A chief AI officer is countable. Who decides is not

Two of the most quoted figures on AI leadership in 2026 point in different directions. IBM Institute for Business Value reports that 76% of organisations have a chief AI officer. BCG reports that 72% of CEOs say they are themselves the main decision maker on AI. Both figures are correctly reproduced by the people who cite them. Together they do not describe the same decision.

IBM measures an appointment

IBM Institute for Business Value, working with Oxford Economics, surveyed 2,000 CEOs and equivalent senior leaders across 33 geographies and 21 industries, fielded February to April 2026. The report Rewiring the C-suite puts the finding this way: in 2026, 76% of organisations have someone in this position, up from just 26% in 2025.

It is the second half of that sentence that does not hold.

The comparison figure comes from a different instrument

The 26% does not appear in IBM's CEO survey for 2025. That survey covered 2,000 CEOs across 33 geographies and 24 industries, fielded in the first quarter of 2025, and the report mentions a chief AI officer exactly once, in a sentence about returns.

The figure comes instead from a separate study, Solving the AI ROI Puzzle, in which IBM, together with Oxford Economics and the Dubai Future Foundation, surveyed more than 600 people who are themselves chief AI officers or the equivalent, across 22 countries and 21 industries, during the first quarter of 2025. That this is the study meant is IBM's own statement: endnote 3 of the 2026 report gives Solving the AI ROI Puzzle as the source of the 26% figure.

Two surveys, two populations, two geographic samples. What reads as a tripling over twelve months is two points from different measurements. The number of industries in the CEO survey also went from 24 to 21 between the two years, so not even the recurring survey has an unchanged sample to compare against.

A number from one population set against a number from another is two numbers, not a trend.

That does not mean the share is unchanged. It means the size of the change cannot be read from these two figures, and that an organisation making decisions on the difference between them is making decisions on a difference nobody measured.

BCG's doubling has no published counterpart

BCG's AI Radar 2026 is built on 2,360 respondents, 640 of them CEOs. The question asked what degree of involvement the respondent has in decisions relating to AI, and 72% of the CEOs answered that they are the main decision maker, which BCG describes as double the previous year.

The previous edition, AI Radar 2025, is built on 1,803 C-level executives in nineteen markets, fielded September to December 2024. The published report contains no reporting of who the main decision maker is. The number of CEOs in the two samples is not directly comparable.

The sixteen markets the report lists are the US, France, Japan, Germany, India, the UK, mainland China, Saudi Arabia, Spain, Italy, Morocco, South Africa, the UAE, Qatar, Hong Kong and Taiwan. None of them is Nordic.

The Nordic survey asks a different layer

Tieto commissioned Norstat to field a survey in February 2026 with 623 respondents in Sweden, Norway and Finland, roughly 200 per country, in organisations with at least 100 employees. The respondents are people who contribute to or support decisions relating to AI, which is the layer below the CEO rather than the CEO. It is a small regional sample beside BCG's 2,360 and IBM's 2,000, and it is also the only one of the three that puts its question to the people who prepare AI decisions rather than to the people at the top of the organisation.

Tieto reports the answers by country in its Swedish release, and the figures that follow are the Swedish subset of roughly 200 respondents. Asked whether the organisation has policies or guidelines for responsible AI, 35% of the Swedish respondents answer yes, 39% that the work is under way, 10% no and 16% that they do not know. For the Nordic total only a general figure is given, that around a third report policies are in place, and the highest share of finished policies is Norway with 38%.

So 16% of the Swedish respondents who contribute to AI decisions do not know whether there are any guidelines to decide by, the largest share of uncertainty among the three countries. With the 10% who answer no, a quarter of the Swedish layer prepares decisions without anything written to prepare them against. Asked about readiness for the EU AI Act, 35% of the Swedish respondents answer that they do not know.

The measurements do not contradict each other. They ask different people about different things: one whether a position exists, the other whether a basis exists. That is exactly why one cannot be used as an answer to the other. An organisation can have staffed the role and at the same time have a layer beneath it that does not know what applies.

What a measurement of your own position can and cannot show is covered in What an AI readiness assessment measures. This text concerns the next question: who takes the decisions between the measurements.

What argues against the reading above

The strongest objection comes from IBM's own material on chief AI officers. Organisations with one report 10% higher return on their AI spend, and are 24% more likely to say they outperform their peers on innovation. If the title were only bookkeeping, that association should not appear.

The limit of the objection lies in the method. The figures come from a survey in which the respondents are themselves chief AI officers or the equivalent, and the question concerns the value of the function they hold. It is a self-assessed outcome from a party with an interest in the answer, and it says more about what the role can achieve than about how many have been given it.

The Swedish measurement carries weaknesses of its own, and they change how the figures should be read. Norstat recruited respondents through an online panel rather than from a company register, and the Swedish subset is roughly 200 respondents. That is a small basis for conclusions about the organisations of a whole country, and differences of a few percentage points between the countries should be read accordingly. On top of that, respondents assess their own organisation. The question is whether policies exist, not whether they are decided, known or followed, and the answer is one individual's picture of their own workplace.

The criterion: the last three decisions

The question of ownership can be settled without counting titles, and it is settled on your own organisation's material.

Ownership that leaves no trace in three decisions already taken is a line on an org chart.

Take the last three AI decisions that were actually taken. A vendor proposal accepted or declined. A department that did or did not get a tool. A change in the rules that led to an action. Write down who took each one, and when.

Three different names means the questions find whichever table happens to be free. The same name three times, someone who was briefed when the question arrived, means the ownership exists whatever the role is called in the directory. And if there is nothing on any of the three lines, because the decisions were never taken, that is the answer that counts for most.

The criterion gives different organisations different answers. If AI sits on every leadership meeting and several initiatives run in parallel, it points to a full-time role with a mandate of its own. If the decisions arrive every week without filling a week, it points to a standing counterpart in a defined arrangement. If the lines are empty, what is missing first is neither of those, but a point on the agenda.

Which of the answers applies cannot be read from a percentage measured on 2,000 CEOs across 33 geographies. It can be read from three lines.