The number nobody decks

Open any AI steering committee deck from this year and there is a number near the front. Hours saved per week. Share of staff using the tools. A self-reported productivity uplift. It is usually large, it is usually green, and it is usually the only number there.

Two other numbers exist. One shows what the gain looks like once it reaches revenue. The other shows whether the organisation itself has changed. Both are smaller, both are harder to produce, and both are almost never on the slide.

The first of those two is the number nobody decks. This insight sets out what each of the three numbers is, what the best evidence available in October 2026 says each one currently reads, and how to build a business case on the one that actually reaches the income statement.

1.Three levels, three numbers

Auron describes AI return in three levels: quick wins, differentiation and transformation. The levels are familiar. What is less familiar is that each one has its own proof number, measured in a different place by a different owner.

2.Level one: the number everyone decks

The first number is real and it is high. In McKinsey's 2026 survey of 1,719 participants across 97 nations, fielded in May and June, 80 percent say AI has improved their productivity. In a separate study of nearly 6,000 senior executives in the US, UK, Germany and Australia, 69 percent of firms actively use AI. And when nearly 750 corporate executives, most of them CFOs, were asked by the Federal Reserve Bank of Atlanta what AI had done for labour productivity at their firms in 2025, they reported 1.8 percent.

None of these numbers is wrong. They measure something that exists: people finding the tools useful. The problem is what happens to them on the way to the slide. A share of people who feel faster is presented as if it were a share of the business that has become more productive. Those are different claims, and only one of them is in the numbers.

There is also a quieter signal in the same data. Executives who use AI regularly average about 1.5 hours a week with it. Level one is wide, but for many organisations it is not yet deep.

3.Level two: the number nobody decks

The Atlanta Fed study did something most surveys do not. Alongside asking executives what AI did for productivity, it computed what their own revenue and employment figures imply. The two do not match.

Reported is three times implied for 2025. The authors call it a productivity paradox and read it as delayed revenue realisation: the gains are perceived before they are booked. That reading is generous, and it may be right. Either way, it means the number on the slide is running a year or more ahead of the number in the accounts.

The wider evidence points the same way. Across the four-country panel of nearly 6,000 executives, 89 percent report no productivity impact on their own firm over the past three years. Looking forward, the same executives expect AI to raise productivity by 1.4 percent over the next three years. In the McKinsey survey, 37 percent of respondents attribute any positive EBIT contribution to AI, essentially flat on last year.

The revenue-implied number is the one a CFO recognises, because it is computed from things the CFO already owns. It is smaller than the reported number in every source cited here. That is precisely why it stays off the slide.

4.Level three: the number that requires an admission

The third number is the hardest to produce and the rarest to see. In McKinsey's survey, 6 percent of respondents attribute 5 percent or more of EBIT to AI and describe its value as significant. That share is unchanged from 2025.

What separates that 6 percent is instructive, because it is not what most programmes fund. Roughly 80 percent of both the top performers and everyone else pursue efficiency gains from AI. Efficiency is universal, so it separates nobody. What the top group does differently is re-engineer the work itself: nearly three quarters fundamentally redesigned workflows because of AI, against about a quarter of everyone else, and most pursue growth or innovation alongside efficiency.

European firm-level data, where productivity comes from company accounts rather than from what executives report, is consistent with that picture. Across a survey frame of roughly 12,000 European non-financial firms, the Bank for International Settlements finds AI adoption associated with a one-off productivity level shift of about 4 percent. The effect is amplified most by investment in training, and less by spending on software and data. The input that converts a tool into a result is training, not more tooling.

Level three is rarely decked for a different reason from level two. Producing it means showing which processes were redesigned, and therefore which were not.

5.Why the deck carries the wrong number

None of this requires anyone to be dishonest. Three ordinary mechanisms put level one on the slide and keep the others off it.

1. The instrument is self-report. The cheapest way to measure an AI programme is to ask the people using it. They are telling the truth about how it feels. The Atlanta Fed gap is the size of the difference between how it feels and what the revenue shows.
2. Time saved is absorbed, not released. An hour reclaimed in a process nobody redesigned does not leave the building. The queue shortens, the day fills back up, and nothing reaches the income statement unless someone decides what gets smaller.
3. Level one is the only number available in month three. Business cases are approved before levels two and three can be measured. The first number available becomes the only number reported, and by the time the others exist, nobody has asked for them.

6.Building the case on the number nobody decks

The fix is not a better survey. It is deciding, before anything is built, which number the use case will be judged on, and making that the number that gates the money.

1. Tag every use case twice. By the level it targets, quick win, differentiation or transformation, and by business case type: Defend, Extend or Upend. A Defend case that claims transformation is a mislabelled case.
2. Write the level two number into the case before any code. Name the revenue, cost or output line that must move, its baseline today, the owner of the change and the date it lands. If nobody can name the line, the case is a level one case and should be funded as one.
3. Gate the money on it. At the Fund gate, no baseline means no funding. At the Scale gate, the gain must hold at full volume and appear in the revenue-implied measure, not only in the survey. At the Run gate, the gain is measured net of the AI spend that produces it.
4. Report all three numbers, every time. A deck that shows level one alone should be read as a deck that has not yet measured the other two. Saying so out loud is cheaper than discovering it at budget review.

7.For a Swiss or European board

The evidence base is uneven, and it is worth being exact about where. The two executive studies are anchored in the US, with UK and German panels in the larger one. The European firm-level study excludes financial firms by construction and contains no Swiss firms. No Swiss study we have found measures the gap between reported and revenue-implied AI returns, and the official Swiss firm survey that does cover AI predates generative AI.

The right conclusion for a Swiss bank, insurer or MedTech firm is therefore neither panic nor comfort. The direction is established across every source and every method we hold. The magnitude for your own firm is unknown. That is an argument for measuring level two now, not for waiting until someone else publishes it.

8.The limits of this evidence

• Every figure here except the BIS result is reported by executives. The revenue-implied measure is better anchored, but it is still derived from survey answers.
• The McKinsey survey is opt-in and published by a consultancy. Its top-performer group is 92 respondents, so comparisons built on that group are directional.
• The Atlanta Fed and NBER studies share an institution and one author. They agree, but they are not fully independent.
• The BIS result is a one-off level shift, not an annual rate, and its sample excludes banks and insurers. Its adoption question also groups AI together with big data analytics, so some firms counted as adopters may use no AI at all.

The useful question about your next AI deck is not whether the number is good. It is which of the three it is, and who owns the other two.

Auron Advisory is formed to build the intelligent enterprise: organisations where AI and disruptive technologies not only increase productivity but open new markets and materially improve the ones already there.

Sources