AI ROI by Industry: What the 2026 Public Data Actually Shows

AI ROI by industry does not exist as an audited public dataset, and anyone showing you a table of "average return per sector" is blending perception surveys with consultancy estimates. What does exist, and is far more useful, is two things: official adoption data broken down by sector, and serious surveys of how many companies are actually seeing financial return. With those two pieces you can place your own company and compute your own number — the only one you can decide on.

This article collects the public data available as of August 2026, with a source and a date attached to every figure, and explains what each one actually measures. Then it gets practical: where the concrete return sits in each type of business, and how to calculate yours from four numbers you already have. If you want the general calculation method, it is in our guide to the ROI of artificial intelligence in business; here we deal with the cross-industry comparison and its traps.

What do the aggregate numbers say about AI returns?

They say most companies still see no measurable financial return, and that the minority who do see it see it clearly. It is an uncomfortable data point but the most honest one available, and it is worth having in front of you before signing any project.

• PwC, 29th Global CEO Survey (January 2026). Across 4,454 chief executives in 95 countries: 56% report that AI neither increased revenue nor reduced costs in the preceding twelve months. Around 30% report a revenue increase, 26% a cost decrease, and only 12% both. PwC defines an increase or decrease as a change of 2% or more.
• McKinsey, The state of AI (November 2025). 39% of respondents attribute some EBIT impact to AI use, and most of them put that impact below 5% of their organization's EBIT. A small group — around 6% — clears that threshold.

The correct reading of those two data points is not "AI does not work". It is that aggregate return is heavily concentrated: a few companies capture a lot of value and many capture little or none. What separates them is rarely the sector; it is whether the project attacked a measurable process or a vague ambition to "use AI".

Adoption by sector: the official numbers

Adoption is measured officially and by sector, and there the differences are enormous. These are the most recent public figures from the European, Spanish and US statistical offices.

Three conclusions genuinely hold across these numbers. First: the size gap (55% versus 19% in the EU) is far larger than the gap between most sectors. Second: the sector with the highest declared adoption is the one that sells technology, which tells you very little about the return on automating an accounting firm or a factory. Third: adoption roughly doubled in a single year across the EU, so any figure older than eighteen months is now decoration.

Why adoption is not ROI

Because adopting means "we used some AI technology in the past year" and return means "we earned or saved money attributable to it". Between those two sits a chasm — precisely the chasm the PwC and McKinsey numbers describe.

The three reasons a high-adoption sector can show low return:

1. Individual use without process redesign. Every employee using an assistant to write better does not change the P&L; it changes the experience of working. Return appears when a step of the process is removed, not when it is accelerated.
2. Pilots that never scale. A trial that works in one team and never leaves it consumes budget and produces no saving. It is the most common pattern in large companies.
3. Uncapitalized savings. Freeing 200 hours a year is not a saving until those hours go into something valuable or a hire is avoided. If nobody decides what happens to the freed time, the return evaporates.

Where the return sits in each type of business

This table is not public data: it is our own reading, drawn from the projects we build and operate, of which process concentrates the return in each type of company. We use it to prioritize, not as a promise.

The recurring pattern: short paybacks sit where volume is high and unit cost is clear (a contact, a booking, a document); long paybacks sit where the saving is a problem avoided (downtime, a penalty, a defect), because measuring what did not happen always takes longer. If you need the unit-level calculation for a specific agent, we break it down in how much an AI agent really saves.

How to calculate your ROI from four numbers

You do not need an industry study: you need four figures from your own company and one hour. This is the calculation we run in every diagnosis before proposing anything.

1. Volume. How many times a month the process you want to automate actually happens (contacts, invoices, quotes, case files).
2. Real unit cost. Minutes per unit multiplied by the fully loaded hourly cost of whoever does it. Fully loaded means salary plus employer costs plus overhead, not gross pay.
3. Genuinely automatable share. Be conservative: 50-80% in processes with clear rules, less where exceptions dominate. If a vendor tells you 95%, ask them to put it in writing.
4. Total three-year cost. Build, plus annual maintenance (10-20% of build cost), plus model usage, plus your own team's internal hours — which almost nobody counts and which are entirely real.

From there the return is arithmetic. And the interesting output is usually not the percentage but the break-even point: under twelve months and the decision is easy; over twenty-four and you are probably attacking the wrong process. For the ongoing measurement side — which indicators to watch afterwards — see our guide on how to measure AI productivity.

The three mistakes that inflate ROI on paper

Nearly every optimistic business case we have reviewed fails in the same three places, and all three are fixable before spending a euro.

• Counting hours nobody will reassign. A saving is only real if it becomes more output with the same headcount, or a hire avoided. Write it into the business case: what the freed hours will be used for.
• Forgetting the supervision cost. An agent needs someone to review, correct and maintain it. Budgeting zero supervision hours is the fastest way to inflate a return.
• Benchmarking against the ideal process instead of the real one. Your starting point is not how your process should work, but how it works today, with its errors and rework. Measuring the baseline before automating is what avoids this trap.

If you would rather run this exercise with a team that does it weekly and will also tell you when it is not worth it, that is exactly the diagnosis we open with from our AI consulting practice.

Frequently asked questions

Is there any industry where AI does not pay off?

Not by industry, but by process profile. It does not pay off where volume is low, exceptions are the norm and every case demands expert judgment: there the cost of building and supervising exceeds the saving. A firm handling twenty cases a month does not have the same business case as one handling two thousand.

Why do adoption figures vary so much between studies?

Because they measure different things. Eurostat and Spain's INE survey enterprises from ten employees up, using a narrow definition of AI technologies; other studies include sole traders, or count individual assistant usage, or ask executives for their perception. Before quoting a percentage, always check who was surveyed and what counted as use.

How long does return typically take to appear?

In high-volume processes with a clear unit cost, 2 to 9 months. Where the value is avoiding problems, 6 to 18. Any promise of return in weeks refers to perceived time savings, not to P&L impact.

Is being an SMB a disadvantage for getting return?

For adopting it is — the large-versus-SME adoption gap Eurostat measures is over 35 points — but for getting return it is often an advantage: projects are smaller, one person decides, and the process actually gets redesigned instead of negotiated between departments.