Earlier this year McKinsey published its State of Organizations 2026 report, and one line has stayed with me since. It comes from an executive the report quotes: for every dollar spent on AI technology, roughly five should be spent on people. I have thought about that line every time I have looked at an AI budget since, because almost all of them are built the other way around. The typical European AI budget for the year runs to a respectable seven figures, nearly all of it on platform licences, integration work, and a handful of agent pilots, while the training and adoption line sits near the bottom of the page at roughly a tenth of the technology spend. It usually gets signed off without much discussion, partly because the technology figures are easier to defend and partly because nobody in the room is quite sure what a serious investment in people on AI would actually look like.
This is the contradiction sitting inside most AI budgets today: companies spend heavily on access and very little on the conditions that make that access useful. The pattern in most companies is close to the inverse of that five-to-one ratio, and the consequences show up in the data, where eighty-eight percent of organisations now use AI in at least one function yet fewer than one in five report a significant impact on the bottom line. The difficulty is largely one of interpretation, because most leaders read “invest in people” and translate it into a training budget, which is far too narrow a reading. The ratio is pointing at something more structural, and understanding what it actually contains is the difference between a budget that buys real movement and one that buys licences nobody ends up using well.
Why the ratio is the shape it is
The ratio reflects a basic property of AI as a technology. Software in its earlier forms could be deployed and the productivity would arrive more or less on its own, because most of the design work had been done before the tool ever reached the user. AI inverts this, since the model is general and the value it delivers depends almost entirely on the judgement of the person sitting in front of it. When people are handed that general model with no role-specific direction, the natural thing to do is point it at the easiest, lowest-stakes work of the day, which is why so many organisations report widespread AI use and very little business value to show for it. The gap sits in how the tool was introduced, well before it says anything about the people using it.
That is why the ratio sits at five-to-one rather than one-to-one or two-to-one, since the cost is rarely in the licence itself. It sits in the surrounding work that turns access into adoption, adoption into skill, and skill into the kind of business outcomes that show up on the income statement. PwC’s 2026 AI Performance Study found that 20% of companies capture 74% of AI’s economic value, and that the companies pulling ahead do so by reinventing more of the organisation around AI, using the same tools everyone else has bought. Those leaders are 2.6 times more likely than their peers to say AI has improved their ability to reinvent their business model. Reinvention is a people activity, the part the tools cannot do on their own, and a company that buys only the platform ends up optimising the edges of a process it never redesigned.
What the five actually buys
If a leader genuinely committed to the ratio, the technology line would absorb roughly one sixth of the total AI envelope, covering licences, integration, security, and platform engineering, and the remaining five sixths would sit across five categories that most organisations currently fund either accidentally or not at all.
The largest of these is role-specific capability building.
The largest of these is role-specific capability building. McKinsey estimates that around 75% of roles will need to be reshaped as AI moves through the organisation, and that reshaping is rarely glamorous work, because it means sitting with a credit analyst, a customer success manager, or a procurement lead and rewriting the actual content of their week with AI inside it. This is why generic prompt courses rarely move the needle, and why the training that works is built around a real task from the person’s own job and ends in a use case they can keep using.
A second and often neglected category is manager enablement.
Middle managers are the layer that converts strategic intent into daily behaviour, and they are also the layer most exposed when AI starts to change how work is done, facing the simultaneous task of using AI themselves, evaluating its use by their teams, and explaining to their reports what now counts as competent work. Without explicit investment in this group, adoption stalls at exactly the level where it carries the highest cost to the business, which is why the organisations that take AI seriously treat it as the next priority after capability building rather than an afterthought.
The remaining three categories carry the rest.
Role and workflow redesign is the slow work of rewriting standard operating procedures, decision rights, and handovers once AI sits inside a process, and it is the structural layer that stops an organisation sliding back into pre-AI habits the moment the training stops. Change and community work covers internal learning networks, peer mentoring, visible-wins programmes, and the psychological safety people need before they will admit that AI is not yet helping them, without which the learning stays private and never compounds. Governance and measurement covers leadership literacy on AI risk, decision frameworks for agent deployment, and the basic mechanisms for tracking whether any of this is producing business value, an area where Deloitte’s State of AI in the Enterprise 2026 found that only 21% of organisations have mature governance for AI agents, which is a fair signal of how much remedial investment most companies still owe themselves. The proportions matter less than the principle, which is that the five in the five-to-one ratio works as a portfolio spanning several distinct categories of work, each of which needs its own plan.
What I have seen up close
In my work with organisations through ZeroTo100, the budgets that perform best share one habit above all: the leadership team runs the people work and the technology work in parallel. The people work begins before the platform arrives, runs alongside the build, and carries on for at least a year after it is in place. The organisations that run the two in sequence instead, and that is most of them, end up paying the people cost twice, once as wasted licence spend and again as the remediation work needed to recover momentum after the first wave of adoption fails to land. In practical terms, the five-to-one ratio is an argument for starting the people work early and keeping it running alongside everything else.
The diagnostic question
Most of these budget conversations get easier when a single question is allowed to do the work. If a member of your leadership team can point at the AI line in next year’s budget and describe, in one sentence per category, what the people money is funding and how its outcome will be measured, you have a credible plan. If the only specifics in the document are the names of the platforms being bought, the budget is in the inverted shape, and its EBIT impact will follow the same disappointing pattern as the rest of the eighty-eight percent. The ratio asks for a different shape of spending rather than a larger one, since the total can stay exactly the same while the money sits in different places, and that shape is visible from the document itself.
The harder question, and the one worth carrying into the next leadership meeting, is whether the organisation has the internal capability to spend the five well even if it were available. If it does not, the first investment to defend is the one that builds that capability, because every euro that follows depends on it.
Sources