Who owns the AI line in the leadership team's budget

In organisations that budget by calendar year, the budget for 2027 is set in the autumn of 2026. AI can sit in it as a standing line or as an experiment, and the difference decides more than the amount. A business case for AI is usually read as a calculation of what the investment costs and what it saves. The most important thing it does is something else: it names who is accountable for the investment delivering anything.

That question is often settled before anyone asks it, by the line the item is entered on.

The experiment money is shrinking

The venture capital firm Andreessen Horowitz asked a hundred CIOs across fifteen industries how they buy generative AI, and published the answers in June 2025. A year earlier, a quarter of spending on language models came out of innovation budgets. Now the share was seven per cent. More and more is paid through central IT budgets and business unit budgets instead, which the firm reads as a sign that generative AI is no longer treated as an experiment.

Two things belong with those figures. The survey was run by a venture capital firm, and it does not say which countries the respondents are in. The figures say something about the direction, and less about the level.

Use is also growing in Sweden, measured in an entirely different way. According to Statistics Sweden, 35 per cent of Swedish enterprises used AI in 2025, ten percentage points more than the year before. These are official statistics, collected by a very different method from a poll of CIOs. The figure covers enterprises with at least ten employees in the industries the statistics include, and the financial sector is not among them.

When money moves from an experiment to a standing line, it also changes owner. An experiment has a project lead and an end date. A standing line has a budget holder and is renewed every year.

The line decides who owns the AI question

One way in is the licence. If IT buys it, because that is where the contracts, the access rights and the security sit, it lands on IT's line. The follow-up comes with the line. Next autumn the question asked is whether the licences are being used, because that is what IT can see. Whether anything in the work changed is rarely asked, because no one on that line runs the work.

At that point, no one has decided who owns the question. The leadership team approved an amount, and the placement followed from who signed the contract.

Where an AI item is placed is an ownership decision, even when nobody makes it.

That makes the budget item the leadership team's first AI decision, whatever has already happened in the organisation. Tools may have been bought and guidelines written. All of that can happen without the leadership taking a position on who owns the result. The budget is the first document in which the question cannot be avoided, because every line has to sit somewhere.

How a consumption cost behaves differently from a licence, and what that does to the same line during the year, is covered in AI is no longer a licence cost. The question that comes first is whose line it is.

Those who get the most from AI have redesigned the work

McKinsey's global survey, fielded in June and July 2025 among 1,993 respondents in 105 countries, singles out a small group: those who attribute more than five per cent of operating profit, and significant value, to AI. The group is about six per cent of respondents. Two of the differences from the rest lie outside IT.

The group is 109 respondents, and both the result and the ways of working are self-reported. An association between two self-reported answers does not show that one produces the other. The same pattern appeared in the previous wave, fielded in July 2024 among 1,491 respondents, with the same limitation. There McKinsey writes that the redesign of workflows is, of the attributes tested, the one with the biggest effect on the ability to see an EBIT impact from generative AI. The wording is McKinsey's, and it rests on the same kind of self-reported answers.

The measurement furthest from an executive survey is Danish. The economists Anders Humlum and Emilie Vestergaard, with surveys conducted in collaboration with Statistics Denmark, linked the answers to administrative records for about 25,000 workers at 7,000 workplaces. The sample is eleven occupations highly exposed to AI, not Danish workers in general. Two years after the launch of ChatGPT there is no effect on earnings or recorded hours, and the study rules out effects larger than two per cent. What has moved is the content of the work. According to the researchers, employers absorb the technology by reorganising tasks, and 85 per cent of users say the time the tool saves goes to other job tasks. Use is highest where the employer encourages it, provides its own tool and trains staff: there, 93 per cent of workers have used AI chatbots at work. That is an association and not evidence of what causes what. The study is a working paper, last revised in March 2026.

The time freed up does not disappear, then, but neither does it turn into a result by itself. What it is used for is decided in the business, and that is one of the decisions a licence on IT's line does not make.

