EAINE Architecture Blueprint Builder

1 Choose your perspective

ExecutiveOutcome, affected parties, ownership, risk, investment, operating cost and the scale, adapt or stop decision.ArchitectureDomains, trust zones, shared components, interfaces, deployment, identity and failure boundaries.DeliveryAI SDLC stages, evidence, release digest, evaluation, operations, rollback and a 90-day path.

2 Select a governed industry scenario

Banking and insuranceSuggested pilot: Policy-grounded operations or decision-support assistantHealthcare and life sciencesSuggested pilot: Bounded administrative or professional-support workflowTelecom and manufacturingSuggested pilot: Maintenance or service-operations assistantPublic sectorSuggested pilot: Internal caseworker or knowledge assistantSaaS and digital productsSuggested pilot: Bounded customer-facing AI capabilityConsulting and servicesSuggested pilot: Engagement preparation or quality-review assistant

Scale AI-assisted decisions without weakening fairness, model risk control, explanation or audit readiness.

3 Personalise the decision document use non-confidential labels

Organisation or programmePrepared forUse-case labelBusiness outcomeOperating contextConstraints and assumptions one per line

4 How will AI be used? choose every option that applies

Choose the situations a reader would recognise. Each selection adds safeguards to your blueprint; one selection never removes the requirements of another.

We use AI inside our organisationChoose this for employee assistants, internal knowledge search, analysis, or other AI that is not offered to customers.Example: an internal policy or engineering assistant.Framework reference: P-INT · Internal AIAI is built into a product or serviceChoose this when customers, partners, or employees experience AI as part of a delivered application or workflow.Example: AI-assisted drafting inside a SaaS product.Framework reference: P-EMB · Embedded AIAI may materially affect peopleChoose this when AI may influence access, rights, safety, employment, finance, healthcare, public services, or another consequential outcome.Example: decision support for credit, care, benefits, or hiring.Framework reference: P-HI · High-impact AIAI can take actions through toolsChoose this when AI can call systems, change records, send messages, spend money, or perform another action—not only provide an answer.Example: an agent that updates a service ticket after approval.Framework reference: P-AGT · AgentWe train or fine-tune our own modelChoose this when your organisation changes model behaviour using its own training data, features, or fine-tuning process.Example: a specialised model trained on permissioned domain data.Framework reference: P-MD · Custom modelWe rely on an external AI providerChoose this when another organisation supplies the model, AI service, platform, or API used by the solution.Example: a commercial model accessed through an external API.Framework reference: P-TP · Third-party AI

5 What could happen if this AI is wrong, unavailable, or misused?

Choose a planning starting point based on the most serious credible consequence. This does not calculate or declare a formal risk classification.

Limited consequenceUse as a planning starting point when failure is contained, reversible, and unlikely to materially affect people or essential operations.Planning reference: EAINE-R1 · R1 — Lower consequenceMaterial business consequenceUse when failure could meaningfully affect customers, business operations, privacy, security, cost, or reputation.Planning reference: EAINE-R2 · R2 — Material consequenceHigh consequence for people or operationsUse when AI may materially affect rights, safety, access, livelihoods, regulated decisions, or critical business services.Planning reference: EAINE-R3 · R3 — High consequenceCritical safety, rights, or systemic consequenceUse when failure could cause severe or difficult-to-reverse harm and demands qualified independent validation and executive authority.Planning reference: EAINE-R4 · R4 — Critical consequence

6 Set the human authority boundary

Human decidesAI advises; an accountable human makes the consequential decision.Human approves before actionAI may prepare an action; a named human authorises execution.Human supervises with interruptAI acts under live supervision with a tested stop path.Automated within a pre-approved boundaryAutomation is limited by purpose, permission, consequence, time and transaction limits.

7 Where should the solution run, and how should it recover?

Choose the closest starting point. Architecture review must still confirm data location, provider, resilience, and recovery requirements.

Operate in one approved regionChoose this when approved data and providers fit one region and the service can safely wait for regional recovery.Architecture reference: DEP-001 · Single-region bounded deploymentKeep a secondary region ready for recoveryChoose this when continuity requires a pre-approved backup region with equivalent controls.Architecture reference: DEP-002 · Active-passive regional recoveryKeep enterprise systems private while using an external modelChoose this when enterprise data and tools stay private while minimised prompts use an approved external model.Architecture reference: DEP-003 · Private runtime with external provider gateway