The Detail Gap: Why Your AI Strategy Stalls in the Middle

Why AI adoption stalls in the middle of every organization

There is a pattern I see in organizations that have fully committed to AI transformation. The people at the top are talking about it constantly. The people at the very bottom ,  the curious ones, the early adopters, the engineers who stay late reading model documentation,  are using it every day, building things in an afternoon that would have taken a week before. And then there is everyone else. The vast middle. Showing up, doing their jobs, mildly aware that something is happening, not sure what it means for them, and not, in any practical sense, changing how they work.

This is what happens when an organization underestimates how much detail is involved in actually doing a job.

John Salvatier wrote a short essay in 2017 called “Reality Has a Surprising Amount of Detail.” His central observation was deceptively plain: when you try to do something for the first time — build a deck, make bread, write a piece of software — you discover it is much harder than it appeared from the outside. Not harder in the sense of requiring special talent. Harder in the sense of containing layers of specificity you never knew existed. The wood warps in ways you didn’t anticipate. The flour absorbs water at a rate that changes based on humidity. The codebase has a deployment constraint nobody thought to document. Expertise, Salvatier argued, is largely the accumulated internalization of detail. The gap between an expert and a novice is not mostly about intelligence. It is mostly about how much detail the expert has lived through.

This is also, I want to argue, the gap between your AI super-users and everyone else in your company. And until organizations understand it as such, their transformation efforts will keep producing the same result: great demos, pilot programs, a handful of power users, and a middle layer that remains essentially unchanged.

The Work Contains More Detail Than the Tool Anticipates

Ask a senior marketing manager to use AI to draft a campaign brief and she will encounter, in rapid succession, every piece of institutional knowledge she has spent years absorbing: the particular sensitivities of this client, the unwritten approval chain, the brand voice that lives nowhere in any document but that her VP will immediately flag if it is off, the competitive context that changed last quarter, the way certain words read in certain markets. None of this is in the prompt. None of it is in the model. The model returns something that looks like a brief and reads like no brief she would actually use. So she sets it aside. She is not resistant to technology. But the tool did not know what she knows, and bridging that gap took more effort than just writing the brief herself.

This is the fundamental dynamic that AI adoption frameworks miss. They measure adoption as usage:  did the employee open the tool, did they generate output. They do not measure the quality of the output relative to what the employee actually needed, or the cognitive overhead of getting there.

A 2023 field experiment run by Harvard Business School in collaboration with Boston Consulting Group put this dynamic into sharp empirical relief. The researchers ran a pre-registered randomized experiment with 758 BCG consultants — about 7% of the firm’s individual contributor workforce — and measured performance across realistic, complex consulting tasks with and without access to GPT-4. For tasks inside what the researchers called the “jagged technological frontier,” the results were striking: consultants using AI completed 12.2% more tasks, worked 25% faster, and produced output rated 40% higher in quality. But for a task specifically designed to fall outside AI’s current capabilities, consultants using the tool performed 19 percentage points worse than those working without it. The AI didn’t just fail to help. It actively degraded performance, because people trusted its output in a domain where its output was wrong.

The implication is that using AI requires specificity that most organizations have not done the work to develop. The super-users figured it out because they had the combination of domain expertise, self-reflection, the ability to articulate processes and flows, and intellectual curiosity to build their own bridges. They wrote their own system prompts. They learned to decompose their tasks. They iterated until the model output matched the standard their work actually required. That iteration is itself a skill, and it is a skill built on top of deep knowledge of the domain. You cannot prompt your way to a good deliverable in a domain you do not understand. The highest-value AI users are, almost by definition, your most experienced people  using the tool to accelerate, redesign and redefine work they already knew how to do well.

The Adoption Problem Is a Translation Problem

In 1962, Everett Rogers published Diffusion of Innovations, mapping how new ideas spread through populations. He identified five adopter categories — innovators, early adopters, early majority, late majority, laggards — and described how the psychology of each group shapes what it takes to reach them. Three decades later, Geoffrey Moore built on Rogers’ framework in Crossing the Chasm, identifying a structural gap between the early adopters and the early majority that many technologies never successfully bridge. The chasm is not primarily a gap in attitude. It is a gap in what each group needs to see before they commit.

Early adopters are comfortable with abstraction. They can see the potential of a technology independent of a fully worked use case. The early majority, Moore’s pragmatists, need to see how the thing works in their specific situation, for their specific job, producing their specific outputs. They are more rational than resistant. They are waiting for evidence that this tool actually does what they need it to do, in the way they need it done.

