Buying AI tools is not an AI strategy

Somewhere in your organisation there is a licence nobody is using. It was bought with good intentions, probably in a bundle, probably at a per-seat price that felt reasonable when multiplied across the workforce and terrifying when you saw the annual total. The rollout email went out. There was a lunch-and-learn. For a few weeks the dashboards showed a spike. Then the line drifted back down, and now a large share of the seats you pay for sit cold.

This is not a story about one vendor or one bad quarter. It is the central pattern of the enterprise AI era, and the numbers are now clear enough to stop pretending otherwise.

The gap between what was bought and what was used

In August 2025, MIT's NANDA initiative published The GenAI Divide: State of AI in Business 2025. Its headline finding, drawn from roughly 150 executive interviews, 350 employee surveys and an analysis of 300 public deployments, was that about 95 percent of enterprise generative-AI pilots delivered no measurable impact on profit and loss. Only 5 percent produced significant value. This was after $30 to $40 billion in enterprise spending on the technology.

McKinsey's State of AI survey, updated through 2025, tells a compatible story from the other end. Some 88 percent of organisations now report using AI in at least one business function, up from 78 percent a year earlier. Adoption, in other words, is close to universal. Yet only 39 percent report any effect on earnings at the enterprise level, and among those, most say AI accounts for less than 5 percent of EBIT. Roughly 6 percent of firms qualify as genuine high performers.

Boston Consulting Group frames the same divide as a value gap. In its 2025 work, only about 5 percent of companies were generating real value from AI at scale, while around 60 percent had captured little or none. BCG's follow-up carried a title that reads like a confession from the whole industry: Why Strategy Matters More Than Tools.

Three different research teams, three different methods, one conclusion. The tools are everywhere. The value is almost nowhere. Adoption has been mistaken for transformation, and the receipts do not match the results.

Why the money did not turn into anything

The comforting explanation is that the technology is not ready. That is mostly wrong, and believing it will cost you another year. The models are more capable than the work most people ask of them. The MIT researchers were direct about this: the failures are organisational, not technical. They called it a learning gap. Generic assistants perform well for an individual because they are flexible, but they stall inside a company because they do not learn the company's workflows, context or standards, and the company does not change its workflows to meet them.

Look closely at any stalled rollout and you find the same three absences.

The first is a missing answer to a basic question: what is this for? A licence is not a use case. Handing someone a general-purpose assistant and hoping they discover value is like handing out spreadsheets with no mention of accounting. The people who thrive are the ones who already knew what they wanted to automate. Everyone else opens the blank box, types a question they could have answered themselves, feels underwhelmed and quietly closes the tab.

The second is that the work never changed. Real value comes from redesigning a task so the tool sits inside it, not beside it. If your underwriters, analysts or support agents run the same process they ran in 2023 and now have a chatbot in an adjacent window, you have added a step, not removed one. McKinsey's high performers are distinguished less by which tools they bought than by having fundamentally redesigned workflows around them.

The third is trust, and this is the one leaders consistently underestimate.

People do not resist AI because they are stupid

There is a name for what happens next, and it predates ChatGPT by a decade. In 2015, Berkeley Dietvorst, Joseph Simmons and Cade Massey published Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err in the Journal of Experimental Psychology: General. Across a series of experiments, people watched an algorithm and a human each make forecasts. The algorithm was demonstrably more accurate. Yet once participants saw the algorithm make a mistake, they abandoned it far faster than they abandoned a human who made the same mistake. We forgive people for being wrong. We do not extend that grace to machines.

This is the behaviour sitting underneath your usage numbers. An employee tries the assistant on something they know well, catches it in one confident error, and writes the whole category off. The single failure outweighs a dozen quiet successes. It is not irrational, exactly. It is a deep, well-documented human reflex, and no procurement decision changes it.

The same researchers found the way through. In later work on overcoming algorithm aversion, they showed that people become markedly more willing to rely on an imperfect algorithm when they are allowed to adjust its output, even slightly. Control restores trust. This has an unglamorous, practical implication: adoption rises when you position AI as a draft to be edited by an expert, not an oracle to be obeyed. People will use a tool they can overrule. They will not use one that asks them to switch off their judgement.

There is a structural version of the same gap. Microsoft's 2025 Work Trend Index, drawn from 31,000 workers across 31 countries, found that 67 percent of leaders felt familiar with AI agents while only 40 percent of employees did. Leadership is running ahead of the people expected to do the actual work. That distance is where mandates go to die.

What actually closes the gap

None of this argues against buying the tools. It argues that the tools are the cheap, easy 10 percent, and the reason so little value has appeared is that most organisations stopped there. Here is where the other 90 percent lives.

Start from work, not from software

Pick two or three high-volume workflows where the cost of the current process is visible and the output is checkable. A first draft of a claims summary. Responses to routine supplier queries. The initial pass of a compliance review. Redesign the workflow end to end with AI inside it, then measure whether the task got faster, cheaper or better. If you cannot name the workflow, you are not doing strategy, you are doing hope.

Enable by role, not by rollout

A company-wide webinar teaches nobody how to do their specific job differently. A lawyer, a recruiter and a financial analyst need entirely different prompts, examples and guardrails. The organisations pulling ahead build role-based enablement: short, concrete, job-shaped training that shows a person the three things that will save them an hour this week. Enablement is not an event. It is a capability you build and keep building.

Feed the tools your context

The reason a generic assistant underperforms in your business is that it does not know your business. It has never read your tone of voice, your pricing logic, your regulatory constraints or your past decisions. Closing that gap, through retrieval, well-built prompts, curated internal knowledge and the accumulated skills of your best people, is what separates a novelty from an asset. Context is the moat, and it is one you have to dig yourself.

Measure behaviour, not licences

Stop reporting seats sold. It tells you what you spent, not what changed. Measure the things that indicate work is actually different: weekly active use inside a redesigned workflow, time saved on a named task, quality scores on the output, the share of a team that has genuinely shifted how they operate. If your AI dashboard shows only how many people you paid for, you are measuring the invoice, not the impact.

Have leaders do the work in public

Algorithm aversion does not yield to memos. It yields to seeing a respected colleague use the tool, catch its errors, correct them and get to a better answer faster. When an executive shows the actual prompt they used to prepare for a board meeting, and is honest about where it was wrong, they give everyone permission to try, fail and keep going. Adoption is social before it is technical.

The Monday version

If you want something to do this week, it is small and it is not buying anything.

Choose one workflow. Sit with the three people who run it most and watch them work for an hour. Find the single step that is slow, repetitive and low in judgement. Redesign that one step with the tool inside it. Agree in advance how you will know if it worked, in minutes saved or errors avoided. Run it for two weeks. Then do the next one.

That is the whole method, and it is deliberately unimpressive. The firms in the 5 percent are not the ones that bought the most licences or moved the earliest. They are the ones that treated AI as a change in how work gets done rather than a thing to be installed, and then did the patient, specific, human work of making that change stick. The licence was never the strategy. It was the receipt for deciding to start.