AI & Intelligence
Governed AI capabilities for Al Manara Properties LLC: a provider-agnostic gateway, versioned models and prompts, permission-aware retrieval, document intelligence, explainable recommendations, anomaly detection — all advisory by default and fenced by the human approval boundary.
Provider profiles, routing by data class / region / capability / cost / latency / health, fallback order, quotas and data-boundary rules. Timeout, retry, circuit breaker and non-AI fallback are mandatory.
Versioned prompts, model policies, tools and schemas with owners, test results, effective dates, pinned provider snapshots and test → staging → production promotion with rollback.
Index sources with permission filters, retrieval traces that prove caller-permission enforcement before model context is built, citations with source versions, and a hard exclusion list.
Processing queue with per-field confidence, source coordinates and validation errors. Low-confidence or high-risk fields hand off to human maker-checker verification; originals stay authoritative.
Explainable lead scoring, next actions, property matches and valuation advice with confidence and main factors. Accept / edit / reject with feedback; out-of-tolerance valuations need review.
Duplicate invoices, unusual commissions, cash/parking exceptions and operational risk patterns — alerts cite contributing records and never reverse, block or accuse without a human decision.
Allowlisted, least-privilege, server-validated tools (read-only by default). Proposed actions are previewed before confirmation, executed idempotently, and high-risk decisions require human approval.
Golden datasets, groundedness and extraction accuracy tests, adversarial suites, plus budgets, rate limits and kill switches per environment / tenant / use case, with full run telemetry.
| Use case | Phase | Required control | Prohibited shortcut | Status |
|---|---|---|---|---|
| Lead extraction, scoring & next action | Phase 2 | Consent, explanation, quality/bias review and override | Automatic exclusion without review | Active |
| Listing copy & translation | Phase 2 | Grounded property facts, disclaimer and approval | Invented features or prices | Active |
| Document OCR / classification | Phase 2 | Confidence, coordinates and maker-checker | Silent master-data update | Active |
| Contract QA & RAG | Phase 2 | Governed content, value/source snapshot, citations, clause rules, legal review | Autonomous clause acceptance or silent source-data override | Active |
| Call transcription / coaching | Phase 2/3 | Recording policy, retention, access and factual grounding | Unreviewed employment action | Pilot |
| Valuation / recommendations | Phase 3 | Data quality, comparable evidence and range/confidence | Guarantee or unexplained value | Planned |
| AML anomaly / graph intelligence | Phase 3 | Explainable alert, evidence and investigator review | Autonomous compliance decision | Planned |
| Copilot / tool actions | Phase 3 | Tool allowlist, preview, least privilege and approval | Open-ended autonomous agency | Planned |
Static render mockup — no live data. Source: ERP Modern Development Specification v2.0 · Covers: AI-01, AI-02, AI-03, AI-04, AI-05, AI-06, AI-07, AI-08 (module hub).