calculation shows it will not pay back, and that is also a result: the money stays in the company.

These are the savings of one department, not the EBIT of the whole company. A company can reach the 5% EBIT threshold from the McKinsey survey when it has several such processes, each checked by a metric.

How to choose processes to invest in is covered on the page about investment in your own AI.

Sources and fact-checking

Checked by the editors on 14 October 2026.

  1. McKinsey, The state of AI in 2026, August 2026 (mckinsey.com)
  2. MIT Media Lab, Project NANDA: The GenAI Divide, State of AI in Business 2025, July 2025 (mlq.ai)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us