AI Should Not Be the Process

I realized recently that I was using AI for something that, at first, felt completely reasonable. I had a work problem involving a process spread across several places. Some of the relevant information existed in previous conversations, some existed in notes and requirements, some depended on what different people understood about how the work was supposed to happen, and some only existed because I remembered why an earlier decision had been made. These were questions that had already been answered, questions that were still open, and pieces of information that made sense individually but had never quite been assembled into one coherent picture.

So I did what many of us increasingly do. I gave the pieces to an AI system and asked it to help me make sense of them. It worked remarkably well. The system helped me connect decisions made at different points in time, identify gaps between what was expected and what appeared to be happening, organize questions that still needed answers, and turn a collection of fragmented information into something I could actually reason about. Somewhere in that process, however, I began to feel uncomfortable, not because the AI was doing a poor job, but because it was doing a very good job at something I was no longer sure it should have needed to do.

There is a difference between improving a process and reconstructing one

A great deal of discussion around workplace AI concerns integration. The question is increasingly not whether employees will use generative AI, but how organizations can move from isolated experimentation toward repeatable, governed workflows. One recent framework, for example, describes a progression from exploration to codification, integration, and eventually institutionalized AI-enabled practices. The broader argument is that organizations create greater value when useful AI practices move beyond individual experimentation and become embedded in recurring workflows and organizational knowledge.¹

That seems reasonable to me, but I think there is another condition worth examining. What happens when AI is not being inserted into a coherent process at all, but is instead helping create the coherence? There is an important difference between asking an AI system to examine a documented workflow and asking it to reconstruct that workflow from fragments. There is a difference between using AI to identify an inefficiency in an existing process and using AI because nobody can easily see the process end to end. There is a difference between making a handoff more efficient and using AI to recover enough historical context to understand what the handoff was supposed to contain.

The outputs may look similar. In both cases, someone is using AI productively, information that was difficult to organize becomes easier to reason about, and work may move faster as a result. But the organizational condition underneath that use is materially different. One situation involves AI augmenting a process that already exists. The other involves AI compensating for parts of the process that do not exist coherently enough without it.

AI can become a kind of compensatory infrastructure

Research on shadow AI already gives us part of the language for describing what happens when employees begin using AI outside formal organizational structures. One line of research examines the risks that emerge as employees use generative and agentic AI tools because they perceive them as useful and efficient, including disclosure risk, governance problems, knowledge dilution, and organizational feedback loops.² Related work describes shadow user innovation, in which employees use readily concealable technologies such as generative AI to create value outside formal managerial or IT oversight. The authors argue that these hidden practices can generate local productivity and learning gains while also creating organizational blind spots when the resulting knowledge never becomes visible enough to be incorporated into broader routines.³

Those concerns matter, but I think there is another possibility inside the same phenomenon. AI can become a kind of compensatory infrastructure. It does not have to be an official system, a formal workflow, or even something the organization recognizes as infrastructure. It can simply become the thing that allows a person to connect knowledge the organization has failed to connect for them.

A missing explanation can sometimes be reconstructed from several conversations. A weak handoff can be pieced back together from historical context. Scattered requirements can be synthesized into something operational. A process that is difficult to see can become legible enough for someone to continue working. None of those uses is inherently irresponsible. In many cases, they are exactly what a competent employee should do when confronted with incomplete information. The harder question is what happens when that temporary cognitive aid becomes the mechanism through which organizational continuity is achieved.

Individual capability is not organizational capability

This distinction becomes especially important when we talk about AI adoption as though greater individual capability automatically means greater organizational capability. Research on collaborative AI adoption has found that employees can make substantial individual use of generative AI while those practices remain disconnected from shared team routines. Individual adoption and collective integration are not necessarily the same thing, and organizational structures can determine whether personal experimentation ever becomes part of collaborative work.⁴

That distinction seems applicable well beyond any one industry. If one employee can use AI to reconstruct enough information to complete a task, it does not necessarily follow that the organization possesses that knowledge in a durable form. If one employee can bridge a weak handoff, the handoff itself may still be weak. If one person can synthesize several scattered sources into a usable representation of the process, the underlying process may still be fragmented.