Most AI rollouts fail this test. They provide general training . Here is how to write a prompt, here is what a large language model is, here is a demo of someone writing a poem, and then release employees into an environment where they are expected to self-discover the relevant applications. This works for people who enjoy self-directed exploration of new tools. It does not work for people who are already managing a full workload and cannot afford to experiment their way to productivity. The organizational cost of that gap compounds quietly. The middle of the organization keeps working the way it always has. The efficiency gains accrue to a small group of self-starters. The board gets a report about AI adoption that counts licenses rather than outcomes.

There is also a social dimension that rarely appears in the adoption literature. Knowledge workers have built their professional identities around what they know how to do. A tool that implies they might do it faster, or differently, carries an implicit challenge to that identity. Stanford psychologist Claude Steele spent decades documenting how threats to a domain central to one’s self-concept trigger defensive responses rather than curiosity — a dynamic he called identity threat. The employees who were not already curious about AI are not going to become curious because their company bought licenses. They need to see someone they respect, doing work they recognize, producing results that are clearly better. They need the translation done for them, in context, before they can evaluate whether it is worth their while.

What Super-Users Actually Know

Spend time with the super-users in any organization and what becomes clear is that their advantage is not primarily technical. They are not using capabilities that others lack access to. They are doing something more fundamental: they have developed a mental model of what the tool can and cannot do, and they have aligned that model with a detailed understanding of their own work. They know which parts of their job involve judgment that the model cannot replicate. They know which parts are essentially synthesis and formatting that the model can accelerate by an order of magnitude. They have learned, through iteration, exactly how to specify what they need so that the output is usable.

This is what Salvatier would call the accumulated detail of working with the tool. It is not something that can be transmitted in a training session. It has to be lived. And the time required to live it is time that most employees, under normal workload conditions, are not going to invest speculatively. The return on that investment only becomes visible after you have made it.

Which means the organizations cannot close the adoption gap by running more trainings. They must create conditions under which more people can afford to invest the time. Microsoft’s own internal work rolling out Copilot to more than 300,000 employees consistently points to the same finding: designated experimentation space, manager permission, and peer-led learning — not general training content — are what actually move employees from awareness to fluency. Not the most tech-forward employees. Not the youngest. The ones whose managers made it psychologically and practically safe to spend time on the learning curve.

The Organizational Implication

The standard response to uneven AI adoption is more enablement: more training, more content, more internal communications about why AI matters. This treats adoption as an information deficit, and the evidence suggests it is not. The employees who are not using AI meaningfully know AI exists. They have heard the boardroom language. They have sat through the all-hands. They are not waiting for more information about the general case. They are waiting for a specific case — their case — to be worked out well enough that the tool is actually useful to them without requiring an unreasonable investment of time and cognitive energy.

This means the highest-leverage intervention is not broad. It is narrow and deep. It is taking the workflow that a specific team does repeatedly  ( the thing they do ten times a week that follows the same basic structure)  and working out, in granular detail, how AI fits into that workflow. What the prompt should contain. What context it needs. What parts of the output require human review and revision. What the output should look like when it is right. This is not work that generalist trainers can do. It requires someone who both understands the tool and understands the work. In most organizations, that person is one of the super-users. The most valuable thing you can do with your super-users is to deploy them as translators.

Salvatier ends his essay with a quiet observation: before concluding something is broken, it is worth asking what detail you might be missing. The companies that look at their AI adoption numbers and see failure are often missing exactly this. The technology is not the constraint. The organizational capacity to translate general capability into specific, contextual, role-level usefulness — that is the constraint. And it requires the same patient accumulation of detail that expertise in any domain requires. There are no shortcuts. There is only the work of actually understanding what your people do and building the bridges, one workflow at a time.

The gap between your super-users and everyone else is not a technology gap. It is a detail gap. The good news is that detail is learnable. The bad news is that it takes time, and attention, and the willingness to get specific in ways that most transformation programs are too impatient to sustain.

That impatience is the real adoption problem.

— Liat Ben-Zur

Sources:

John Salvatier, “Reality Has a Surprising Amount of Detail” (2017);

Dell’Acqua, McFowland, Mollick et al., “Navigating the Jagged Technological Frontier,” Harvard Business School Working Paper No. 24-013 (2023);

Everett M. Rogers, Diffusion of Innovations (1962);

Geoffrey A. Moore, Crossing the Chasm (1991);

Claude M. Steele, Whistling Vivaldi (2010); Microsoft Digital, Copilot deployment research (2024–2025).