Part 2 of 3 in The Language of AI Progress
I ended the last piece by saying I had broken the broad idea of “AI adoption” into four milestones. I got there after spending a lot of time trying to answer what sounds like a fairly simple question. If “adoption” can describe everything from getting access to a tool to materially changing how work gets produced, what would make the differences between those conditions easier to see?
I kept coming back to two questions about the work itself. Has the way the work gets done changed? Has the work being produced changed with it?
I like those questions because they pull the conversation away from the technology for a minute. They don’t ask which model someone uses, how many prompts they write, whether they attended a training session, or how enthusiastic they feel about AI. All of those things can be useful to know, but none of them tells me much on its own about what has actually changed in the work. I’m more interested in whether the process someone relies on today looks meaningfully different from the one they relied on before, and whether that change has altered the kind or scope of work they can realistically produce.
Once I started looking at progress through that lens, some of the conditions we casually group together as “adoption” became much easier to separate. Someone can use AI regularly while leaving most of their underlying way of working intact. Someone else can still be producing familiar work, but through a process that would have looked completely foreign to them a year earlier. Go far enough and even the work itself can start to change, because new capabilities make previously unrealistic tasks, outputs, or problems suddenly manageable.
I use four milestones to describe those differences. I don’t think there are objectively four stages of AI progress hiding in the universe waiting for someone to name them, and I’m not especially attached to the labels themselves. Three stages, six stages, different names entirely. Fine. What I care about is whether the distinctions are useful enough to help us describe where someone is, recognize when something meaningful has changed, and have a more precise conversation about where they may be trying to go next.
Access comes first
Before those two questions become useful, one more condition has to be true. Someone needs access to AI in the first place. I treat access as its own milestone because there is quite a difference between a tool being unavailable and it suddenly sitting in front of you, approved, provisioned, and ready to use. For a lot of organizations, crossing that threshold used to require a deliberate rollout. Now access can spread across hundreds or thousands of employees remarkably quickly, sometimes simply because AI has been added to software people were already using.
That speed makes access one of the easiest forms of progress to see, which probably explains why it is also one of the easiest to measure. How many licenses have we assigned? What percentage of employees have signed in? How many people activated the tool at least once? Those are clean numbers, and I understand the appeal. If the question is whether AI has been made available across an organization, they can tell us quite a lot.
I become much less comfortable when we start treating those same numbers as evidence of adoption. A first sign-in tells me that someone crossed the front door. It does not tell me whether they came back the next day, found anything useful once they were inside, or changed a single thing about how they work. Even widespread access can coexist with very little meaningful change in the work being produced.
Access still belongs in the picture. We just have to be precise about what it tells us. It establishes the opportunity for everything that follows. Once someone has the tool in front of them, we can finally start asking the questions I care more about. Are they actually using it? Has the way they work begun to change? Has the work itself changed with it?
Two questions, three very different conditions
Once access is established, those two questions start doing most of the useful work. I’ve found that the answers create three conditions that are meaningfully different from one another, even though we might casually describe all three as someone “using AI.”
The first question is about the production process. Someone can add AI to an existing workflow without really changing the workflow itself, or they can begin producing familiar work through a process that now depends on AI in a much more fundamental way. The second question looks at what comes out the other side. Are they still producing the same general kind and scope of work, or has the change in how they work expanded what they can realistically produce?
Put those two dimensions together, and the framework becomes fairly simple.
I use same and new pretty deliberately here, but not literally. “Same work” does not mean the output is identical to what someone produced before AI showed up. It means the work still falls within the general range of what they already knew how to produce and were reasonably capable of producing. “New work” begins when AI expands that range enough that something previously impractical, inaccessible, or outside the old production system becomes realistic.
The table is useful to me because it forces us to look past the fact that AI appears somewhere in the process. All three conditions involve AI. The more interesting question is what changed after it arrived. Sometimes the answer is very little. Sometimes the process changes while the work remains familiar. And sometimes the change reaches far enough that the boundaries of the work begin to move, too.
Those differences are easiest to see one milestone at a time.
Bring the work to AI
I think of usage as bringing your work to AI, and I mean that quite literally. You already have a way of getting something done, and somewhere along the way you hand a piece of it to the model. Maybe you bring a question into a chatbot, paste in a draft and ask for help improving it, summarize a document you would otherwise have read yourself, or ask AI to troubleshoot a problem before you go back to whatever you were doing. The interaction can be useful, frequent, and even sophisticated. The larger production system still looks mostly the way it did before AI arrived.
That is why usage gets same under “way of working” in the table. The person may have added a new tool, but the underlying path from problem to finished work remains recognizable. A marketer still opens the document, writes the campaign, and moves it through the same process, even if ChatGPT now helps with the first draft. An analyst still works through the same spreadsheet and delivers the same analysis, even if AI helps explain a formula or clean up a paragraph along the way. A manager may use AI every morning to summarize notes or pressure-test an idea, then carry the answer back into the workflow they already had.
The second half of the row matters just as much. Usage also gets same under “work being produced.” I’m using “same” in the sense we established earlier. The output still sits inside the person’s existing range of work. They could already write the campaign, analyze the spreadsheet, prepare the meeting notes, or solve the problem. AI may help them get there faster, reduce some friction, improve the quality, or save them from staring at a blank page for twenty minutes, but the work itself remains familiar.
None of this makes usage trivial. A person can get enormous value from AI while staying here, and regular usage can represent a substantial change from having access and doing nothing with it. I only want to be careful about what we infer from it. Someone can use AI dozens of times a day and still organize their work around essentially the same production system they used before.
The easiest way I’ve found to recognize usage is to watch the direction of travel. The work keeps leaving its existing environment, visiting AI for some help, and then coming back. The next milestone starts to look different because AI begins moving in the other direction.
