For the past two years, companies have been asking the wrong question: how do we use AI in our processes?
That question made sense at the beginning. When large language models first appeared, the instinct was natural: take what already exists, from workflows to functions, decision chains, etc., and try to accelerate them. Add copilots. Add assistants. Add automation layers. Improve productivity.
But as we’ve seen, that approach doesn’t scale. As I’ve argued in previous pieces, enterprise AI hasn’t failed because the technology doesn’t work. It has failed because we tried to place it in the wrong layer. Large language models were never designed to run a company, and embedding them into existing processes doesn’t change that structural mismatch.
Now that the initial enthusiasm has collided with reality, a different question is starting to emerge, quietly, but unmistakably: what if the problem is not how to use AI in our processes, but that our processes were never designed for AI in the first place?
The return of an old idea (this time for real)
In the 1990s, business process reengineering (BPR) promised something radical: redesign companies around information systems instead of layering technology on top of existing workflows. The idea was compelling, but the execution was uneven. Many initiatives became expensive reorganizations with limited long-term impact, partly because the underlying systems were still rigid, fragmented, and unable to adapt in real time.
This time is different.
Back then, systems were passive. They stored information, enforced rules, and supported decisions made by humans. Today, systems are becoming active: they can generate, evaluate, coordinate, and increasingly, act. That shift changes the equation entirely. It means we are no longer just digitizing processes: we are redefining what a process is.
McKinsey’s latest research on AI adoption reinforces this point: while usage is widespread, real impact correlates strongly with workflow redesign, not just tool deployment. Organizations that rethink how work is done, not just how it is assisted, are the ones seeing measurable gains.
In other words, the original promise of BPR is resurfacing, but now the technology can finally support it.
Why most processes are incompatible with AI
The uncomfortable truth is that most enterprise processes today are not just inefficient. They are structurally incompatible with the kind of systems AI is becoming.
They are:
- Fragmented: spread across tools, teams, and data silos
- Sequential: built around handoffs and delays
- Context-poor: dependent on individuals to reconstruct state
- Decision-latent: optimized for review, not action
- Human-centric by design: assuming that cognition, memory, and coordination are scarce
These characteristics made sense in a world where humans were the limiting factor. They don’t make sense in a world where systems can maintain context, apply constraints, and operate continuously.
Deloitte captures this tension clearly in its recent analysis of agentic AI: many organizations are trying to automate processes designed for humans instead of rethinking the work itself. The result is predictable: complexity increases, but outcomes don’t improve proportionally.
That’s not a tooling problem: that’s a design problem.
AI doesn’t optimize processes: it exposes them
One of the most consistent patterns across enterprise AI initiatives is this: the more you try to apply AI to an existing process, the more visible that process’s limitations become.
What was previously hidden behind human effort becomes explicit:
- missing data
- inconsistent rules
- unclear ownership
- duplicated work
- delayed feedback loops
In that sense, AI behaves less like an optimization layer and more like a diagnostic tool. It reveals the gap between how a company thinks it operates and how it actually operates.
This is why so many pilots stall. Not because the model fails, but because the process it is inserted into cannot absorb what the model produces. As MIT Sloan has argued, the challenge is not simply adopting AI, but redesigning organizations so that they can actually use it effectively.
And that leads to a much more uncomfortable conclusion: the limiting factor is no longer the technology. It’s the company.
From processes to systems
If the previous phase of enterprise AI was about adding intelligence to tasks. The next one will be about redesigning systems so that intelligence is embedded from the start.
That shift changes everything. Instead of asking:
- “How do we automate this step?”
Companies will have to ask:
- “Why does this step exist at all?”
- “What would this process look like if it were designed around continuous context?”
- “Where should decisions actually happen?”
- “What constraints should be enforced automatically?”
These are not incremental improvements. They are structural questions. And they point toward a different kind of organization: one where processes are no longer static sequences of actions, but dynamic systems that maintain state, integrate data, operate under constraints, and continuously adapt based on outcomes. The same characteristics that define the systems described in my previous article.
The companies that move first will look very different
This is where the shift becomes visible. The companies that successfully redesign their processes around these principles will not just be faster or more efficient. They will operate differently:
- decisions will happen closer to data
- coordination will require fewer handoffs
- feedback loops will shorten dramatically
- execution will become more continuous
- roles will evolve around systems, not tasks
Microsoft’s Work Trend Index already hints at this transition, describing organizations moving toward more dynamic, outcome-driven structures where humans and AI collaborate around goals rather than functions.
From the outside, these companies may not look dramatically different at first. But internally, their operating logic will have shifted. And that shift compounds.
This is not optional
It’s tempting to think of this as an opportunity. It is, it may well be. But it’s also something else: a constraint.
Because once some companies begin to operate this way, the others are not competing against better tools. They are competing against a different kind of system.
A system that:
- learns faster
- adapts continuously
- coordinates more efficiently
- executes with fewer delays
That is not something you can match by adding another copilot or deploying another model. It requires redesign.
The next phase of enterprise AI is organizational
If the first phase of AI in the enterprise was about experimentation, and the second about realization, the next one will be about transformation.
Not transformation driven by models, but by structure. We are not moving from “worse AI” to “better AI,” we are moving from companies built for humans, to companies that must operate with machines as part of their core logic. And that requires something many organizations have avoided for decades: rebuilding how they actually work.
The real question
So the question is no longer “how do we use AI?” It is: “are we willing to redesign our company so that AI can actually work?” Because if the answer is no, the outcome is already clear:
AI will not fail. Your processes will.
