SYSTEMS.

AI entered critical national infrastructure faster than the governance that was supposed to bound it. In CNI environments — energy, defence, finance, health, telecoms — the institutional accountability a regulator, a board, or an investment committee expects is not met by a capability with a compliance document attached. It is met by a system where the risk management, human oversight, and audit trail are structural properties of the build. The Systems service exists to invert that order: governance first, capability second.

Governed AI built to the high-risk standard by default. AI systems for CNI operating environments engineered against the EU AI Act's high-risk requirements — risk management (Article 9), human oversight (Article 14), and data governance — as the build specification, not a post-build compliance check.

Human-in-the-loop architecture. Oversight designed as a structural property of the system: defined points of human review, decision records that are reviewable after the fact, and named accountability roles that are not nominal sign-off steps.

Compliance-by-architecture, not retrofit. The governance, audit trail, and documentation structure built as part of the system from inception, so an assessment finds evidenced governance rather than a capability with a compliance annex.

Sector-calibrated deployment. AI built for the specific CNI operating environment it serves — its regulators, its resilience expectations, its data-governance obligations, and its procurement scrutiny.

Sovereign private deployment. For CNI operators where classified data, regulatory constraints, or institutional security policy prohibit cloud-based inferencing: hardware and software installed within the client's physical perimeter, running a private large-language model closed from public networks and fine-tuned on the client's own operational data. Satisfies ISO 27001:2022 network-segregation controls (Annex A.8.20 to A.8.22).

Governance-first scoping

A build begins by defining the system's governance perimeter before its feature set: where human oversight must sit, what must be logged and reviewable, what the accountability structure is, and which regulatory instruments bind the deployment.

Architecture to the high-risk standard

The system is architected against the EU AI Act's high-risk requirements as the build specification: risk management (Article 9), human oversight (Article 14), and data governance. This is the default, regardless of whether the specific system is formally classified as high-risk — because the standard a CNI environment will hold it to is the high-risk standard. EU AI Act Article 50 transparency obligations take effect 2 August 2026. High-risk obligations were provisionally deferred in the 7 May 2026 Digital Omnibus agreement — standalone Annex III to 2 December 2027, product-embedded Annex I to 2 August 2028 — but this deferral is subject to formal adoption expected July 2026. Build to the substance, not the date.

Human-in-the-loop construction

Oversight is built as a structural property: defined control points, decision records that are reviewable after the fact, and accountability roles with real authority. The method that governs this component is the DI Systems Flywheel: humans are trained first, their operational knowledge is codified into the AI, and AI outputs are fed back into human understanding. The AI augments the humans; the humans govern the AI. The cycle runs continuously.

Evidenced delivery

The system is delivered with its governance evidenced: risk-management records, oversight design, data-governance documentation, and audit-trail architecture are all part of the deliverable. The test a CNI environment applies is whether the evidence exists at the point of assessment, not at the point of request.

Systems maps to BEACON, the Execution Engine — one of the six pillars of the DIRECT framework. BEACON builds governed AI to human-in-the-loop and EU AI Act risk-management standards from inception.

Scope the governance before the feature set. A system whose oversight, logging and accountability are designed after its capabilities are built is a governance retrofit, not a governed system.

Build to the high-risk standard by default, not only where legally mandated. The EU AI Act's high-risk timeline has moved (see below), but the standard a CNI environment will hold an AI system to has not.

Treat the EU AI Act timeline as live and moving, and build to the substance rather than the date. As of 7 May 2026, EU lawmakers provisionally agreed to defer high-risk obligations — standalone Annex III to 2 December 2027, product-embedded Annex I to 2 August 2028 — but this is subject to formal adoption expected July 2026. Article 50 transparency obligations take effect 2 August 2026 as originally scheduled.

Make human oversight architectural, not nominal. A sign-off step is not human-in-the-loop. Oversight that survives institutional scrutiny is designed into the system, not added to it.

Deliver the system already evidencing its governance. The test a CNI environment applies is whether the evidence exists at the point of assessment, not at the point of request.