Reading some of the current debates on AI, one might think that everything hinges on the sophistication of models or the speed of tool deployment. But the BCG and McKinsey reports I presented to you last week show something quite different: businesses are not lacking in technology or ambition, they are mainly lacking in an understanding of what AI really transforms within the organization (Adoption and impact of AI: lessons (and limitations) from the latest McKinsey and BCG studies). Most companies seek to move quickly, launching pilot projects and accumulating tools, but without ever addressing what truly determines collective performance: how work is designed, distributed, orchestrated, and learned.
The promises of AI reveal a reality that we like to ignore, namely that work as it is done differs from work as it is described, roles reflect history more than current needs, decisions are made in a labyrinth that no one has a complete view of, and internal infrastructures sometimes resemble a patchwork more than an intentional architecture. In other words, AI acts as a revealer of organizational flaws that have long been tolerated for lack of anything better (If your business isn’t designed for AI, it will end up being designed by AI and How management let systems do the thinking for them).
What sets the most advanced players apart is not access to the most powerful technologies, but an ability to fundamentally rethink the very structure of the business. They approach AI as an organizational transformation before making it a technological transformation. This is precisely what the reports highlight, and it gave me the idea to try to offer some runways for approaching the organizational transformation made necessary by AI.
In short:
- AI is not only transforming tools, but also requiring a fundamental rethinking of how work is organized, distributed, coordinated, and collectively learned.
- The most advanced businesses approach AI as an organizational transformation, starting by mapping actual work and redefining roles, decision-making processes, and the purpose of activities.
- It is necessary to reverse the traditional logic by starting with objectives and operating models before choosing the appropriate technological tools, in order to avoid automating existing processes without adding value.
- Successful adoption of AI relies on resolving systemic frictions, building a governed modular IT system, and clear governance of uses to ensure consistency, interoperability, and trust.
- AI only produces value if it is part of a continuous organizational learning process, where each experiment feeds into a structural improvement in collective functioning.
Transformation must begin with work design, not technology
Reports refer to workflows as if they were the core of the problem, when in fact they are only the visible surface. Behind each series of steps lies work that is often very different from what is described in official processes (Work about work: when the reality of work consists of making things that don’t work work). The risk, when limiting oneself to this cartographic vision, is to apply a logic of automation to practices that are already flawed. This amounts to reinforcing mechanisms that, in some cases, no longer have any reason to exist.
It is essential to distinguish between what is truly productive work and what feeds into this layer of peripheral activities that clutter up everyday life. Until we clearly separate the useful from the superfluous, AI will only streamline useless work without creating value. Clarifying the purpose of each process also means agreeing on what needs to be done, by whom, and why, before even discussing how a machine could be involved.
Rethinking the distribution of work between humans and AI requires a big-picture view that goes far beyond organizational charts. Mapping work becomes the essential starting point. It is on the basis of this reality that we can decide what needs to be eliminated, what needs to be redistributed, and what can be transformed. Without this, technology remains a veneer.
Implication: transformation must start with a map of actual work, not organizational charts or official processes.
Rethinking roles and decision-making structures
The advent of AI is changing the very nature of many tasks. Some are disappearing, absorbed by agents capable of managing a flow of micro-decisions, while others are focusing on more complex issues because they require expertise that automation cannot yet replicate. Still others are changing scale, becoming cross-functional where they were previously confined to a single team.
This reconfiguration requires a review of the granularity of roles. We can no longer think of these roles as a fixed set, but rather as an assembly of skills, decisions, and responsibilities distributed between humans and computer systems. The question is no longer just what each person does, but what room for action and interpretation they retain in the face of agents that are constantly active.
The way in which the organization makes its intentions clear from the outset (not to replace but to redesign the work) and demonstrates this in its decisions plays a decisive role in how teams respond to these changes. When teams perceive that the ongoing transformation puts them at risk, we cannot expect strong internal support. Worse still, there is a real risk of protective and defiant behavior towards the project, as we have already seen with topics whose original philosophy has been misused by putting them at the service of bad objectives (Is AI the new Lean? and Lean Without Layoffs: The Commitment That Makes Continuous Improvement Work). In other words, do not expect any engagement or support from people who think that the success of the project will mean their dismissal.
