Board asks for a number, analysts open 14 dashboards: Get one governed answer, in Arabic, in under 5 seconds
A conversational data platform your CRO, CFO and internal audit can defend in a SAMA review. Runs in-country, honors row-level security end-to-end, and answers questions on your core banking, cards, treasury and AML data with a published Arabic NL2SQL accuracy score, not a marketing number.
- Arabic NL2SQL accuracy
- 94.6%Arabic NL2SQL accuracy
- median answer
- < 5smedian answer
- banking scenarios in eval
- 1,200+banking scenarios in eval
- data leaves Saudi Arabia
- 0data leaves Saudi Arabia
In Saudi Arabia
What a banking reviewer checks before an NL2SQL bot goes live
Every question, Arabic or English, runs through the same governed semantic layer. This is the evidence the sector authority expects before the bot touches production numbers.
SAMA
Metric definitions used by the bot match the definitions البنوك reporting already uses.
Semantic layer built from the existing governed metric catalogue, not invented per question.
Metric-to-source mapping document reviewed by the reporting owner.
SDAIA, PDPL
Row and column-level entitlements applied before an Arabic or English question can return a number.
Identity resolved and entitlements applied at the semantic layer, ahead of query generation.
Entitlement test log showing the same question answered differently per role.
NCA ECC
Read-only execution path, with SQL validated and cost-checked before it runs.
Query validator gate ahead of execution, no write path exposed to the bot.
SQL audit trail attached to every answer, reviewable line by line.
Where it runs
The warehouse connection and the semantic layer run inside Saudi Arabia; the bot never ships raw data outside that boundary to generate an answer.
Arabic, as it is spoken here
Arabic business questions, including mixed Arabic-English phrasing common in Saudi teams, resolve against the same catalogue without a translation step.
Published Arabic NL2SQL evaluation harness, not a demo number
A proof of concept on your sanitized data
SAMA-aligned architecture reviewed by ex-CBUAE / SAMA advisors
Deployed on HUMAIN, Oracle OCI Jeddah, or your private VPC
Questions your team actually asks, answered with governed SQL
In Arabic, English or a mix, the bot converts each question into SQL over a unified semantic layer.
Show branch-level NPL trend for the last four quarters, split SME vs corporate.
أفضل 20 عميلاً من حيث نمو الأصول تحت الإدارة في جدة هذا الشهر
Which AML alerts breached the 24-hour SLA in Riyadh region, ordered by risk score?
قارن نسبة التحويل الرقمي بين قنوات الأفراد والشركات هذا الربع مقابل الربع الماضي
Cost-to-income by business line, month over month, with drivers.
أعرضي محفظة القروض المتعثرة عالية القيمة التي لم تتم مراجعتها منذ 45 يوماً
Where teams use NL2SQL in nl2sql for saudi banks
The board pack, generated in the meeting
CEO or CRO changes a filter mid-slide, NPL by segment, LCR sensitivity, cost-to-income by region, and the bot returns a governed, board-defensible number in seconds. Same definitions your finance team publishes.
AML & fraud analyst, 3× faster triage
Analysts ask the bot to surface breached SLAs, high-risk typologies and PEP exposures, every query is scoped by RBAC, every answer cites the source tables, every action lands in an immutable log ready for SAMA and internal audit.
Branch manager Q&A on WhatsApp
Regional managers ask NPS, cross-sell, cash-in/out and productivity questions from WhatsApp in Saudi Arabic. Row-level security ensures every manager sees only their own portfolio, never a peer's numbers.
Certified semantic layer + complete enterprise security layer
We never wire the bot straight to your warehouse, every question runs through a governed semantic layer and a security stack that respects every row and column.
Certified semantic layer
- Certified metrics: NPL, CASA, NIM, LCR, NSFR, cost-to-income, RoRWA, revenue per FTE.
- Semantic layer over dbt / Cube / LookML, one definition per KPI, everywhere, forever.
- Pre-built connectors for core banking (T24, Flexcube, ICBS), cards (Way4, TSYS), treasury, AML (Actimize, FOCAL) and CRM.
- Graceful failure: if the bot is not sure, it asks a clarifying question, it never returns a confident wrong number.
Complete enterprise security layer
- In-country deployment on HUMAIN, Oracle OCI Jeddah, or your private VPC, zero customer data egress.
- Row- and column-level RBAC enforced INSIDE the query, not just in the dashboard.
- Automatic PII masking on IBAN, national ID, card PAN, mobile, with per-role reveal rules.
- Prompt-injection defenses, safe-SQL sandbox, and rate limits per user, per role, per table.
- Immutable audit (WORM), question, SQL, tables, rows returned, user, timestamp, retained per SAMA & PDPL.
- ISO 27001-scoped platform; NCA ECC-2 aligned; independent pen-test report on request.
Questions Saudi leaders ask before rollout
Will it actually work on OUR core banking schema, not only a demo one?
That's exactly why we start with a proof of concept on a sanitized copy of your actual schema, T24, Flexcube, ICBS, or in-house. You see the accuracy score on your questions, not ours. If it doesn't hit the threshold your team agrees to up-front, you don't pay for the pilot.
Does the bot enforce SAMA data-residency requirements?
Yes. Every deployment runs in-country, HUMAIN, Oracle OCI Jeddah, or your own VPC. Model weights, prompts, embeddings, audit logs and every SQL result stay inside Saudi Arabia. We ship the SAMA and PDPL architecture diagram your risk team can hand straight to the regulator.
We already pay for a Copilot / BI-vendor assistant. Why not use that?
Bundled assistants score well in demo videos and collapse to 10–31% on real enterprise schemas (public benchmarks, 2025). They also send queries through cross-border AI endpoints, a PDPL exposure for banks. Our platform is Arabic-first, SAMA-defensible, runs in-country, and publishes its accuracy on your data. Compare the eval harness side by side and decide.
What happens when the bot misunderstands an Arabic question?
It asks a clarifying question. We built failure-mode transparency in from day one, the bot never returns a confident wrong answer on financial data. Every uncertain query is logged, reviewed weekly by our Arabic linguistics team, and fed back into the model.
How is accuracy actually measured?
We ship a banking-specific Arabic NL2SQL evaluation harness: 1,200+ Saudi banking scenarios (NPL, CASA, treasury, AML, cards, digital) scored on execution accuracy, semantic accuracy, and hallucination rate. Every release runs the harness on YOUR sanitized schema and the score is shared with your data team, not just marketing.
See your first banking question answered on YOUR sanitized data, 20-minute working session, no PoC fees.
No fees, no procurement commitment, no data leaving Saudi Arabia. You walk out of the session with an accuracy evaluation report your team owns, whether you buy or not.