NHS Sovereign AI Replacement Architecture
Chapters 1 – 9, 51 – 59 — The dependency problem, patient sovereignty, structural inversion, and the NMAI Core kernel.
Download the complete NHS Sovereign AI Replacement Architecture master record. The web pages provide section summaries; the PDF contains the full technical architecture.
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.
A Copilot-style deployment is not a simple assistant interface. It is a dependency chain:
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 data is a sovereign clinical asset. Its handling is governed by four pillars:
| Pillar | Requirement |
|---|---|
| Necessity | AI use must be clinically necessary, not merely convenient |
| Consent | Patient data boundaries must be respected and enforced |
| Audit | Every decision must be recorded and reviewable |
| Control | The 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 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?
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 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:
| Decision | Meaning |
|---|---|
| Permit | Task may proceed to approved model route |
| Block | Input must not reach the model |
| Warn | Caution markers detected; conditional proceed |
| Review | Human review required before model access |
| Escalate | Task must move to specialist process |
| Deny model access | No model contact permitted |
| Fail closed | Control failure occurred; system stops safely |
Only a permitted task can move to model execution. That makes the model subordinate to the route.
Systems Being Replaced:
| Existing Dependency | Replacement Component |
|---|---|
| Microsoft Word / Excel / Office | Open-source office layer (LibreOffice) |
| Copilot chat interface | Closed-domain NHS Clinical CUI |
| Raw prompt submission | NMAI deterministic input conditioning |
| Foreign transformer routing | UK-hosted model endpoint (vLLM) |
| Open semantic internet context | NHS-approved document boundary |
| Supplier telemetry / logs | NHS-controlled audit trail |
| Supplier licensing dependency | Open-source deployment stack |
| Unvalidated productivity claims | Route-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
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.
Scaling Stages:
| Stage | Purpose |
|---|---|
| Local prototype | Prove NMAI conditioning, ruleset logic, audit, and safe failure |
| Department pilot | Test bounded workflow tasks using synthetic or approved non-identifying data |
| Trust pilot | Test controlled operational deployment with governance oversight |
| Regional deployment | Test interoperability and wider operational scaling |
| National deployment | Deploy only after evidence, governance, cyber, procurement, and exit controls are approved |