In this context, managers see their scope of intervention evolve and become the guarantors of consistency between human decisions and those produced by systems. Their role shifts towards integration, arbitration, and context adjustment rather than execution (Let’s rehabilitate the role of the manager). This shift requires a new clarification of the decision thresholds between humans and AI in order to avoid gray areas where no one knows who is responsible for what.
Implication: Rewriting roles is an essential task, not a secondary HR deliverable.
Replace the “tool ? use” logic with “objective ? operating model ? tool”
As long as we start with the tool, we are doomed to optimize what already exists. This is exactly what most businesses do: they automate the status quo, then are surprised when they only achieve marginal gains. The right approach is to start with the desired effect, whether it’s savings, fast, quality, consistency, or differentiation.
Once the objective has been clarified, it becomes possible to imagine a coherent operating model. This model, and this model alone, then makes it possible to determine the work to be done. Only at this stage can we decide what requires human input and what can be entrusted to AI. In this logic, the tool appears last, as the realization of a structured intention rather than an improvised response.
This reversal of the sequence profoundly transforms the way the business approaches its own internal architecture. It shifts the focus from technology to structure and decision-making, which is much more demanding but also much more structuring.
This is where we leave the realm of internal mechanics and enter that of enterprise design, where we ask ourselves how the organization wants to function before deciding how it will equip itself (Enterprise design before architecture: putting the company back the right way up).
Implication: AI transformation must be backed by a enterprise design approach, not just technical architecture.
Define a strategy for reducing friction
The concept of friction addressed by Sutton in his latest book (The Friction Project: How Smart Leaders Make the Right Things Easier and the Wrong Things Harder) offers a very interesting approach. Most everyday irritants are not related to motivation or competence, but to systemic constraints that slow down decisions, fragment information, saturate teams’ cognitive capacities, or simply prevent employees from doing their jobs as well as they would like or could. A useful AI transformation must begin with a careful reading of these frictions.
Identifying decision-making processes that are too slow or costly means gaining new visibility into what is hindering collective momentum. Similarly, understanding how information actually flows within the business helps prevent AI from reproducing existing silos. Analyzing cognitive flows means recognizing that one of the most immediate values of AI is to declutter the mind, not to replace judgment. And in general, identifying everything that prevents people from doing their jobs well, or even doing their jobs at all, and tackling organizational complexity are two areas that businesses should have been focusing on for a long time, and which the arrival of AI has brought back to the top of the agenda, in the hope that this time, the issue will not be swept under the carpet (The organizational complication: the #1 irritant of the employee experience) .
At this stage, technology is still only a means to an end. The central problem remains the resolution of organizational dysfunctions that have become entrenched over time and have never really been addressed. AI becomes relevant when it is part of this logic of streamlining.
Implication: AI transformation must begin with the resolution of systemic problems, not with technology.
Building modular IT and a governed data model
Reports focus on infrastructure, but they often do so through a purely technical lens, whereas the issue is broader: it is about designing IT that can evolve without being weakened by each new component. Modularity becomes essential for isolating, connecting, or replacing AI building blocks as needs change.
Interoperability, meanwhile, prevents the emergence of model silos that each behave like a small black box (Digital workplace, AI, and interoperability: a problem that remains unresolved). Clear governance prevents the organization from ending up with a patchwork of unspoken rules, introduced along the way in prompts, exceptions, or local adjustments. Finally, observability makes system decisions understandable, which is essential when these decisions impact critical chains and, above all, to build trust in the tools.
This change in posture is transforming the IT department. It can no longer limit itself to technical execution, but is now responsible for the overall consistency of internal capabilities and the quality of their articulation.
Implication: The IT department must move away from its role as a technical executor to become the architect of internal capabilities.
