88 Percent Use AI, 6 Percent Earn From It

On August 25, 2026, McKinsey published this year’s edition of “The State of AI.” The figure that will travel from it into every slide deck: 88 percent of companies use AI regularly in at least one business function. Three months earlier, the ifo Institut (ifo Institute) had measured 54.5 percent for Germany. In July, the KfW — Germany’s state development bank — arrived at 20 percent.

More interesting than any single one of these numbers is the distance between them. Between 20 and 88 lies a factor of more than four, and none of these surveys is badly made. They measure different things. Knowing which saves you the obvious question of whether your own firm is “falling behind” — and leads you to a second number in the same study that says more than the 88 does: 6 percent.

At a glance

• What holds: McKinsey measures 88 percent regular use — but only 37 percent of companies can attribute any EBIT effect at all, and only 6 percent one of at least 5 percent.
• What follows: For Germany, usage rates in the same window run from 20 percent (KfW) through 26 (official statistics) to 54.5 percent (ifo). The spread has four explainable causes, and none of them is a measurement error.
• The catch: None of these numbers says anything about your firm. The only solid rate is the one you collect yourself.

Five surveys, five numbers

What the studies actually asked — and when:

Source	Usage rate	Survey period	What was asked
McKinsey, State of AI (2026-08-25)	88 %	May 4 – June 8, 2026	Regular use in at least one business function; 1,719 respondents from 97 countries
OECD, D4SME survey (2026)	61 %	2026, date not clearly stated	At least one AI application in use; 2,018 SMEs in 12 countries
ifo Institute (2026-06-05)	54.5 %	May 2026	AI use in the company; self-reported by German firms
Federal Statistical Office (as of 11/2025)	26 %	2025 reporting year	Official survey, companies with 10 or more employees
KfW Research, Fokus No. 554 (July 2026)	20 %	Panel waves 2022–2024	Deployment of at least one AI technology in the Mittelstand

A sixth survey is missing here on purpose. The report on artificial intelligence in Germany by Bitkom (the German digital industry association) is reproduced in secondary sources with two different usage rates, and I could not resolve which one is the original’s. A figure whose primary source cannot be verified does not belong in a comparison table. One finding those same sources do report consistently matters more for the Mittelstand anyway: 33 percent of AI users report unexpectedly high costs — token consumption, GPU hosting, systems integration.

Why the numbers diverge this far

Four causes, working together.

First: the question is not the same. Whether somebody somewhere in the firm uses a chatbot is a very low threshold; whether a model runs inside a defined process with a defined output is a very high one. McKinsey asks about regular use in at least one business function — deliberately broad. What sits behind such rates is visible in the OECD data: of its 61 percent of AI-using SMEs, 76 percent count as novices running basic applications. High adoption, low maturity.

Second: one source measures the past. The KfW’s 20 percent come from its Mittelstand panel, waves 2022 to 2024. The report appeared in July 2026; the reality inside it is roughly two years old — not a flaw, but the price of a large panel structure kept consistent over years. The comparison with a May 2026 snapshot is therefore meaningless. Official statistics look similar: the most recent figure from the Federal Statistical Office covers the 2025 reporting year, and the next survey is scheduled to appear only in autumn 2027.

Third: this is self-reporting, not measurement. None of these studies reads log files or invoices; they ask people what they believe they are doing. That distorts in both directions — whoever owns AI projects tends to overstate maturity, and whoever has shadow AI in the building understates it, usually without knowing. And the McKinsey figure is a statement about respondents, not about companies: the view from one division of a large group is not the situation of a firm with 40 employees.

Fourth: who gets asked shapes the result. Participation is voluntary, and those affected by a topic answer more often — which lifts every AI rate systematically. In vendor and consultancy studies the respondent list often comes from the firm’s own orbit; whoever is already digitalizing is overrepresented. Failed projects disappear twice over: nobody likes reporting them, and discontinued projects no longer count as AI use.

The second gap: from use to result

The actual news sits inside the McKinsey survey itself:

What is measured	Share
Uses AI regularly in at least one business function	88 %
Can attribute any effect on EBIT at all	37 %
Achieves an EBIT contribution of at least 5 percent	6 %

The gap between 88 and 6 is not a contradiction; it is the distance between tool access and operating result. The most frequently cited counter-study is harsher still: the MIT study on generative AI in business from summer 2025 — still in circulation unchanged in 2026 — finds that 95 percent of the pilots examined had no measurable effect on the profit and loss account, against 30 to 40 billion US dollars invested. MIT located the cause explicitly not in missing talent or in regulation, but in missing integration and context adaptation.

Where it fails in practice shows in BCG’s survey on AI at work (June 2026): 42 percent of regular users save eight hours a week, but 66 percent receive no guidance on what to do with the time gained. Time saved without a purpose appears in no set of accounts.

One object, five shadows: rates from 20 to 88 percent describe not five realities but five questions. The mark sits where the object touches the ground — the one point that does not depend on the viewing angle.

