AI Strategy vs AI Adoption

In the spring of 2024, Microsoft and LinkedIn surveyed 31,000 knowledge workers in 31 countries. Seventy-five percent said they were already using generative AI at work. Of those users, 78 percent were bringing their own AI tools rather than waiting for their employer to provide them. The leaders in the same survey told a different story. Seventy-nine percent agreed their company needed to adopt AI to stay competitive, and 60 percent worried that their leadership lacked a plan and a vision to implement it¹.

Note what each figure counts: the 75 percent is all knowledge workers surveyed, the 78 percent only those who use AI, and the 60 percent only leaders. A year later Gallup found the same gap from the employee side. Forty-four percent of all US employees surveyed said their organization had begun integrating AI, but only 22 percent said it had communicated a clear plan or strategy for doing so².

Put those numbers in front of a leadership team and a tempting conclusion follows: our people are using AI, so our AI strategy must be working. It does not follow. An organization can have very high usage, a dozen overlapping tools, unknown data flows and no measured change in any result that matters. High usage is evidence of interest. It is not evidence of strategy, and it is not yet evidence of value.

The core idea

Strategy and adoption answer different questions, and confusing them is a common and costly mistake in enterprise AI. Strategy is choices. It decides why AI matters to the organization, where it will be applied, what comes first, who owns the decisions and how much will be invested. As What Is an Enterprise AI Strategy? showed, its value lies as much in what it refuses as in what it funds. Adoption is behavior. It is the degree to which people, teams and processes actually change how they work using AI.

Each failure has a recognizable look. The plan on a shelf has clear priorities, a funded portfolio and well-built products, and very little change in how work gets done. The strategy may be sound; execution never converted it into behavior. Fragmented activity is the reverse. Employees experiment enthusiastically, teams buy their own tools, several vendors hold the same kind of data, and spending is duplicated. There is plenty of activity and no way to add it up. The survey numbers above suggest that many organizations sit in that corner today.

Richard Rumelt’s warning about bad strategy applies directly. One of its hallmarks, he argues, is mistaking goals for strategy³. “Every employee using AI by year end” is a goal. It may be a reasonable one, but it says nothing about which work should change, why, or what the organization will stop doing to make room.

A chain from choice to result

The two ideas meet in a chain. The previous chapter traced every AI initiative upward to the business strategy. This chapter follows the same line downward, past the strategy document and into the work.

Strategy funds capability: a small portfolio of products and tools chosen for the priorities. Capability is only potential until the target people use it in their real work. Use matters when the workflow itself changes, meaning the steps, the hand-offs and who checks what. And a changed workflow is what produces an outcome a board will recognize. Michael Porter made the underlying point four decades ago: competitive advantage comes from the activities a firm performs, not from the tools it owns⁴. AI creates value where it changes an activity that matters.

The test of the chain is traceability. If no one can say which priority a tool serves, or which workflow a usage number belongs to, the chain is broken somewhere. The chain also runs backwards. Heavy use with no improvement suggests the use case was wrong or the outputs are weak. A valuable tool that is barely used points to friction, low trust or poor enablement. A strategy that listens to that evidence, and updates its priorities, is the management loop the previous chapter described.

Adoption is a ladder, not a login

People rarely jump from no AI to a changed way of working in one step. It helps to picture adoption as a ladder.

Awareness means people know which tools exist, which policies apply and where to get help. It is necessary and proves nothing. Experimentation covers drafting, summarizing and brainstorming; it is good for learning and rarely moves a business result. Regular use makes AI part of one person’s routine, and even here the right question is whether the routine has improved.

The rung that matters strategically is workflow integration. Picture an illustrative finance team closing the month. AI drafts the variance commentary from the ledger, the analyst checks it against what she knows about the business, and the controller signs it off. The team’s process has changed, not just one person’s habit, and that is what makes the next rung, a business outcome such as a faster close with fewer corrections, possible.

Many adoption dashboards stop at the third rung. That is understandable, because the lower rungs are easy to count. It is also why so many programs report success without being able to point at any changed result.

Usage is cheap; changed work is not

Read the evidence in this chapter rung by rung and one shape keeps appearing. The Microsoft survey counts the lower rungs: three in four knowledge workers use AI, which says nothing yet about any workflow¹. In Gallup’s data, only 16 percent of AI users strongly agree that the tools their organization provided are useful for their work²; that is a share of AI users, not of all employees. The physician group in this chapter’s story enabled its tool for 10,000 physicians and staff, and 968 physicians reached 100 uses or more in the first ten weeks⁵. Each rung up, the numbers shrink. Use spreads easily; changed work does not. A large Danish study found the same gap across a whole economy: widespread chatbot use, no measurable effect on earnings⁶.

That is why measurement has to climb the ladder with the program. Why AI use so often fails to become productivity is the subject of AI and Workforce Productivity; the point here is how to read adoption itself.

The weakest measure is reach: how many people have access and have switched the tool on. Better is usage among the people who are supposed to use it. Better still is the share of target workflows where AI is used well. The measure that matters in the end is the outcome: cycle time, quality, cost, revenue, or what customers experience.

Two warnings belong next to the stack. First, high adoption can be bad adoption. Very high usage alongside falling quality and rising rework is not success, so a usage number should always be read next to a quality measure; Leading vs Lagging AI Metrics develops these guardrails. Second, mandates. An instruction that every employee must use the assistant raises usage quickly. It does not, on its own, change a workflow, earn acceptance or create value.

Think of a town that wants more music in its life and gives a piano to every household. It can count pianos delivered and lids opened, and both numbers will look wonderful in the first month. But nobody plays better because a lid is open. Playing improves when someone chooses a piece to perform and takes lessons from a teacher who corrects their hands. In an AI program, licenses and logins are the pianos and the open lids. The chosen piece is the strategy, the lessons are workflow integration with managers as the teachers, and the recital is the outcome.

