Have you deployed AI tools in your development teams but your delivery times have not changed? The problem is not the tool. It is elsewhere in the chain and existed before AI.
AI compresses the codifiable links in the software production chain. It does not resolve the bottlenecks that were already there: testing, integration, production, coordination. It makes them visible by sending more volume into already saturated pipes.
Code was never the only bottleneck
In November 2025, Anthropic released a new version of Claude Code. Developers spent their Christmas vacation experimenting with it. Many came back troubled: the tool had built projects in a few hours that would have taken them weeks to code.
The numbers give the measurement of acceleration. At Google, more than 30% of new code is generated by AI (Pichai, Q1 2025). At Microsoft, Nadella estimates the proportion at 20-30%, with unequal results depending on the language. These numbers are not scientific measurements. These are estimates from executives in front of their shareholders. But the trend is undeniable.
Yet two-thirds of organizations that have heavily integrated AI tools have seen no reduction in their workforce. And if we broaden the look beyond the human cost, time to market, actual delivery rate, stability in production; the observation is often the same: the tool is there, the systemic gain is not. For what ?
Because writing code has never been the only constraint. Anyone who has been through a software project knows where time is really wasted: in manual testing and regression, in waiting for a testing environment, in integration cycles, in release procedures, in coordination between teams. The entire DevOps movement was born from this observation. AI compresses the most codifiable link in the chain: writing. But the bottlenecks that already existed in testing, integration and MEP do not disappear. They become more visible, because the rest goes faster.
The image is simple: it’s like doubling the flow of a highway whose exit ramp remains one lane. The result is not fluidity. This is the displaced traffic jam.
This is exactly what happened with the cloud. Companies that did “lift and shift” paid more for the same result. Those who have rethought their architecture have gained a lasting advantage. AI asks the same question: do we automate one link without touching the rest, or do we restructure the delivery system as a whole?
To answer this, we must look at where AI actually affects the company. On four strata.
Stratum 1: software production
Current agents (Claude Code, Cursor, Codex, Copilot) excel in a specific register: easily verifiable tasks. Standard code generation, bug fixes, testing, limited scope refactoring. What requires judgment (architecture, design arbitration, ambiguous business needs) remains human.
The attractive idea of a multi-agent orchestration where each agent would play a business role (PO agent, UX agent, dev front agent) does not correspond to reality. Agents specialize by ability and scope, not by job description. On the other hand, supervised multi-agent systems already operate in production on decomposable and bounded tasks: due diligence pipelines, data engineering workflows, predictive maintenance. What remains immature is complete autonomy on open creative projects: those that require arbitration, business context and judgment. Confusing the two is either overselling or rejecting everything altogether.
The model that emerges is simpler: delegate, verify, assume. The human sets the intention. The agent implements. Humans validate and bear responsibility. This process, which would have taken three days, takes two hours. But it requires a quality of judgment that seniority alone does not guarantee.
This point deserves to be emphasized. What increases in value is not seniority as such. It is the ability to pose the right constraints, to detect what the agent has misunderstood, to decide in ambiguity. A routine senior who only validates without reading remains mediocre with or without AI. A less experienced profile but with real system awareness can create much more value.
For a manager, the consequence is direct: the skills to be evaluated change. The writing speed no longer differentiates. What differentiates is the ability to break down a problem into delegable subtasks, to specify constraints precise enough for an agent to respect them, and to identify in a diff of 500 generated lines the defect that the agent did not see.
Stratum 2: information systems architecture
If AI amplifies what exists, then the quality of what exists determines everything. The DORA 2025 report shows real gains in individual efficiency and product performance. But it also notes a persistent signal of instability in production, which confirms that the bottleneck is not disappearing, it is moving towards stabilization. AI acts as a mirror and a multiplier. Solid foundations: compound gains. Fragmented organization: amplified chaos.
Data quality, API modularity, application consistency are no longer secondary technical subjects. These are the prerequisites without which agents produce noise. An agent connected to a poorly governed database does not generate insights. It generates errors at high speed.
Who is carrying this topic? Not necessarily the classic IT department, often too operational. Depending on the profile of the organization, it will be a platform department, a product department, a CTO positioned at the COMEX. What matters is that there is clear authority responsible for the foundation on which the AI operates. In the most successful companies, this shift has already been made: nearly two thirds of their technology managers participate in the development of the corporate strategy (Global Tech Agenda 2026).
