Provider-agnostic AI Gateway

Console for AI administrators and security: provider profiles, routing policies by capability / data class / region / cost / latency / health, fallback order, quotas and data-boundary rules. Modules only call internal capability contracts — never vendor-specific APIs.

AI-01AI-08

Provider health Model registry

Advisory by default — the gateway never bypasses domain controls AI responses returned through this gateway are advisory. Payments, postings, payroll, permission changes, legal commitments and compliance decisions always pass deterministic validation and human approval via the safe action boundary (AI-07). Timeout, retry, circuit breaker and a non-AI fallback are mandatory on every route; every fallback activation is recorded.

Requests today

12,483

▲ 6.2% vs 7-day avg

Fallback activations (24h)

31 provider swap · 3 deterministic path

Circuit breakers

1 half-open

PRV-VTX-03 · probing every 60s

p95 latency (generation)

1.9 s

SLO ≤ 3.0 s

Spend MTD

AED 18,540

of AED 25,000 budget · 74%

Provider profiles Adapter configuration

Profile	Capabilities	Region / residency	Max data class	p95 latency	Error rate 24h	Circuit	Health	Status
PRV-AZO-01 · Azure OpenAI UAE North · pinned snapshots only	Generation, embeddings	UAE (in-country)	Confidential	780 ms	0.4%	Closed	Healthy	Active
PRV-BED-02 · AWS Bedrock (Claude) me-central-1 · zero-retention agreement	Generation, tool use	UAE (me-central-1)	Confidential	1,120 ms	0.7%	Closed	Healthy	Active
PRV-VTX-03 · Google Vertex AI europe-west4 · translation models	Translation, generation	EU (cross-border)	Internal	2,940 ms	3.8%	Half-open	Degraded	Active
PRV-OCR-04 · Self-hosted OCR/IDP On-prem · Dubai HQ datacenter	Extraction, classification	On-prem (no egress)	Restricted-PII	3,410 ms	1.1%	Closed	Healthy	Active
PRV-EMB-05 · In-cluster embeddings bge-m3 · on-prem GPU pool	Embeddings	On-prem (no egress)	Restricted-PII	95 ms	0.1%	Closed	Healthy	Active
PRV-SBX-06 · Sandbox provider Synthetic data only · test env binding	Generation	UAE	Synthetic only	640 ms	0.2%	Closed	Healthy	Test only
PRV-DET-00 · Deterministic engine Non-AI fallback: rules, templates, manual queues	Fallback for all routes	On-prem (no egress)	Restricted-PII	12 ms	0.0%	n/a	Always available	Active

Routing policies & fallback order Routed by capability · data class · region · cost · latency · health

Route	Capability · use case	Data class	Primary	Fallback 1	Fallback 2	Non-AI fallback (mandatory)	Timeout / retry
RT-GEN-LST	Generation · listing copy EN/AR	Internal	PRV-AZO-01	PRV-BED-02	—	Approved template library	8 s · 2× backoff
RT-EXT-INV	Extraction · invoices & vouchers	Restricted-PII	PRV-OCR-04	None — data boundary	—	Manual keying queue (Finance)	30 s · 1×
RT-EXT-KYC	Extraction · KYC documents	Restricted-PII	PRV-OCR-04	None — data boundary	—	Maker-checker manual entry	30 s · 1×
RT-EMB-KB	Embeddings · knowledge base	Confidential	PRV-EMB-05	PRV-AZO-01	—	Defer & queue re-index	5 s · 3×
RT-TRN-LST	Translation · listing AR variants	Public	PRV-VTX-03	PRV-AZO-01	—	Human translation queue	10 s · 2×
RT-TOOL-CP	Tool use · staff copilot	Internal	PRV-BED-02	PRV-AZO-01	—	Decline with explanation	15 s · 1×
RT-GEN-SUM	Generation · lead & call summaries	Confidential	PRV-AZO-01	PRV-BED-02	PRV-DET-00	Show raw activity timeline	8 s · 2×