Adoption is a change problem

Why do so many programs stall at regular use? Because adoption is a change problem as much as a technology problem, and much of the change is invisible from the launch plan.

The survey evidence points below the waterline. In Gallup’s 2025 data, employees who strongly agreed that leadership had communicated a clear plan were 2.6 times as likely to feel comfortable using AI². In the Microsoft survey, only 39 percent of AI users had received AI training from their company, and 52 percent were reluctant to admit using AI for their most important tasks¹. People who hide their use cannot share what works, and a team cannot redesign a workflow around practices nobody admits to.

John Kotter’s study of transformation efforts named undercommunicating the vision as one of the classic errors, often by a factor of ten⁷. If only 22 percent of employees say a clear plan has been communicated, most of the organization has no plan to act on.

Incentives matter just as much. Suppose, as an illustration, an analyst saves two hours a week, and her manager fills those hours with more of the same work. She sees no benefit, so she quietly stops, or keeps using AI and stops mentioning it. Workload, recognition and role design have to change with the tool. This is why adoption cannot be handed entirely to IT. Executives set the reason and fund enablement. Managers make adoption real, team by team, because they decide what the work looks like and what happens to the time it frees. The aim is not to make people use AI. It is to use AI to improve how important work gets done.

Individual usage is not enterprise adoption

Leaders often blur one more distinction. Individual usage grows on its own: people find a tool, try it and keep it if it helps. That is the 78 percent bringing their own tools. It is real demand and useful learning, but nobody can see what data goes in, what quality comes out or what it costs.

Enterprise adoption is an organizational capability. It needs common tools, training, support, clear policy, workflow integration and measurement, and no individual builds those at a desk. This is the sense in which Marco Iansiti and Karim Lakhani describe AI as an operating-model change rather than a technology purchase: the gains come when the firm, not the individual, is rebuilt around it⁸.

Unapproved use deserves a calm response. A blanket ban feels safe, but the demand does not disappear; it goes out of sight. The better sequence is to understand what people need, offer a safe alternative that is genuinely useful, set clear policy and educate. The data risks are the subject of Privacy and Confidential Data. The opposite failure is just as common: an audit finds dozens of chat assistants, coding assistants and automation tools, bought team by team. That is high tool adoption and rising complexity. Strategy has to say which capabilities are standardized and where local choice is allowed, a question Centralized vs Federated AI takes up later in this module.

Story: a physician group that chose one burden

A well-documented example of strategy and adoption working together comes from health care. A large physician group that staffs the hospitals and clinics of an integrated health system in Northern California had a specific, named problem: documentation. Physicians were spending long hours on clinical notes, much of it outside clinic hours, in what clinicians call “pajama time”.

The group did not buy a general assistant for everyone and wait. It chose one workflow, the clinical note written after a patient visit, and one kind of tool, an ambient AI scribe that listens to the visit and drafts the note. The physician reviews, corrects and signs every note. In October 2023 the group enabled the tool for 10,000 physicians and staff, and supported the rollout with several training formats, at-the-elbow peer support, materials for patients and continuous monitoring⁵.

The early numbers show the ladder at work. Within ten weeks, 3,442 physicians had used the scribe in up to 303,266 patient encounters, and 968 had used it in 100 encounters or more⁵. Reach was broad; workflow integration was concentrated. The group measured the rungs that mattered. Physicians rated draft notes, joined feedback forums and scored note quality on a standard documentation instrument, giving an average of 4.35 out of 5. The weakest area was thoroughness, and some users reported missed details and trouble with conversations involving several speakers, which shaped guidance by specialty⁹.

By the end of 2024, 7,260 physicians had used the scribe in more than 2.5 million encounters. The group estimated 15,791 hours of documentation time saved, compared with non-users, the equivalent of 1,794 eight-hour working days. That is a modeled estimate, and the satisfaction figures that follow are survey responses, not measured outcomes. The most active third of users accounted for most of the use. Eighty-four percent of physicians reported a positive effect on their communication with patients and 82 percent said their work satisfaction had improved, while 47 percent of patients said their doctor spent less time looking at the computer¹⁰.

Every link in the chain is visible. The strategic choice was a named burden. The capability was one tool for one workflow. Adoption was supported by peers, not just announced. The workflow changed, with a clinician still signing every note. And the outcome was measured where the strategy said it would be, next to quality and patient experience. This is one organization’s experience, not a guarantee, and the clinical safety questions it raises belong to later modules. But it shows what adoption looks like when a strategy has decided what adoption is for.

What this means for leaders

Ask for the chain, not the count. A usage number is useful only when someone can say which priority it serves and which workflow it belongs to. Choose a few workflows and move them up the ladder deliberately, rather than spreading licenses and hoping. Give managers the job of adoption, with the authority to redesign the work and decide what happens to the time it frees. And treat the evidence from adoption as input to the strategy: heavy use without results and valuable tools without users are both telling you something.

Check yourself

1. If most employees use AI, the AI strategy is working.MythFact
2. Strategy and adoption are two names for the same thing.MythFact
3. Mandating AI use guarantees adoption.MythFact
4. In its first ten weeks, most physicians who tried the group’s AI scribe became heavy users.MythFact
5. High usage with falling quality is not successful adoption.MythFact
6. Adoption is mainly an IT responsibility.MythFact

Reflection: find your piano count

What comes next

The physician group’s strategy did not begin with AI. It began with a business problem, the documentation burden on its clinicians, and only then asked where AI could help. That order is the subject of the next chapter, Start With Business Strategy, which shows how business objectives, competitive priorities and operating realities should decide where AI fits.