AI Uncovers Hidden Bottlenecks in Teams

AI Exposes Hidden Bottlenecks

The integration of artificial intelligence (AI) into development teams has not led to a reduction in delivery times, as many had hoped. The issue lies not with the tool itself, but with the pre-existing bottlenecks within the production chain, which are now being highlighted by AI.

AI has the ability to compress the codifiable links in the software production chain. However, it does not resolve the bottlenecks that already existed, such as testing, integration, deployment, and coordination. On the contrary, it makes them more visible by increasing the workload in already saturated systems.

Code Is Not the Only Bottleneck

In November 2025, Anthropic launched a new version of its tool Claude Code. Developers who tested it over the Christmas holidays were surprised by its ability to complete projects in a few hours that would normally take weeks. This acceleration is illustrated by impressive figures: at Google, more than 30% of new code is generated by AI, according to Sundar Pichai in the first quarter of 2025. At Microsoft, Satya Nadella estimates this proportion to be between 20% and 30%, although results vary by programming language. These figures, while estimated by executives in front of their shareholders, indicate a clear trend.

Yet, despite the massive integration of AI tools, two-thirds of organizations have not seen any reduction in their workforce. Expanding the analysis beyond human impact, time-to-market, actual delivery throughput, and production stability also show no significant gains. Why? Because writing code has never been the only limiting factor. Anyone who has participated in a software project knows that time is often lost in manual testing, regressions, waiting for a staging environment, integration cycles, deployment procedures, and coordination between teams. The DevOps movement was born from this realization. AI compresses the most codifiable link in the chain: writing. But the bottlenecks that already existed in testing, integration, and deployment do not disappear. They become more visible as the rest of the process accelerates.

The image is simple: it’s like increasing the flow of a highway while the exit ramp remains a single lane. The result is not increased fluidity, but a shifted bottleneck.

This is exactly what happened with the cloud. Companies that simply moved their infrastructures to the cloud without rethinking their architecture ended up paying more for the same result. Those that rethought their architecture gained a sustainable advantage. AI poses the same question: should we automate one link without touching the rest, or restructure the entire delivery system?

To answer this question, we need to examine where AI concretely impacts the business, across four layers.

Layer 1: Software Production

Current agents, such as Claude Code, Cursor, Codex, and Copilot, excel at easily verifiable tasks: generating standard code, fixing bugs, testing, and limited-scope refactoring. Tasks requiring judgment, such as architecture, design arbitration, and ambiguous business needs, remain in the human domain.

The enticing idea of a multi-agent orchestration where each agent holds a business role (PO agent, UX agent, front-end dev agent) does not correspond to reality. Agents specialize by capability and scope, not by job description. However, supervised multi-agent systems are already functioning 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 outright rejection.

The emerging model is simpler: delegate, verify, assume. The human defines the intent. The agent implements. The human validates and takes responsibility. This process, which would have taken three days, now takes two hours. But it requires a quality of judgment that seniority alone does not guarantee.

This point deserves emphasis. What is increasing in value is not seniority per se. It is the ability to set the right constraints, to detect what the agent misunderstood, to decide in ambiguity. A routine senior who merely validates without reading remains mediocre with or without AI. A less experienced profile but with a true system sense can create much more value.

For a manager, the consequence is direct: the skills to evaluate are changing. 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 adhere to, and to spot in a 500-line diff the flaw that the agent missed.

Layer 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 highlights a persistent signal of instability in production, confirming that the bottleneck does not disappear; it shifts towards stabilization. AI acts as a mirror and a multiplier. Solid foundations: compounded gains. Fragmented organization: amplified chaos.

Data quality, API modularity, and application consistency are no longer secondary technical issues. They are 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 responsible for this issue? Not necessarily the traditional IT department, often too operational. Depending on the organization's profile, it could be a platform management, a product management, or a CTO positioned at the executive committee level. What matters is that a clear authority is responsible for the foundation on which AI operates. In the most successful companies, this shift is already happening: nearly two-thirds of their technology leaders are involved in developing corporate strategy (Global Tech Agenda 2026).

The build vs. buy question also arises. Building one's agentic infrastructure internally offers a competitive advantage but is costly. Buying third-party platforms accelerates deployment but creates dependency. Failing to make a decision allows each business unit to purchase its own tools without overall coherence. This is already happening.

Layer 3: Team Organization

This is the most mined terrain.

In some contexts, part of the gross 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, existing debt, and level of legacy, this model may or may not work. There is no general law. Today, headcounts have not changed in most organizations because structures have not changed. But the emerging model in the most advanced product companies suggests smaller, more expensive teams that demand higher quality judgment.

What is documented, however, is the collapse of the training pipeline. A Stanford study (Brynjolfsson et al., ADP data) shows that employment for developers aged 22 to 25 has declined by nearly 20% from its peak at the end of 2022. Among the Magnificent Seven, recent graduates now account for only 7% of hires, down from 15% before the pandemic (SignalFire). The history of computing invites caution: each increase in abstraction has ultimately increased the total demand for developers. But even if employment volume rebounds, the nature of junior work is changing. And if no one reinvents the path to skill development, the pool of senior judgment will mechanically dry up in five to ten years. This is not an IT problem. It is a general management problem.

At the same time, AI is overflowing into non-technical teams. Customer service is deploying conversational agents. Marketing is connecting tools to the CRM. Finance is automating reporting. Shadow AI is shadow IT on steroids: same causes, same risks, multiplied speed.

Layer 4: Governance

Without it, the previous three layers produce instability.

Four key areas emerge:

• Define codified policies that frame the autonomy of agents: levels of authorization, safeguards, human approvals, auditability. Not all tasks deserve the same level of supervision. Automatically generated documentation can operate autonomously. A production deployment requires multi-person validation.

• Responsibility: when an agent pushes a change that brings down a critical system, who is accountable? AI does not bear responsibility. The human who validated does. 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 making decisions impacting customers without traceability of its reasoning exposes the company to concrete legal risk.

• Cross-functional governance: who decides when AI crosses product, data, security, legal, and business lines? Who arbitrates? Who bears the risk? This is precisely where companies stumble: everyone wants AI, but no one wants the responsibility for what it does. Today, according to Gartner, 17% of IT leaders manage this coordination. 69% expect to do so by 2030. The gap measures the extent of the work.

The one-for-one replacement: replacing a developer with an agent without changing the process does not create gains. This is why the majority of equipped companies have not saved anything.