News \ The 90% behind AI: engineering for reliability

The 90% behind AI: engineering for reliability

Leo Simons & Julian Hessels
March 27, 2026 · 3 min read English

AI has matured, now it must be engineered

Artificial intelligence has entered a decisive decade. It is no longer confined to innovation teams or exploratory proofs of concept, but steadily becoming embedded in the everyday digital infrastructure of enterprises. AI now touches workflows, customer journeys, financial analysis, operational processes, and software delivery. In doing so, it has moved from optional capability to structural dependency.

This shift requires a different conversation at board level. The central question is changing from “What can AI do?” to “How do we ensure AI behaves predictably, transparently, and under human control when decisions matter most?” The enterprise adoption of AI asks for architectural clarity and responsible integration. The technology has matured. Our responsibility around it must mature as well.

What we are witnessing is not simply a technology shift, but the emergence of a production gap. A small group of organizations are industrialising AI structurally, embedding it into core operations. The majority remain in pilot mode. The gap between these two groups compounds over time, in cost base, speed, knowledge accumulation, and decision quality. AI maturity is no longer a question of experimentation, but of competitive positioning.

Our mission-critical heritage gives us a clear position: AI is not a silver bullet. It is an enabler whose value depends overwhelmingly on the engineering, governance, and operational integrity that sits beneath it. We call this perspective the 90% behind AI: a reminder that reliable AI is defined far less by the power of its models than by the trustworthiness of the systems around them.

The next phase of enterprise AI

AI is no longer a collection of pilots or innovation experiments. It has become a structural component of our business that is influencing customer experience, continuity, compliance, and operational excellence. The question is no longer what AI can do, but how we stay in control as intelligence becomes part of our core infrastructure.

As Jensen Huang (CEO Nvidia) recently observed, every company will soon operate two factories: one that produces its products or services, and one that produces intelligence. The second factory is not metaphorical. It is an engineered infrastructure for data, models, orchestration, and control. Organizations that treat AI as a project build experiments. Organizations that treat it as a factory build capability.

This shift is visible across sectors: organizations that succeed are those who treat AI not as experimentation, but as infrastructure that requires discipline, governance, and transparency.

Predictable performance

AI promises acceleration, insight, and new possibilities. But as it becomes intertwined with decisions that affect customer trust, revenue, and compliance, its reliability becomes the defining factor of success. When AI fails, the impact is no longer isolated. It becomes operational, financial, and reputational.

Europe’s governance-first approach emphasizes explainability, accountability, lineage, security, and sovereignty. Rather than slowing innovation, this provides clarity, helping organizations scale AI responsibly rather than experimentally.

This shift resembles earlier transformations in mission-critical IT. Reliability becomes the frontier of competitiveness. Control becomes a design principle. And AI begins to move from being an exciting capability toward becoming a dependable component of enterprise architecture.

As the pace of automation accelerates, reliability, transparency, and human authority become the foundation of trust. Across industries, we see the same pattern: success comes not from having the most AI, but from ensuring AI behaves predictably when stakes are highest.

Enterprises that fail to build governance, oversight, and engineering discipline into AI systems expose themselves to operational, compliance, and reputational risks that no longer stay contained.

The 90% behind AI

Turning AI from prototype to production demands engineering discipline. Systems must be observable, testable, explainable, and always under human supervision. As AI systems evolve toward more autonomous behavior, governed memory and controlled state management become equally critical. Without clear boundaries around what an AI system can remember, modify, or trigger, small deviations can compound across workflows and interacting agents. Enterprise reliability therefore depends not only on model accuracy, but on disciplined control of state, permissions, and escalation paths.

This is the essence of the 90% behind AI: the data foundations, software design practices, governance structures, integration patterns, and security measures that collectively determine how AI behaves.

Data you can trust is the anchor. You need to know where it comes from, how it’s been handled, and whether anything has shifted along the way. That clarity ensures every result can be explained, defended, and improved. Quality isn’t something you check after the fact, it’s a daily standard, not a line buried in documentation.

The same discipline applies to the AI models themselves. They need clear versions, constant monitoring, real-world testing, and the ability to quickly switch back if something goes wrong. Visibility shouldn’t be optional; systems must make behavior obvious, not assumed. And governance should be built into how things work, through transparency and reversibility, rather than added as extra bureaucracy.

