Illustrative use cases

How a ministry, a central bank, and a grid operator would run it.

Fictional institutions, realistic data. Each scenario shows the sensitive information involved, what fails without a control layer, and exactly how the Alliance Intelligence Control Plane mediates the work.

The institution

Ministry of Economy and Industry, Republic of Novara

The ministry drafts industrial policy, reviews investment proposals, and answers parliamentary inquiries under statutory deadlines. Analysts spend most of their time locating and reconciling documents rather than forming judgements.

Staff: 2,400 civil servants
Analysts in scope: 38 policy analysts
Document corpus: 1.2M pages, 4 languages
Deployment: Sovereign cloud, national territory

Sensitive data in scope

Draft legislation and policy memorandaRestricted — pre-decisional
Leakage would move markets and pre-empt cabinet deliberation.
Investment proposals and company filingsConfidential — commercial
Contains trade secrets submitted under legal confidentiality.
Inter-ministerial correspondenceRestricted
Attribution of positions to named officials.
National statistics before publicationEmbargoed
Early disclosure is a market-abuse and legal exposure event.
Published law, treaties, and public consultationsPublic
Safe for external frontier models.

Without a control layer

Analysts paste restricted excerpts into consumer AI tools to meet deadlines.
No record of which documents informed a given policy recommendation.
Answers differ between departments because each keeps its own document set.
Legal counsel cannot certify how an AI-assisted brief was produced.

Deployment configuration

Mode: Sovereign cloud inside national territory
Keys: Held by the ministry's CISO office
Models: Local model for restricted, external for public
Egress: Denied by default; exceptions named and logged

How a single request is handled

Every request passes the same institutional path.

Identity

An analyst authenticates through the ministry directory. Their department, clearance, and current case assignment become policy inputs.

Classification check

The request touches a pre-decisional memorandum. The control plane tags the request Restricted before any model sees it.

Policy decision

Policy forbids Restricted content leaving national infrastructure. External providers are removed from the candidate set.

Retrieval scope

Retrieval is limited to the corpora this analyst may read — the embargoed statistics set is excluded, not merely ranked lower.

Model routing

The most capable permitted model runs on sovereign compute. Public-only sub-questions may route to a frontier model separately.

Human review

Because the output feeds a ministerial brief, it enters review with citations to source paragraphs before release.

Evidence

Identity, policy decision, retrieval scope, model version, reviewer, and output are written to an immutable audit record.

Days → hours

Parliamentary inquiry turnaround

Illustrative target for the pilot workflow.

Restricted content sent to external providers

Enforced by routing policy, not staff discipline.

100%

AI-assisted briefs with source lineage

Every output reconstructable for internal audit.

Guardrails that hold in this scenario

Refusal rather than downgrade when no permitted model can handle the classification.
Embargoed statistics unreachable until their publication timestamp passes.
Ministerial-brief outputs cannot be released without a named human reviewer.

These institutions are fictional. Names, datasets, and figures are illustrative constructions used to describe intended system behaviour — they are not customers, deployments, or performance guarantees.

Next step

Bring your own workflow to a 90-day pilot.

We run the same path against one real institutional workflow, inside your infrastructure and your classification rules.

Security Architecture