Organizing for AI: from experimentation to industrialization

SECTION 1

Why 88% of AI POCs never reach scale

Artificial intelligence had its "internet moment" with the arrival of mainstream generative AI, then a second wave with autonomous agents. The sense of urgency has never been stronger. Yet beneath the hype, a stubborn reality persists: while a large majority of companies say they are piloting AI initiatives, roughly 88% of POCs never reach large-scale deployment. The pattern repeats itself: promising pilots launched with no clear path to industrialization, eroding trust and feeding skepticism.

It is rarely the technology alone that blocks progress. Misaligned priorities, scattered experimentation, legacy data, talent shortages, and a shifting regulatory landscape slow things down, while concerns around bias, privacy, and transparency add their own complexity. Leaders face a paradox: the cost of inaction is rising, but moving forward without the right foundations is costly too.

### The lesson that stands out

Unlocking AI value depends less on algorithms than on how the company organizes itself to use them, responsibly and at scale. Who builds the models, how teams collaborate, and how risks are managed determines whether AI stays stuck in pilot mode or becomes a repeatable engine of value. At Adservio, we regularly see organizations investing heavily in models and platforms while neglecting structures and processes: the result is dozens of pilots that never move past experimentation, for lack of organizational capacity, not technical capability.

SECTION 2

AI maturity: a three-stage journey, not a big bang

For most organizations whose core business is not technology, becoming an AI-enabled enterprise is a gradual evolution, not an overnight switch. But before launching teams or allocating budget, there must be a clear vision of how AI will support the business: without that clarity, even the best AI teams risk being underused, then disbanded, a pattern we see far too often.

Once strategic intent is in place, capabilities are built over time. AI maturity reads as a three-stage journey: a dedicated AI squad that creates momentum and the first proofs of value; distributed squads that integrate AI into key business domains, where value is created and decisions are made; and finally an industrialized AI Factory, where capabilities are embedded in the operational DNA and continuously improved across the enterprise.

Each stage prepares the next: the technical foundations, talent, and proofs of value from the first squad condition the success of distribution, and the discipline gained during distribution conditions industrialization. Skipping stages almost always amounts to stacking up pilots that lead nowhere.

SECTION 3

Stage 1: The dedicated AI squad: prioritize and learn fast

Innovation efforts need a safe space to experiment. A dedicated AI squad, typically an agile team of five or six people mixing software engineers, data scientists, ML engineers, and a product manager, builds capabilities and delivers the first use cases, in close contact with business stakeholders. When generative and agentic AI is the focus, strong software engineering and architecture skills matter more than pure modeling expertise.

### A disciplined prioritization framework

Early success depends on demanding prioritization: too many AI teams chase appealing ideas with no strategic value. We use multi-criteria scoring that evaluates each use case on four axes, measurable business impact (revenue, cost, customer experience), technical feasibility (data, technology, available skills), regulatory, ethical and operational risk, and strategic alignment. This transparent framework avoids the trap of projects that are technically interesting but strategically insignificant.

### Measure the speed of learning, not just deliverables

Equally important: providing a sandbox where experiments run fast without threatening production. In AI, uncertainty is a feature, not a flaw, the metric that matters is the speed of learning. Early squads should be measured on hypotheses tested per sprint, time from hypothesis to validated insight, the share of pilots reaching a defined ROI, and adoption speed among end users. Done well, this first squad does more than deliver a few wins: it seeds the AI platform's foundations and creates reusable artifacts, data pipelines, evaluation frameworks, integration patterns, governance rules, that will accelerate every squad that follows.

SECTION 4

Stage 2: Distributed squads: embedding AI into the business

Once value is demonstrated, the next step is anchoring AI more deeply in the business: forming domain-aligned squads, marketing, sales, operations, risk, product, able to build and integrate AI solutions where value is created. But scaling is not just hiring: which domains are ready? Is the stack modular enough to avoid starting from scratch? Are local leaders ready to sponsor adoption and drive change? Not every function will move at the same pace, and forcing uniformity too early often backfires on the organization.

### Choosing the domains that are ready

Two criteria make this stage executable. First, a clear problem owner, ideally with P&L accountability: without a leader whose objectives depend on the initiative's success, projects stall. Second, pre-existing demand for analytical solutions that AI can accelerate: go where teams are already asking for better insights, rather than forcing AI onto domains that do not see the value. Start small and focused, then build momentum through visible wins.

