The Gap Between Using AI and Changing How We Work

Part 3 of 3 in The Language of AI Progress

In the last article, I proposed four milestones for describing AI progress: access, usage, incorporation, and leverage. Having a vocabulary for these different conditions makes it easier to notice a tension in how AI is showing up at work. Usage continues to grow, and many employees are reporting benefits. In Gallup’s 2026 workplace research, 65% of employees at organizations that had implemented AI said it improved their productivity and efficiency, yet only about one in ten strongly agreed that AI had transformed how work gets done across their organization. Those questions measure different things, and organization-wide transformation is a much higher bar than changing a particular workflow. The contrast makes me curious about how often people are finding value in AI while continuing to produce familiar work much as they always have.

I don’t want to discount the value people are already getting from AI. Usage can make familiar tasks faster, easier, and better, sometimes considerably. But we’re talking about technology capable of much more than improving individual steps in an otherwise unchanged process. Someone who prepares a recurring report might already use AI to analyze an unusual result or draft a summary, then return to assembling everything manually. There may be an opportunity to reconsider how the information is gathered, evaluated, and prepared for review, improving the entire process. I suspect most knowledge workers have at least some familiar work that could benefit from that kind of reconsideration. The opportunity won’t always justify the effort, but given what AI has made possible, I would expect more of those opportunities to be finding their way into how we work.

My own experience is partly responsible for that expectation. I found generative AI surprisingly easy to start using. I recognized its utility, became comfortable working with it, and found plenty of ways to make my existing work easier. By most conventional interpretations, I was “adopting” AI quite successfully. Yet it took much longer for the technology to meaningfully change how I produced familiar work. I was benefiting from AI regularly while relying on production processes that looked remarkably similar to the ones I’d used before. I don’t consider that time wasted, but looking back, I can see opportunities to work differently that I hadn’t yet pursued. What interests me now is why so much time passed between recognizing the technology’s value and changing how I actually worked.

What “slow adopter” can’t explain

Our usual understanding of technology adoption struggles to account for what’s happening here. When someone takes a long time to embrace a new technology, we have a well-established explanation for that timing. Some people are early adopters, others need more convincing, and the rest may take years before something earns a place in their routine. I’ve always found that explanation plausible, but it becomes less satisfying when we’re talking about people who have already embraced AI. They’re using it. They’re getting impressive results and finding increasingly sophisticated applications for it. Calling them slow adopters doesn’t tell us much about why their underlying ways of working remain largely unchanged. In fact, by most measures of adoption, they may be progressing quite nicely.

Deloitte’s 2026 State of AI in the Enterprise makes the distinction between gaining productivity and changing processes more explicit. Drawing on a survey of business and IT leaders, it categorized 37% of organizations as using AI at a surface level, with little or no change to existing processes, even while capturing productivity and efficiency gains. Where Gallup asked employees about transformation across their organizations, Deloitte examined the extent to which organizations reported changing their processes.

Separating usage from incorporation helps expose something peculiar about the time we spend with AI. Someone might spend six months becoming exceptionally proficient at prompting, learning new models, experimenting with features, and finding useful applications for the technology. They could emerge from those six months substantially more capable than when they started, with measurable improvements in the quality or speed of their work. But those accomplishments may leave the production process itself unchanged. The milestones we established in the previous article describe different kinds of progress, and greater proficiency in usage doesn’t guarantee movement toward incorporation. We can keep getting better at using AI while spending months or even years producing familiar work through essentially the same processes.

I refer to the period between those two conditions as the AI Assimilation Gap.¹ It’s the period when someone is already using AI but has yet to meaningfully change how they produce familiar work. Usage describes a condition, while the Gap draws our attention to how long that condition can persist, even as proficiency and productivity continue to improve. Where worthwhile opportunities for incorporation exist, the period separating usage from incorporation merits attention. Naming the Gap doesn’t explain why it lasts so long, but it gives us something more precise to investigate. Before we start looking for explanations, though, we ought to establish whether we’re even trying to reach the same destination.

Once the expectations are clear

Maybe we haven’t been particularly clear about what we’re asking employees to accomplish in the first place. Organizations have spent the last few years encouraging team members to embrace AI, experiment with new tools, develop their skills, and find ways to become more productive. Those requests make sense, but they’re also broad enough to accommodate wildly different interpretations of success. Someone might hear “we want everyone adopting AI” and conclude that using ChatGPT throughout the day, attending training sessions, and finding useful applications for the technology is what the organization had in mind. And perhaps it was. Without agreeing on what success actually looks like, it’s difficult to know whether anyone is falling short of expectations.

Imagine an employee who has enthusiastically embraced AI. They’re using it regularly, saving several hours a week, producing better work, and sharing useful techniques with colleagues. Ask whether they’ve adopted AI and they’ll tell you they absolutely have, with plenty of evidence to support the claim. Their manager, meanwhile, is wondering why the team is still producing the same reports, following the same processes, and relying on the same manual activities it did a year ago. The manager expected the technology to change how the work gets done, while the employee understood the assignment as becoming an effective AI user. I can sympathize with both perspectives. The employee has made legitimate progress, and the manager may have identified opportunities for more substantial improvements. They’re evaluating different outcomes under the same broad expectation, which makes it easy for both to feel justified in their assessment.

