01 — Architecture Overview

NHS Sovereign AI Replacement Architecture

Chapters 1 – 9, 51 – 59 — The dependency problem, patient sovereignty, structural inversion, and the NMAI Core kernel.

Full Master Architecture PDF

Download the complete NHS Sovereign AI Replacement Architecture master record. The web pages provide section summaries; the PDF contains the full technical architecture.

Download PDF
In This Section
Overview The Dependency Problem Patient Sovereignty The Structural Inversion NMAI Core Replacement Stack Final Position
Overview

The NHS does not need to replace one chatbot with another chatbot. The real problem is not which generative AI system should be used. The real problem is dependency.

When AI is embedded into office tools, documents, email, meetings, summaries, search, and administrative workflow, the NHS must ask a more serious question: Where does the NHS workflow go when AI is added?

A foreign SaaS AI dependency chain does not only process prompts. It also involves identity systems, document context, file names, metadata, staff instructions, operational patterns, logs, telemetry, model inference, licensing control, administrative access, service continuity, and supplier-controlled updates.

Governing Rule
No NMAI Core decision, no model call.
The Dependency Problem

A Copilot-style deployment is not a simple assistant interface. It is a dependency chain:

NHS staff user
Microsoft 365 / office environment
AI assistant interface
Cloud processing
Model inference
Telemetry and logging
Supplier licensing
Supplier update control
Supplier continuity risk

Once staff begin relying on that chain for routine document, administrative, clinical-support, and operational work, the supplier becomes embedded inside the workflow. That creates public-sector risk.

The NHS must retain control of:

  • Patient-data boundaries
  • Workflow routing
  • Prompt exposure
  • Model access
  • Audit evidence
  • Output status
  • Human review
  • Emergency stop
  • Supplier exit
  • Service continuity

This is not a productivity issue alone. It is an infrastructure-sovereignty issue.

Patient Sovereignty as the Governing Principle

Patient data is a sovereign clinical asset. Its handling is governed by four pillars:

PillarRequirement
NecessityAI use must be clinically necessary, not merely convenient
ConsentPatient data boundaries must be respected and enforced
AuditEvery decision must be recorded and reviewable
ControlThe NHS must retain operational control, not delegate it to suppliers

Public-sector control must come before model convenience. The architecture places NHS governance at every layer, ensuring patient-data sovereignty is never compromised by supplier integration.

Closed Medical Environment Principle:

  • No open internet context in clinical processing
  • NHS-approved documents only
  • NHS templates and policies as the semantic boundary
  • Controlled clinical and administrative context
  • Semantic-noise reduction through domain locking
The Structural Inversion

The NHS Sovereign AI Replacement Architecture reverses the model-first route.

A model-first system asks: What can the model do with this input?

This architecture asks first: Should this input reach a model at all?

Proposed Route
Open-source office layer
Closed-domain NHS Clinical CUI
NMAI Core deterministic input conditioning
Audit and decision state
UK-controlled model endpoint (only where permitted)
NHS-controlled output, review, and governance

The model is not first. The model is not the authority. The model is not the governance layer. The model is a post-conditioning execution module.

It may draft, summarise, rewrite, restructure, format, or build a document only after NMAI Core has permitted the task.

NMAI Core — The Deterministic Input-Conditioning Kernel

NMAI Core is not a chatbot. It is not the transformer. It is not the GPU layer. It is the pre-transformer control function that governs every model-access decision.

NMAI Core performs:

  • Input conditioning
  • Task classification
  • Marker detection (patient data, clinical risk, safeguarding, legal, governance)
  • Patient-data boundary checking
  • Clinical-risk marker checking
  • Safeguarding marker checking
  • Legal and complaint marker checking
  • Governance marker checking
  • Ruleset decision
  • Audit-reference creation
  • Model-access decision

NMAI Core can return:

DecisionMeaning
PermitTask may proceed to approved model route
BlockInput must not reach the model
WarnCaution markers detected; conditional proceed
ReviewHuman review required before model access
EscalateTask must move to specialist process
Deny model accessNo model contact permitted
Fail closedControl failure occurred; system stops safely

Only a permitted task can move to model execution. That makes the model subordinate to the route.

This Is Not GenAI Governance
A GenAI governance wrapper starts with the model and then attaches policy, filters, logs, prompts, or review controls around it. This architecture starts before the model. It prevents the model from becoming the first operational boundary. That decision must be deterministic, auditable, controlled, and reviewable.
Replacement Stack Overview

Systems Being Replaced:

Existing DependencyReplacement Component
Microsoft Word / Excel / OfficeOpen-source office layer (LibreOffice)
Copilot chat interfaceClosed-domain NHS Clinical CUI
Raw prompt submissionNMAI deterministic input conditioning
Foreign transformer routingUK-hosted model endpoint (vLLM)
Open semantic internet contextNHS-approved document boundary
Supplier telemetry / logsNHS-controlled audit trail
Supplier licensing dependencyOpen-source deployment stack
Unvalidated productivity claimsRoute-level productivity measurement

Why This Matters for the NHS:

  • Block unsuitable inputs before inference
  • Prevent raw workflow material going straight to a model
  • Show warning and review states
  • Preserve human approval at every decision point
  • Audit model-access decisions with full provenance
  • Avoid external SaaS fallback paths
  • Continue working when generation is disabled
  • Replace model endpoints without redesigning the whole system
Final Architecture Position

This is not a chatbot replacement plan. It is not a model-selection guide. It is not a supplier-swap proposal. It is a deterministic NHS workflow-control architecture with optional model execution after permission.

NMAI Core governs access. The model executes only after permission. The model is replaceable. The route is sovereign.

Scaling Stages:

StagePurpose
Local prototypeProve NMAI conditioning, ruleset logic, audit, and safe failure
Department pilotTest bounded workflow tasks using synthetic or approved non-identifying data
Trust pilotTest controlled operational deployment with governance oversight
Regional deploymentTest interoperability and wider operational scaling
National deploymentDeploy only after evidence, governance, cyber, procurement, and exit controls are approved
← Home 02 Implementation →