The build vs buy question also arises. Building your agentic infrastructure in-house offers a competitive advantage but is expensive. Buying third-party platforms speeds up deployment but creates dependency. Not deciding means letting each profession buy its own tools without overall coherence. It’s already happening.
Stratum 3: team organization
This is the most minefield.
In certain contexts, part of the raw production capacity can be absorbed by a smaller core of experienced profiles equipped with agents. Velocity increases, but the economic model changes: four senior profiles cost more than six juniors. Depending on the product, the existing debt, the level of legacy, this model may or may not work. There is no general law. Today, the workforce has not changed in most organizations because the structures have not changed. But the model that is emerging in the most advanced product companies prefigures teams that are smaller, more expensive, and more demanding in terms of quality of judgment.
What is documented, however, is the collapse of the training pipeline. A Stanford study (Brynjolfsson et al., ADP data) shows that employment of developers aged 22 to 25 has fallen by almost 20% compared to its peak at the end of 2022. Among the Magnificent Seven, recent graduates now only represent 7% of hires, compared to 15% before the pandemic (SignalFire). The history of computing calls for caution: each rise in abstraction has ended up increasing the total demand for developers. But even if the volume of employment is recovering, the nature of junior work is changing. And if no one reinvents the path to increasing skills, the pool of senior judgment will automatically dry up in five to ten years. This is not a CIO problem. This is a general management problem.
At the same time, AI is beyond technical teams. Customer service deploys chatbots. Marketing plugs tools into the CRM. Finance automates reporting. Shadow AI is shadow IT on steroids: same causes, same risks, increased speed.
Layer 4: governance
Without it, the three previous strata produce instability.
Four projects are required:
- Define codified policies that govern the autonomy of agents; authorization levels, guardrails, human approvals, auditability. Not all tasks deserve the same level of supervision. The automatically generated documentation can run autonomously. A production deployment requires multi-person validation.
- Responsibility: when an agent pushes a change that brings down a critical system, who responds? The AI bears no responsibility. The human who validated, yes. And if no one validated, that’s the problem.
- Regulatory compliance: the EU AI Act imposes transparency and explainability obligations on companies deploying these systems. An agent who makes decisions impacting customers without traceability of his reasoning exposes the company to a concrete legal risk.
- Transversal governance: who decides when AI crosses product, data, security, legal and businesses? Who referees? Who bears the risk? This is precisely where companies are failing: everyone wants AI, no one wants responsibility for what it does. Today, according to Gartner, 17% of CIOs manage this coordination. 69% expect to do so by 2030. The gap measures the scale of the work.
The traps
- One-for-one replacement: replacing a developer with an agent without changing the process does not create a gain. This is why the majority of equipped companies have saved nothing.
- The invisible debt: “vibe coding”, non-specialists who generate functional but fragile applications, produces systems that are less maintainable and more vulnerable. Apiiro documented a tenfold increase in security breaches per month at Fortune 50 companies between December 2024 and June 2025. The cost won’t be seen this quarter. See you in two years.
- The false economy of speed: producing three times more code that piles up waiting for review, saturates test environments and generates more incidents in production is a clear step backwards disguised as progress. The acceleration of a link only makes sense if the rest of the delivery system absorbs it.
What is decided now
64% of technology executives plan to deploy agentic AI in the next 24 months (Gartner). Only a minority have the foundations to do so.
Three decisions cannot wait any longer and their order is important:
- Audit the reality of shadow AI in the organization, because it is the only way to know where we really stand before deciding anything else.
- Designate a clear authority responsible for agentic governance and technical foundation. Without governance, every AI initiative adds chaos.
- Change metrics. Stop counting POCs and start measuring the time to market, the defect rate in production, the capacity to absorb a change in specification.
AI does not replace teams. It compresses one link in the chain and exposes all the others. The bottlenecks were already there: testing, integration, production, coordination, clarification of needs. The general slowness masked them. The acceleration of AI reveals them. Organizations that restructure around these bottlenecks will gain a structural advantage. Others will discover, project after project, that they are accelerating in the wrong place.