Developing an orderly adoption model
It is tempting to move quickly, often because external pressure is strong and leaders want to see progress. However, experience shows that speed is only an advantage if it is done in an orderly manner. Starting with high-value, low-friction use cases creates a credible foundation before tackling more complex projects.
Redesigning workflows then becomes an essential step. Until they are stabilized, any attempt at mass deployment exposes the business to setbacks. A governance model creates a common framework that limits local interpretations and maintains the consistency of the system. In addition, capitalizing on successes and failures and developing skills must follow suit, otherwise the organization moves forward without really understanding what it is implementing.
Only once these conditions are met can agents be deployed in stabilized flows. The gain is not in immediacy but in reliability.
Implication: order matters more than speed.
Create an organizational learning system
The businesses that will benefit from AI will not be those that have accumulated the most pilots, but those that have been able to learn in a structured way (95% of Enterprise AI Pilots “Fail”–Just Like Lean? Not So Fast and The experimental organization). Testing, analyzing, adapting, stabilizing, and then industrializing requires a new discipline, almost a routine, that goes beyond the scope of traditional projects.
This learning becomes collective when lessons are shared, what works is documented, and reusable workflow patterns are created. AI can also contribute to this capitalization by detecting patterns, revealing inconsistencies, or suggesting improvements.
Organizational design must therefore integrate this logic of continuous adjustment into the very functioning of AI systems so that the business can build on what it learns instead of falling back into the same mistakes.
Implication: Organizational design must include a continuous improvement loop integrated into AI.
Bottom Line
Both reports converge on the same observation: AI is not a technological issue but an organizational design challenge. Deployments carried out without reshaping work, roles, and decision-making mechanisms produce only limited value. The gap between leaders and followers cannot be explained by the technology available or the budget allocated, but rather by the ability to redefine how the organization actually works.
AI is certainly accelerating, but it accelerates what it is given to accelerate, and as long as the foundations remain unchanged, speed only magnifies the flaws.
To answer your questions…
AI highlights internal dysfunctions that are often overlooked: discrepancies between actual work and job descriptions, fragmented decision-making, rigid roles, and fragile infrastructure. Successful businesses don’t just add tools; they rethink how work is distributed, coordinated, and learned. AI accelerates what already exists, which is why it’s important to review the foundations before deployment. Without this organizational redesign, gains remain marginal. Rethinking the structure allows objectives, decisions, and daily operations to be aligned.
Official workflows do not reflect the reality of work. Relying on them leads to the automation of inefficient practices. Mapping actual work allows us to identify useful tasks, those that should be eliminated, and those that should be redistributed. AI then becomes a lever for transformation rather than a simple accelerator of flawed routines. This transparency helps us to clearly decide on the respective roles of humans and systems before introducing tools.
The automation of micro-decisions is changing the nature of tasks: some are disappearing, while others are becoming more complex or cross-functional. Roles must be redefined as a combination of skills and responsibilities shared with systems. It is becoming essential to clarify who decides what in order to avoid gray areas. Managers are evolving toward a role of integration and arbitration. Without clear intentions, teams fear for their job security, which slows adoption.
Moving quickly without order results in fragile deployments. First, choose low-friction use cases, then stabilize workflows before any large-scale deployment. Governance ensures consistency and limits local deviations. Skill development and capitalization on learning ensure a real understanding of what is being implemented. Once these steps have been followed, speed becomes an asset because it is based on solid foundations.
The most successful organizations are those that systematically learn from their tests. Each pilot allows them to adjust their work, document what works, and stabilize practices before industrialization. AI also helps identify inconsistencies and recurring patterns. By integrating a continuous improvement loop into its systems, the business avoids repeating the same mistakes and strengthens its ability to evolve with AI.
Image credit: Image generated by artificial intelligence via ChatGPT (OpenAI)
Related articles:
Adoption and impact of AI: lessons (and limitations) from the latest McKinsey and BCG studies
AI adoption: methods put to the Test
EDGY: a common language to align identity, experience, and operations
Who is handling your artificial intelligence projects? Probably not the right people.
How management let systems do the thinking for them
Automation is done with people, not against them.