At this stage, change management becomes critical: employees need support to adapt their workflows, trust AI-generated results, and experiment within safe boundaries. The winning patterns we observe: embedded AI champions who live within business teams and speak their language, training contextualized to the domain rather than generic courses, quick wins celebrated publicly, and guardrails that are clear but not punitive.

@cite:naviguer-le-scaling-de-l-ia

SECTION 5

The AI center of excellence: accelerate without bureaucratizing

As adoption spreads, a well-designed center of excellence (CoE) supports responsible, efficient scaling. Its purpose: aligning governance, technology, skills, and business priorities, so that AI is deployed safely, sustainably, and with measurable value. In the GenAI and agentic context, early efforts focus on foundations, shared platforms, tooling, secure deployment practices, before advanced capabilities such as fine-tuning or custom architectures emerge with maturity.

Designed well, the CoE does not slow teams down; it accelerates them. Its key responsibilities cover standards and governance (how AI is built, tested, deployed, and monitored), shared platforms and tools that reduce duplication, talent development, not only data scientists but also engineers, product managers, and business stakeholders, measuring and spreading the learnings, and expert support for challenges the distributed squads cannot solve on their own.

At Adservio, we help our clients design CoEs that are "thin but mighty": light enough to avoid creating bureaucracy, strong enough to create real value, a few senior people, reusable artifacts, and a veto right limited to genuine risks.

@cite:les-cinq-dimensions-de-l-adoption-de-l-ia

SECTION 6

Stage 3: The AI Factory: industrializing decisions with agentic AI

At full maturity, an AI-enabled organization operates like a modern factory, but for decision-making and learning. Rather than producing physical goods, the AI Factory systematically turns raw data into predictions, insights, and automated actions that spread into every process. The emblematic examples still speak for themselves: TikTok's feed learns preferences in real time from every interaction; Amazon continuously adjusts millions of prices and logistics routes; John Deere deploys agricultural equipment that self-optimizes from field data.

As "Competing in the Age of AI" highlights, the AI Factory breaks the link between growth and headcount by embedding intelligence into workflows and products. It acts as a digital operating system, data, algorithms, experimentation, infrastructure, that learns from every interaction to drive faster, more consistent decisions at scale. It is this continuous-learning loop that separates true AI factories from organizations that have merely deployed a few models.

### From prediction to autonomous action

Since 2025, this idea has crossed a threshold with agentic AI, which has moved from emerging to operational deployment: agents that no longer just predict but act in context, handling a customer request end to end, remediating incidents, orchestrating processes. The AI Factory of 2026 therefore combines predictive models, GenAI, and agents under a single requirement: explicit guardrails, human supervision proportionate to risk, and full traceability of actions. Autonomy without governance does not create a factory; it creates systemic risk.

@cite:mlops-de-l-experimentation-a-la-production-a-grande-echelle

SECTION 7

Governance, the AI Act, and principles that last

The journey now unfolds within a regulatory framework that is taking shape. In Europe, the AI Act applies in stages: obligations on general-purpose models have been in force since August 2025, and the "Digital Omnibus" package adopted in spring 2026 postponed most high-risk system obligations to late 2027, a delay to be used for building governance, not for putting it off. System inventory, risk classification, documentation, and human oversight are becoming organizational capabilities in their own right, naturally carried by the CoE.

### Five principles that hold up

Beyond compliance, a few principles separate organizations that industrialize from those that stack up pilots. Start from business value: tie AI outcomes to executive objectives, because AI for AI's sake gets neither budget nor attention. Fund experimentation with intent: AI is probabilistic and not every pilot will succeed; a balanced portfolio, safe bets, medium initiatives, a few moonshots, secures short-term wins while exploring transformative potential. Steer from the top and the bottom: leadership sets direction and accountability, while the ground needs the skills, tools, and space to build. Design for scale from the start: every pilot should leave a reusable asset behind. And rewire how work gets done: do not insert AI into existing processes, rethink processes around what it makes possible.

In AI, there is no finish line: new architectures, agentic capabilities, and regulations keep moving the goalposts. What matters is not reaching an endpoint but building the organizational muscle to adapt, learn, and industrialize at pace. At Adservio, we start with an honest assessment of AI maturity, identify the critical gaps, organization, technology, data, skills, and build a pragmatic roadmap that balances ambition and feasibility. Success in AI does not belong to those with the best algorithms, but to the most adaptable organization.

FAQ

Frequently asked questions