And above all, human oversight must remain absolute. AI augments human judgment and, where appropriate, can execute within clearly defined boundaries under human supervision. Clear escalation paths, role-based accountability, and human-in-the-loop principles ensure that responsibility lives with people, not with models.

This is how intelligence becomes infrastructure rather than risk.

Our direction: reliability by design

We focus intentionally on the 90% behind AI, which is the engineering, governance, and human oversight that ensure dependable outcomes.

Our approach is simple and strategic:

  • Use AI where it creates measurable business value
  • Keep humans in the loop for all meaningful decisions
  • Protect our data, our customers, and our integrity
  • Scale what works, responsibly and visibly

These principles mirror our internal values: clarity of purpose, responsible data use, transparency, risk awareness, and continuous improvement.

This alignment ensures that as AI capabilities evolve, control remains with us, not with the system.

Why Narrow AI works for the enterprise

The last decade centered on monolithic, general-purpose models: powerful but opaque systems attempting to solve many problems at once. In business-critical contexts, reliance on monolithic, general-purpose models introduces ambiguity, unexplainability, and operational risk that many enterprises are not prepared to accept. Even when powerful general-purpose models are used, enterprises increasingly constrain them within domain-specific boundaries, policy layers, and orchestration frameworks to ensure accountability and operational control.

Enterprises are gravitating toward narrow, domain-specific AI systems — not because general models lack capability, but because organizations need control. Domain-focused AI operates within clear tasks and defined boundaries. That makes it easier to govern, validate, explain, and align with regulatory frameworks such as the EU AI Act. In that sense, it behaves more like enterprise software: testable, observable, and accountable.

General models like Claude are being deployed at scale, but often within constrained use cases. Even powerful foundation models require context, guardrails, and orchestration layers to become enterprise-ready. Raw capability alone does not equal operational reliability. At scale, AI needs a harness.

This is not a retreat from general intelligence. It is architectural discipline. Enterprises are learning that adoption depends less on model breadth and more on clarity of scope, accountability, and integration into existing systems of record and control.

Over the next six months, the balance may shift in terms of tooling and vendor positioning, but the underlying enterprise logic is unlikely to change: organizations will continue to combine general models with domain-specific layers that constrain, contextualize, and operationalize them.

As with IT modernization, the winning architecture is rarely the most powerful in isolation. It is the one that can be governed, scaled, and trusted.

The scaling problem is the real problem

Many organizations succeed in deploying AI in a single use case. A chatbot works. A document analyzer works. A coding assistant works. The mistake is assuming this success transfers automatically to the next domain.

Today’s AI agents reason from scratch for every task. Context is reconstructed repeatedly. Successful procedures are not stored as reusable runbooks. Token consumption escalates. Costs rise. Latency increases. And unpredictability becomes the hidden tax of scale.

Without shared knowledge infrastructure, AI remains a collection of islands rather than a coherent enterprise capability.

Modularity instead of monoliths

Instead of relying on a single, large model, organizations are shifting toward multi-agent systems composed of many smaller, specialized components. Each agent performs one task transparently, with explicit inputs, outputs, and responsibilities.

Inspired by approaches such as the 12-Factor Agents framework, this modular intelligence allows systems to evolve without becoming fragile. Agents can be upgraded, replaced, or versioned independently. Failures remain local. Boundaries stay clear. Human oversight remains intact.

This is AI designed for how enterprises actually work: distributed, federated, supervised, and resilient. As agent-based architectures mature, orchestration becomes as important as model performance. Agents increasingly execute tasks, call tools, and coordinate with other systems. This requires runtime governance: explicit permissions, scoped memory, continuous observability of actions, and the ability to intervene or revoke capabilities instantly. Without this layer of control, autonomy scales risk faster than value.

The cognitive backbone: shared intelligence at scale

Scaling AI requires more than modular agents. It requires a shared cognitive backbone: a layered knowledge infrastructure that formalizes domain entities (ontology), current operational state (context graph), and standardized business definitions (semantic layer).

Above this backbone sit reusable runbooks, which are proven procedures that agents can execute without recomputing every decision from first principles. Instead of improvising each task, the system learns structurally.

The result is not only lower token consumption and latency, but compounding organizational intelligence. Each execution strengthens the system rather than remaining an isolated interaction.