The milestones from the previous article can help distinguish those expectations. We can stop treating adoption as a single condition that applies to someone’s entire job and look at the work itself. An employee might have rebuilt how they produce one recurring deliverable around AI, still bring individual questions to ChatGPT while working on another, and have little reason to change a third. Each piece of work presents a different opportunity, so the conversation becomes much more useful when we choose a particular workflow and agree on what improvement we’re actually pursuing. Perhaps the existing approach is delivering everything the organization needs, in which case continued usage may be a perfectly acceptable outcome. Or perhaps there’s an opportunity to reconsider the production process itself, with benefits significant enough to justify the effort. In that case, the manager and employee can agree that incorporation is the intended objective, rather than continuing to talk past one another about adoption. They can then evaluate progress against the change they’ve agreed to pursue.

Imagine that conversation goes as we’d hope. The manager and employee identify a worthwhile opportunity to change how a recurring report gets produced, agree that incorporation is their objective, and understand what success would look like. They now share an expectation that was previously open to interpretation, and perhaps that agreement is enough to get the change moving. But the conversation hasn’t redesigned the report or established a different way of producing it. If the work continues to happen as it did before, we can no longer explain the delay as a misunderstanding.

Poor communication can absolutely contribute to disappointing outcomes, and organizations have a responsibility to explain what they’re asking people to accomplish. But there’s a limit to what clarity can achieve on its own. Agreeing that a different way of working would be beneficial doesn’t mean someone knows how to create it. Once the destination is understood, the next question is whether we’ve given people the knowledge and skills required to get there.

Even when training works

Once we’ve agreed that incorporation is a worthwhile objective, the next explanation seems fairly obvious. Maybe people simply don’t know how to get there. AI is evolving at an extraordinary pace, and keeping up with its capabilities has practically become a job of its own. New models, tools, features, and techniques arrive faster than most of us can evaluate them, much less become proficient with them. I can understand why organizations respond by investing in training. If employees aren’t changing how they work because they don’t understand what the technology can do, then helping them develop that understanding seems like a sensible place to start.

And sometimes, I suspect it’s what they need. Someone who has only experimented with basic prompting may have no idea that AI could help orchestrate a recurring process, work directly with existing files, or create a simple application that replaces something they’ve been doing manually for years. A well-designed training program can expose those possibilities, demonstrate relevant techniques, and help employees build enough confidence and competence to experiment on their own. That knowledge can help them recognize opportunities they might otherwise overlook. For employees whose primary obstacle is limited exposure to the technology, training may be enough to get them moving toward incorporation.

One difficulty is that training can meet all its objectives while leaving open the question we’re investigating. Someone might finish a program substantially better at prompting, working with files, and applying AI to different parts of a recurring report. They can demonstrate new abilities, use them successfully, and produce better results. Those gains can happen within an otherwise familiar production process. The employee has become more capable of using AI, which doesn’t tell us whether they’ve changed how the report itself gets produced.

Of course, a thoughtful training program might go further. Being comfortable with AI isn’t the same as knowing how to reconsider a familiar workflow, and there’s no reason training couldn’t address that distinction. An employee could learn to evaluate dependencies, identify opportunities for redesign, and develop a workable alternative to their existing process. They might even demonstrate that alternative using the previous month’s data. At that point, we’ve addressed more than a lack of knowledge about the technology. The employee has the relevant skills and a plausible approach to producing the report differently. Whether that approach becomes the normal way the report is produced is still a separate question, one that successful training alone doesn’t necessarily answer.

We could keep investing in training, and it might help. But once someone can recognize a worthwhile opportunity and demonstrate a credible way to pursue it, insufficient knowledge becomes a less satisfying explanation for any continued delay. If incorporation remains elusive, we need to understand what stands between that demonstrated capability and a change in how the work gets produced.

The weight of what already works

We’ve considered two reasonable explanations for why incorporation might take longer than expected. Perhaps employees and organizations haven’t agreed on what they’re trying to accomplish, or perhaps employees lack the knowledge and skills needed to change how they work. Both problems are real, and addressing them can help people move toward incorporation. But notice what those explanations have in common. Each looks for something that might be missing, whether it’s a shared objective or a necessary skill. Neither directs much attention toward what’s already there, an established way of producing work that continues to meet expectations. I wonder whether we’ve underestimated the significance of asking someone to change that.

Most of us have spent years developing ways of working that reliably produce results. We know how to meet deadlines, anticipate problems, and deliver work we’re comfortable putting our names on. Incorporating AI can mean reconsidering processes that already work, often while we’re still expected to deliver the same work to the same standards. And the systems, expectations, and demands surrounding that work don’t necessarily change just because we’ve identified a potentially better way to produce it. The established process already fits those conditions, while a different approach may require experimentation, new approvals, or a temporary disruption to reliable results.

Which brings me back to the time we spend inside the AI Assimilation Gap. When someone recognizes a worthwhile opportunity, understands what needs to change, and has the skills to pursue it, what happens when that possibility encounters the processes and expectations already governing their work? Why might months or even years pass before a different approach becomes part of how that work gets produced? We now have a vocabulary that makes the period easier to recognize, but understanding what fills it remains a different challenge. I think answering those questions requires us to look more closely at the people doing the work, and the conditions in which we’re asking them to change it.