Redesigning a workflow is not something IT can order. It takes the person who runs the work: who does which step, what can be skipped, what has to be checked and by whom. A licence on IT's line gives the tool to everyone and the responsibility for the change to no one. For a professional services firm this is especially plain, because the workflow being redesigned is often the one that is billed, as the page on professional services sets out.

Four parts of a business case for AI that can get a yes

A case the leadership team can decide on has four parts, and the calculation is only one of them.

A named workflow and its baseline. Which step is to change, how long it takes today, and how often the result has to be corrected afterwards. "Efficiency" is not an answer. Without the baseline there is nothing to compare with next autumn, and what a measure that holds looks like is set out in Do you know whether your AI programme worked.

A named owner for the change. A manager in the business who runs the work in that workflow. IT owns the platform, the licence and the security. The line manager owns the work becoming different, and it is the line manager who reports the outcome.

The costs after the licence. The licence is the item that shows. Behind it sit the data work that makes the material usable, the rollout into daily work, the follow-up, and the rebuilds that follow when a provider changes or retires a model, which is covered in Who decides when your AI agent has to be rebuilt. BCG writes that the companies that have moved beyond the experiments follow a rule: ten per cent of resources to algorithms, twenty to technology and data, seventy to people and processes. The rule is a recommendation for how resources should be allocated, not an account of how they are. What the report measures is where the difficulties lie: about 70 per cent concern people and processes, 20 per cent technology and 10 per cent the algorithms, according to BCG's experience, which the firm writes is corroborated by its survey. That is where a case that only counts licences has its blind spot.

A decision point. When is it measured again, and what decides whether the investment continues, is scaled up or is stopped. In McKinsey's 2024 wave, fewer than one in five said their organisation tracked KPIs for its generative AI solutions. An item without a decision point renews itself.

The case contains no promised return as a percentage. It cannot be calculated before the baseline is measured, and a percentage from a supplier who has not seen the organisation's data is a story. Producing the baseline, the prioritisation and a decision basis for the leadership team is what an AI Readiness Assessment does. Which use cases come first, and with which owners, is the strategy work that follows.

The objection: part of the item belongs with IT

There is an opposing consideration that carries real weight. Not everything should move out of IT, and anyone who reads the argument above as a reason to hand the AI budget out to the business units makes a new mistake.

A licence spread across ten units produces ten contracts, ten access models and data flows no one has an overview of. The same thing happens when tools are bought on the business units' own accounts without anyone deciding it. Ownership is then as undecided as on IT's line, only scattered.

Two lines need two owners: one for the platform and one for the change.

A similar boundary shows in how organisations divide responsibility between a central function and the business units. In its July 2024 wave, McKinsey asked how different parts of AI work are organised.

The table is based on 1,229 respondents whose organisations use AI in at least one function, with "don't know" answers removed. Risk and data are most often held centrally. Adoption is what is most often shared. A central function is not always IT, but the pattern points the same way as the objection. It does not bring the argument down, it bounds it: the licence, the security and the data belong with IT, and the change in the work belongs with whoever runs it.

Strike the item and see who notices

The criterion can be applied to the budget draft as it stands today. Take each AI item and imagine it gone. Who notices first?

If the answer is IT, because a platform, an access right or a security control disappears, the item is infrastructure. It sits where it should.

If the answer is a manager in the business, because a workflow slows down or a step has to be done by hand again, that manager should own the change the item pays for. The case should be the manager's, and it is the manager who reports the outcome next autumn.

If the answer is no one, not even after a quarter, the item is a licence without a change. It needs a case before it is renewed, or no place in the budget.

The same question gives different organisations different budgets. One that gets the answer IT for every item has a platform and no change to own yet. One that gets three different answers has found its first AI decision. Who made the most recent AI decisions shows who has decided so far, which is tested in A chief AI officer is countable. The budget item decides who decides in 2027.