When intelligence becomes infrastructure, control must be intentional. Observability, lineage, and explainability turn opaque automation into transparent digital teamwork. These systems behave like co-engineers and are visible, inspectable, and aligned with policy. Nothing runs wild. Everything runs under governance.

Governance as an asset

European AI maturity is defined by a governance-first mindset. Regulation such as the EU AI Act, NIS2, and GDPR is not a brake on innovation but a framework that institutionalizes trust. It creates clarity on what “good AI” looks like: explainable, reversible, accountable, and under human oversight.

This regulatory foundation aligns naturally with our approach. Our work brings together AI engineering, risk and compliance, security, and digital sovereignty to create systems that are controlled by design and reversible by architecture. Evidence replaces claims. Observability replaces assumptions. Compliance becomes an enabler rather than an obstacle.

Organizations that embrace this approach gain a strategic advantage. They build AI that withstands scrutiny, adapts under stress, and scales with confidence.

From pilots to production: how value is created

AI creates value when it is safe to scale. The most successful organizations follow a disciplined path: one that balances experimentation with governance and engineering.

Internal productivity accelerates first. Tools such as GitHub Copilot and Claude Code shorten development cycles, reduce cognitive load, and free teams to focus on higher-value work. These gains scale further through governed, low-code AI environments that allow safe experimentation within enterprise guardrails.

Customer-facing and domain-specific AI solutions follow. Whether in code modernization, financial insights, customer operations, or legacy transformation, value emerges when AI is embedded into architectures built for lineage, explainability, and auditability.

Risk, compliance, security, and sovereignty play an accelerating role rather than a limiting one. Digital sovereignty does not eliminate tradeoffs between capability, cost, and control, but it ensures those tradeoffs are made consciously, not inherited by default. With clear exit strategies, model portability, and sovereign deployment options, organizations maintain freedom of action and avoid the lock-in that plagued earlier technology waves.

AI value becomes measurable not just through productivity but through reduced cost of failure, improved resilience, fewer incidents, better audit outcomes, and faster modernization cycles. Reliability transforms from a risk-mitigation exercise into a growth asset.

Organizations that industrialize AI structurally experience measurable shifts:

  • Reduced token and compute costs through reusable procedures
  • Lower incident frequency due to deterministic orchestration
  • Faster cycle times in engineering and operations
  • Improved audit outcomes through built-in traceability
  • Reduced vendor lock-in through model portability

Over time, these effects compound. AI ceases to be a cost center experiment and becomes a structural margin enhancer.


Three pillars of control

As AI becomes part of enterprise architecture, three pillars reinforce one another:

  • Resilience, ensuring continuity even when systems degrade.
  • Sovereignty, ensuring freedom of choice as dependencies evolve.
  • Reliability, ensuring trust as decision-making becomes augmented.

These pillars share the same DNA: distributed readiness, reversibility, and engineering discipline. They define the difference between organizations that deploy AI and organizations that control it.

Control is the architecture

Our mission-critical experience gives us a unique perspective. We have spent two decades ensuring that life-critical, financial, and high-integrity systems perform predictably, fail safely, and remain under human command. We bring that same philosophy to AI.

We design AI systems as engineered, observable, governed components — not black-box ambitions. We build with sovereignty in mind, so customers retain control across cloud, on-premise, air-gapped, and hybrid environments where required. We architect for reversibility, so decisions made today do not trap organizations tomorrow. And we ensure that intelligence supports human authority rather than undermining it.

The degree of autonomy granted to AI must always be proportional to its certified trustworthiness. Systems that advise tolerate uncertainty. Systems that execute autonomously require verifiable controls, drift detection, and escalation mechanisms. That calibration is not a technical detail. It is a board-level governance decision.

In a world where intelligence runs through everything, control is not a policy but the architecture. And reliability is not a feature, it is the precondition for trust, value, and leadership.

The question for leadership is no longer whether AI will mature. It will. The question is whether your organization matures architecturally before competitive disadvantage becomes structural.

AI has matured. It is becoming part of the enterprise landscape. Those who engineer it to behave reliably under pressure will define the next era of digital leadership.

We build AI the same way we build everything mission-critical: with discipline, transparency, and the people in charge.

Read more