A0 · Deterministic · financial services, telecommunications, public sector, retail, technology

AI governance plane

One gateway, one budget ledger, one release gate for every AI workload in the company.

Customer job: Let twenty product teams ship AI features fast while the company can answer what each one costs, which data it sends where, and whether it got worse.
Today: Each team holds its own provider keys; spend appears as one invoice line; nobody can stop a regressed prompt before customers see it.
Accepted result: A workload release that passed its evaluation gate, with its traffic attributed and inside its budget.

A shared platform sits in front of all model traffic. Product teams get approved models, keys and budgets through a portal; every call is routed, cached, scanned and attributed; releases pass an evaluation gate; finance sees cost per workload and per accepted result.

Google Cloud / primary provider

What carries each step.

Apigee policies enforce token quotas and routing; Model Armor screens prompts and responses; BigQuery holds the call lake for Looker Studio chargeback.

Data and model estate in one place: Gemini on Agent Platform, Agent Search and RAG Engine, AlloyDB and BigQuery, Cloud Run for everything else.

Component	Service	Evidence
	Person
	Cloud Run	prior, ga
	Identity Platform	prior, ga
	Apigee (AI gateway policies)	prior, ga
	Model Armor	prior, ga
	Firestore	prior, ga
	Gemini 3.8 Flash (Agent Platform)	source, ga
	Secret Manager	prior, ga
	BigQuery + Looker Studio	prior, ga
	BigQuery + Looker Studio	prior, ga
	Evaluation service (Agent Platform)	confirm
	Review console on Cloud Run	prior, ga
	External system

Customer view / judgement, not benchmark

Weigh what this customer values.

Remove providers that fail a hard constraint first: residency, data rights, identity estate, latency, operating owner, budget. Then weigh the rest. Scores are architect judgement on a 1-5 scale for the pattern as specified, recorded with a rationale. They are not benchmarks. A provider's consumption opportunity never changes a customer score.

Performance×1Does the default path meet the latency and quality budget for this workload?Time to value×1How quickly does a team reach a first accepted result with managed parts?Speed to execution×1How much undifferentiated wiring stands between the design and a running slice?Cost per accepted result×1How predictable and low is the all-attempt cost at the stated volume?Exit cost×1How cheaply can the workload move models, data or runtime later?Governance fit×1Do identity, residency, audit and policy controls exist without custom building?

Performance 5Time to value 5Speed to execution 5Cost per accepted result 4Exit cost 4Governance fit 3
Fastest path for teams already on Vercel; enterprise data-residency and audit needs confirmation.
Performance 5Time to value 4Speed to execution 4Cost per accepted result 5Exit cost 4Governance fit 4
Strong edge control plane at low cost; release gating is application work.
Performance 4Time to value 3Speed to execution 3Cost per accepted result 4Exit cost 4Governance fit 5
Enterprise-grade controls; Apigee configuration is a skill the platform team must hold.
Performance 4Time to value 4Speed to execution 3Cost per accepted result 4Exit cost 3Governance fit 5
Mature pattern on Azure; tier choice drives cost.
Performance 4Time to value 3Speed to execution 3Cost per accepted result 4Exit cost 3Governance fit 5
Native attribution for Bedrock traffic; non-AWS models need a separate gateway.
Performance 3Time to value 3Speed to execution 4Cost per accepted result 4Exit cost 5Governance fit 2
Portable and cheap; operational risk sits with your team.

Provider ledger / shown after the choice

What grows on Google Cloud.

These meters are the provider team’s consumption view. They explain the cloud workload; they never raise a customer score.

Gemini input and output tokens
Cloud Run vCPU-seconds and requests
Agent Search queries and index storage
AlloyDB or Cloud SQL instance hours
BigQuery slots or bytes scanned

Take it with you / generated from one specification

Exact views, and a running start.

Google Cloud views

SVG diagram for decks and documents
diagrams.net file editable, shared IDs
React Flow JSON for your own canvas
Mermaid flowchart

Every provider

Structurizr workspace C4 containers, one deployment view per provider
Journey sequence (Mermaid)

One-click template

No Google Cloud template yet.

The same answer contract runs on these runtimes today; its conformance suite is the porting checklist.

Vercel: anthropic/claude-sonnet-5.5 with gateway fallbacks
Cloudflare: @cf/meta/llama-3.3-70b-instruct-fp8-fast
Google Cloud: gemini-3.8-flash on the global endpoint

Customer ledger / assumptions you can move

Cost per accepted result.

Every attempt is charged, so failed runs count. Value comes only from accepted results. Illustrative assumption: tasks are workload releases per month; minutes saved are setup and review time per release. Model spend savings from caching and routing are a separate measured ledger.