The work getting done and the process being healthy are not the same proposition. That sounds obvious when stated directly, but it becomes much harder to see when the outcome looks good. A deadline gets met, a coherent deliverable appears, the missing context is recovered, and the next person receives something usable. From the outside, all of that can look like organizational success. Yet some of the apparent functionality may actually be coming from an invisible layer created by a person using AI to compensate for structural gaps around them.

AI can make a weak process easier to survive

One of the most impressive things about generative AI is its ability to absorb friction. Information can be scattered, inconsistently written, created at different times, stored in different formats, or understood differently by different people, and a capable model can often help someone bring those pieces into a common frame remarkably quickly. That is a real capability, and it is precisely why the distinction matters.

Imagine two organizations producing essentially the same result. In the first, the relevant process is documented, roles are understood, decisions are recorded, and people know where to find the information required to perform the work. AI is used to examine that system, identify patterns, explain anomalies, and suggest improvements. In the second, much of that structure is incomplete. A knowledgeable employee uses AI to recover enough of the missing continuity that the work can still happen. The visible result may be nearly identical, but the organizational meaning is not. In the first case, AI may genuinely be improving a functioning process. In the second, AI may simply be making the cost of a weak process easier to absorb. Those are different outcomes, and I suspect organizations increasingly need to learn how to tell them apart.

This is where AI adoption can become deceptive. Productivity can improve without the underlying process improving with it. A capable employee using a capable AI system can absorb missing context, inconsistent documentation, poor handoffs, and organizational ambiguity well enough that the dysfunction becomes less visible. The process may appear healthier because someone has become unusually good at surviving it.

This is not an argument against using AI this way

There would be something slightly absurd about arguing that someone confronted with fragmented information should refuse to use a tool capable of helping them understand it. I will continue using AI this way because the capability is too useful not to. The more important question is what we infer from the fact that it was necessary.

If I use AI to reconstruct context that should have been captured elsewhere, perhaps the lesson is not simply that AI made me more productive. Perhaps the interaction also produced a signal about the process. Something was difficult to see. Something depended too heavily on memory. Something was distributed across people rather than represented clearly. Something about the documentation, ownership, workflow, or handoff required reconstruction before the work could continue.

The AI may help solve the immediate problem while simultaneously revealing another one. That second problem should not disappear merely because the first one was resolved. In fact, AI may be particularly useful if we treat the need for that reconstruction as evidence rather than allowing the successful output to erase the conditions that made the reconstruction necessary.

AI should probably sit above the process

Research on organizational AI governance increasingly treats AI as part of a larger sociotechnical system rather than as an isolated model. One recent approach, for example, describes AI governance in terms of policy, technical controls, workflow mechanisms, monitoring, and escalation, explicitly locating AI inside an organizational system of norms and accountability.⁵ I think that framing becomes particularly useful here because it asks us to think about where AI belongs relative to the process rather than merely whether the technology can perform a task.

The most interesting organizational role for AI may not be to become the missing connective tissue inside every imperfect workflow. It may be to help us see those workflows more clearly. AI can compare what people say a process does with what available evidence suggests actually happens. It can surface inconsistencies, reveal information concentrated in one person, make hidden dependencies easier to see, and help examine why the same handoff repeatedly fails or why the same decision has to be reconstructed every few months.

In other words, AI can be extraordinarily useful as a layer over process. That is different from quietly becoming the layer that allows the process to exist. This distinction may seem subtle until we begin measuring AI adoption primarily by outputs. If work gets done faster, we may reasonably conclude that AI improved productivity. But perhaps another question belongs beside that metric: what exactly did the AI improve? Did it augment a healthy process, expose a problem that can now be fixed, or simply make an unhealthy process easier for one person to navigate?

I started with a practical problem and a tool that helped me solve it. I ended with a question I did not expect to be asking. If I am using AI to weave together a process that does not otherwise exist coherently, the important question may not be how impressive the AI is, but why I needed it to do that in the first place. AI should help us understand the system, not become it.