Data-boundary rulesPrivacy & classification

Data class	Allowed providers	Retention eligibility	Redaction before send
Public	Any active profile	Provider default	None
Internal	UAE + approved EU routes	Zero-retention preferred	Party contact details masked
Confidential	UAE-resident profiles only	Zero-retention required	IBAN, TRN, salary masked
Restricted-PII	On-prem only (PRV-OCR-04, PRV-EMB-05)	No provider retention	Emirates ID 784-19●●-●●●●●●●-●, passport masked in logs
Client money / trust	Not routable — deterministic only	n/a	n/a — see client money
AML case material	Not routable — excluded at index	n/a	n/a — see exclusion list

Quotas & budgets (per environment · tenant · use case)Cost observability

Scope	Monthly budget	Used MTD	Utilization	Rate limit	State
prod · Al Manara · document extraction	AED 9,000.00	AED 6,120.40		120 rpm	OK · 68%
prod · Al Manara · listing copy & translation	AED 4,500.00	AED 3,890.15		60 rpm	Warning · 86%
prod · Al Manara · campaign drafting	AED 2,000.00	AED 2,000.00		30 rpm	Capped — fallback path
prod · Al Manara · lead & call summaries	AED 5,500.00	AED 3,214.90		90 rpm	OK · 58%
prod · Al Manara · staff copilot tools	AED 3,000.00	AED 2,145.65		40 rpm	OK · 72%
staging · all use cases	AED 1,000.00	AED 168.20		20 rpm	OK · 17%
Total (production)	AED 24,000.00	AED 17,371.10	AI-08: at cap, noncritical requests use the lower-cost/fallback path or are deferred with an explanation

Spend by provider — July MTD

PRV-AZO-01

8,420

PRV-BED-02

5,315

PRV-OCR-04

3,180

PRV-VTX-03

1,140

PRV-EMB-05

485

Provider	Spend MTD (AED)	Requests	Avg cost / request
PRV-AZO-01 Azure OpenAI	8,420.00	148,210	0.057
PRV-BED-02 AWS Bedrock	5,315.00	61,882	0.086
PRV-OCR-04 Self-hosted OCR	3,180.00	9,412	0.338
PRV-VTX-03 Google Vertex	1,140.00	12,067	0.094
PRV-EMB-05 In-cluster embeddings	485.00	402,116	0.001
Total	18,540.00	633,687	—

Fallback & circuit events (last 24h)Audit explorer

17 Jul 2026 · 10:41 · RUN-88412 · corr. c9f2-4a

PRV-VTX-03 timeout ×3 → circuit opened, now half-open

Advisory translation request (RT-TRN-LST, Public) permitted fallback → rerouted to PRV-AZO-01. Approved alternative used and recorded. Recovered
17 Jul 2026 · 09:12 · RUN-88231 · corr. b7e1-9d

Campaign drafting quota capped (100%)

Noncritical request deferred with explanation shown to Layla Hassan; queued for 01 Aug window. Deferred
16 Jul 2026 · 22:03 · RUN-87914 · corr. a1c8-2f

PRV-OCR-04 nightly maintenance window

RT-EXT-INV has no external fallback (data boundary) → 42 documents routed to manual keying queue. Deterministic path
16 Jul 2026 · 15:47 · RUN-87655 · corr. e4d0-7b

PRV-AZO-01 HTTP 429 burst

Retry with backoff succeeded on attempt 2; no fallback needed. Recovered
16 Jul 2026 · 11:20 · corr. f2a9-1c

Kill switch drill — staging

Vikram Nair executed quarterly kill-switch test on staging copilot; all routes declined with explanation in 1.2 s. Drill

Risky action — disable provider profile

Approval route: AI administrator (maker) → Security officer (checker). Segregation of duties enforced — Vikram Nair cannot approve his own change. Recovery: re-enable is a new approved activation, see version lifecycle.

Static render mockup — no live data. Source: ERP Modern Development Specification v2.0 · Covers: AI-01 (provider-agnostic gateway, routing, fallback, data boundaries), AI-08 (budgets, rate limits) · §12.3 provider abstraction and controls.