Primary metric: Governed share of model traffic. Model calls routed through the plane with workload attribution, divided by all model calls observed on provider invoices.

Monthly value: €16,416
Monthly cost: €5,192
Net: €11,224
Cost per accepted result: €120

Eligible tasks per month60Share attempted with the system80 in 100Attempts accepted without rework90 in 100Minutes saved per accepted result240 minValue of an hour of that work€95.00Run cost per attempt€4.00Fixed monthly cost (hosting, review, support)€5,000

Experience / drawn from the specification

What the person sees, including no.

Screens are rendered from structured blocks in the catalog, so a refusal or an exception is designed before it happens. Values marked fixture are illustrations, not measurements.

Workload	Models	Budget	Gate
support-agent	claude-sonnet-5, gemini-3.8-flash	EUR 3,000	Pass
invoice-intake	gemini-3.8-flash	EUR 1,200	Pass
sales-copilot	claude-sonnet-5	EUR 2,500	Blocked

Fixture data.

Developer portaldesktop · default

Governed trafficfixture

Cache savingsfixture

Cost per accepted resultfixture

Chargebackdesktop · default

sales-copilot v1.8 blocked: citation precision 0.88 (gate 0.95).

12 failing cases attached
Previous release remains live

Release blockeddesktop · refusal

support-agent reached 80 percent of its monthly model budget.

Cache hit rate fell after the prompt change on Tuesday
Fallback model handled 12 percent of calls

Open workloadRaise budget

Fixture data.

Budget alertmobile · warning

Team workflow / who carries what

Every lane leaves evidence.

Lane	Responsibility	Evidence it leaves
Customer	Product team registers, ships, reads its cost and eval results	Workload record
AI	Models serve calls through the plane only	Routed-call share
System	Gateway enforces policy, budgets and guardrails	Blocked and redacted call log
Reviewer	Review board approves models and data classes	Decision records
Operator	Platform team and finance reconcile invoices against the lake monthly	Reconciliation report

Release evaluations

What must pass before anyone ships.

Governed traffic share
>= 0.95 of invoiced model spend traced to a workload
Monthly reconciliation of the lake against provider invoices.
Gate coverage
100 percent of production AI releases ran the gate
Pipeline audit.
Guardrail false positives
<= 1 percent of benign calls blocked
Labelled benign traffic sample.
Gateway overhead
p95 added latency <= 30 ms
Synthetic load.

Risks and controls

What would go wrong first.

Teams call providers directly.
Provider keys exist only in the plane; egress rules; invoice reconciliation.
Gateway outage stops all AI features.
Multi-region deployment, fail-open policy per workload class, tested runbook.
Spend growth read as success.
Report cost per accepted result beside spend; provider consumption is a separate ledger.

Image recipes / concept art only

When to generate, and when never to.

16:9 · gpt-image-2 or a current Gemini image model

Customer moment for a proposal cover or portal hero.

Cinematic editorial photograph, late afternoon light, shallow depth of field. Let twenty product teams ship AI features fast while the company can answer what each one costs, which data it sends where, and whether it got worse. The moment just after the accepted result: A workload release that passed its evaluation gate, with its traffic attributed and inside its budget. Real workplace, restrained palette of deep navy, warm paper and one cobalt accent. Natural expressions, no screens facing the camera. Concept art only. Do not render product logos, cloud icons, UI text or architecture labels; compose exact labels from the specification in code.

3:2 · gpt-image-2 (consistent characters across panels)

Team workflow storyboard for a workshop or executive briefing.

Five-panel storyboard, clean ink line with flat colour, consistent characters across panels. Panels follow the lanes customer, ai, system, reviewer, operator: (1) Product team registers, ships, reads its cost and eval results (2) Models serve calls through the plane only (3) Gateway enforces policy, budgets and guardrails (4) Review board approves models and data classes (5) Platform team and finance reconcile invoices against the lake monthly Each panel shows hands and decisions, not screens. Concept art only. Do not render product logos, cloud icons, UI text or architecture labels; compose exact labels from the specification in code.

Do not generate

Any image a customer will read for component names, data classes or provider services.

Do not generate. Export the compiled view for the chosen provider. A generated architecture image can invent services, misspell names and draw icons the vendor never published.

A design, not a deployment: no realization on this page carries a provider receipt. Generated from the pattern source at revision c43825efda38. Apache-2.0 for catalog content and exports; provider names are trademarks of their owners; Google Cloud icons retain Google rights.

Next pattern: Document intake pipeline