Bring AI into the work
I think of incorporation as bringing AI into your work. The goal may still be familiar, but the path to getting there starts to change. Instead of repeatedly carrying individual questions or pieces of a task out to a chatbot, AI becomes part of the environment, process, or system through which the work gets produced. At some point, removing AI would mean changing the workflow itself rather than simply losing a helpful tool along the way.
That is why incorporation gets new under “way of working.” The change can be technical, but it does not have to be. A developer working with AI directly inside a terminal is an easy example because the shift is visible. The same idea can show up in much less technical work. Someone might use AI to build a simple local HTML tool that replaces a repetitive manual process, create a repeatable workflow for turning raw information into a finished deliverable, or reorganize a recurring task so that AI handles parts of the production process that used to happen manually. The important part is the change in how the work moves from beginning to end.
The work being produced, though, can still be the same. The developer may still be building the kinds of features they were already capable of building. The marketer may still be producing the same campaign brief. The analyst may still be delivering the same report. Incorporation does not require someone to suddenly become capable of entirely new categories of work. It describes a different production system for work that remains within their existing range.
Usage and incorporation can look surprisingly similar from the outside. Both people are using AI. Both may use it every day. Both may save significant amounts of time. But one person is repeatedly taking pieces of an existing workflow to AI, while the other has started rebuilding the workflow around what AI can do. Frequency alone will not tell you which one you are looking at.
For me, the direction of travel is useful again. With usage, the work goes out to AI and comes back. With incorporation, AI begins moving into the work itself. The output may still look familiar on the other side, but the machinery producing it has started to change.
Expand the work within reach
I think of leverage as expanding the work within reach. It carries forward the new way of working from incorporation, but the second half of the table changes, too. The work being produced starts to move beyond what someone could realistically have created through their old production system. I use new fairly broadly here. The work does not need to be novel to the world, or even especially sophisticated. It only needs to sit outside that person’s previous practical range.
Some of the clearest examples show up when AI lowers a barrier that used to keep someone out of a particular part of the production process. A marketer who could describe the internal tool they wanted but had no realistic path to building it can now produce a functioning version themselves. Someone comfortable analyzing data might create an interface that lets other people interact with it. A person who has always worked in documents and spreadsheets might suddenly find themselves producing small applications, prototypes, automations, or other artifacts that previously would have required a different skill set altogether. The idea may have existed before. The ability to carry it far enough to become real often did not.
Other work used to happen only by crossing an organizational boundary. You had an idea, wrote a brief, handed it to someone with the necessary technical or creative skills, and waited for their portion of the production process to begin on their clock. That division of labor still has plenty of value, but AI can let people participate much farther into work that previously would have been handed off almost immediately. They can explore the idea, learn enough to make decisions they could not previously make, and produce a working first version themselves. If the work still needs to be handed off, they can arrive with more than a description of what they hope someone else can build.
Volume alone does not get us here. If a new AI-enabled workflow lets someone produce ten reports in the time they once produced five, that can be a significant improvement while the work itself remains familiar. I would still describe that as incorporation if the person is producing the same general kind of work they were already capable of producing. Leverage begins when the boundary of what they can realistically produce starts to move.
None of this means AI instantly gives someone the expertise that traditionally comes with years of working in a domain. It can lower the barrier to participating in that work much faster than it can transfer the judgment and context that come with experience.
Leverage is the point where that expanded range becomes visible in the work itself. The production system has changed, and the person can now take part in creating things that would previously have been out of reach, abandoned before they began, or handed to someone else almost immediately.
Usage helps you with work you already do. Incorporation changes how you do it. Leverage starts changing what you can realistically take part in producing.
A way to describe progress, not grade people
I want to be clear about what I think these milestones are describing. They are a way of looking at the relationship between AI and a particular piece of work, not a score assigned to the person doing it. Someone can be at usage for one task, incorporation for another, and leverage somewhere else entirely. The framework gets less useful the moment we start treating one of those labels as a permanent identity.
The same artifact can also mean different things depending on who produced it and what their previous production system looked like. A developer using AI to code a small local HTML app may be incorporating AI into work they already knew how to do. A marketer who had never been able to build software at all might create something very similar and be operating much closer to leverage. The output alone cannot tell us where the milestone sits. We have to know what changed in the way the person worked and whether the work itself moved beyond what had previously been within reach.
Organizations will look even messier. A company can have employees who technically have access but rarely use AI, people who bring questions to a chatbot throughout the day, others who have rebuilt meaningful parts of their workflow around it, and a smaller group already producing work they could not realistically have produced before. All four conditions can exist inside the same organization at the same time, and even inside the same team. A single “adoption” percentage is going to flatten a lot of that variation.
I also would not assume that everyone moves neatly through the milestones in order, once, and never moves backward. New tools appear. Roles change. Some kinds of work lend themselves to incorporation faster than others. Someone may develop a highly integrated way of working in one area while continuing to use AI very casually somewhere else. The value of the framework comes from being able to describe those conditions with more precision, then notice when and where the condition changes.
The transitions become visible, too
Once the milestones are separated, I find myself paying much more attention to the movement between them. Instead of asking whether someone has “adopted AI,” we can ask what kind of change has actually occurred. Did access turn into regular use? Did regular use begin changing the production system? Did that new way of working expand the range of work someone could realistically take on? Those are different kinds of progress, and grouping them together makes it harder to see where movement is happening and where it is not.
I also don’t think we should assume that every transition behaves the same way. Some seem relatively easy to encourage once the right conditions are in place. Others ask more of the person making the change, even when the technology is available, useful, and already part of their routine. Once I started looking at AI progress this way, one transition in particular kept standing out.
That transition is where I want to go next.