What this means for your business

First: stop measuring yourself against the market figure. Whether you are above or below 54.5 percent cannot be answered — your firm counts as a user in one survey and not in the other. What counts is the list of your use cases, with names, owners and status. As a starting point, ten typical use cases in the Mittelstand are more concrete than any rate.

Second: define your own number before you start. The difference between 88 and 6 percent is, at its core, a measurement problem — 37 percent of companies cannot even say whether AI affects their result. Decide up front which quantity is supposed to change (cycle time, cases handled, error rate, hours per case), and record the baseline while it still exists. How to make the output itself verifiable is in Evaluating AI outputs; what a trial that yields a solid answer looks like is in Setting up an AI pilot properly.

Third: put the cost of checking into the business case. A third of users report unexpectedly high costs, and checking the output costs working hours. A calculation that sets only saved hours against the token price is too optimistic; the honest one includes integration, control and maintenance — the same calculation that decides build or buy.

Conclusion

The 88 percent are not wrong, they are simply not your number — and the same goes for the 20 and the 54.5. Setting them against each other compares questions and survey years, not economies. What remains solid is the shared finding: use is rising fast, the demonstrable result follows far more slowly, and large companies scale more often than small ones — in the McKinsey survey, 54 percent of companies above one billion US dollars in revenue against roughly a third of smaller ones.

One number you collect yourself on a single process is therefore worth more than any study. It is small and unrepresentative — and the only one that applies to your decision.

If you are weighing which process lends itself to such a measurement, and how to set it up so that a number rather than an opinion comes out at the end, let’s talk. I work at this intersection as a business lawyer and developer in one person.

FAQ

How many companies in Germany actually use AI?

There is no single number that answers this. The ifo Institute reports 54.5 percent for May 2026, the Federal Statistical Office 26 percent for the 2025 reporting year, and the KfW 20 percent for the Mittelstand — the latter on a data base covering 2022 to 2024. The figures do not contradict each other; they answer different questions about different populations. Anyone quoting one of them should also state the survey period and the question asked.

Why does McKinsey report 88 percent and the KfW only 20 percent?

For two main reasons. McKinsey asks globally about regular use in at least one business function — a low threshold, met as soon as any single team uses a tool regularly. The KfW measures the deployment of at least one AI technology in the German Mittelstand, and its data base consists of panel waves from 2022 to 2024, so a snapshot roughly two years old.

What does the 6 percent figure from the McKinsey study mean?

6 percent of the companies surveyed qualify there as “AI high performers”: they attribute at least 5 percent of their EBIT to AI use. 37 percent can attribute any EBIT effect at all. The gap to the 88 percent of users describes the distance between tool access and operating result. A widely cited MIT study from 2025 points the same way: 95 percent of the pilots it examined showed no measurable effect on results.

Which number should I as a mid-sized company track at all?

Your own. What helps is a metric on one concrete process — cycle time, cases handled per day, error rate, hours per case — measured before AI is introduced and afterwards under the same conditions. That includes the full cost side: licences, token consumption, integration, maintenance, checking the output. This number is unrepresentative and still the only one that can carry your decision.

---

Sources — as of 29/08/2026

• McKinsey, The State of AI (2026-08-25; fielded 2026-05-04 to 2026-06-08) — https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
• ifo Institute, company survey May 2026 (2026-06-05) — https://www.ifo.de/en/facts/2026-06-05/more-half-companies-germany-use-artificial-intelligence
• KfW Research, Fokus Volkswirtschaft No. 554 (July 2026; data base Mittelstand panel 2022–2024) — https://www.kfw.de/PDF/Download-Center/Konzernthemen/Research/PDF-Dokumente-Fokus-Volkswirtschaft/Fokus-2026/Fokus-Nr.-554-Juli-2026-KI_Mittelstand.pdf
• Federal Statistical Office (Statistisches Bundesamt), ICT survey of enterprises — https://www.destatis.de/DE/Themen/Branchen-Unternehmen/Unternehmen/IKT-in-Unternehmen-IKT-Branche/IKT-U-Erhebung/info.html
• OECD, Empowering SMEs in the Age of AI (D4SME survey) — https://www.oecd.org/en/publications/empowering-smes-in-the-age-of-ai_bf5a9816-en.html
• MIT study on generative AI in business, reporting (2025) and its reception in summer 2026 — https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/ ; https://www.forbes.com/sites/terdawn-deboe/2026/07/28/ai-adoption-fails-95-of-the-time-small-business-leadership-is-why/
• BCG, AI at Work (2026-06-03) — https://www.bcg.com/publications/2026/ai-at-work-why-strategy-matters-more-than-tools
• Bitkom, publication on artificial intelligence in Germany — https://www.bitkom.org/Bitkom/Publikationen/Kuenstliche-Intelligenz-in-Deutschland

Note: the usage rate in Bitkom’s 2026 report is reproduced in secondary sources with two diverging percentages and is therefore deliberately not quoted here.

This article is general information and not legal or investment advice. As of August 29, 2026; the surveys cited are updated continuously, so please check the current state before making decisions.