Ask any question. In Arabic. In Saudi Arabia.
Sawal turns Arabic and English business questions into governed SQL against your warehouse, ERP and CRM. Hosted on HUMAIN in Saudi Arabia, aligned to PDPL, with a certified semantic layer, row-level security and agentic actions.
In Saudi Arabia
Asking your data in Arabic, under Saudi rules
A question in Arabic that turns into a query against a core system is an access event. It has to obey the same rules the underlying data obeys.
NDMO data management controls
Answers reconcile to the governed definition, not to an ad hoc query.
A semantic layer with owned metrics, so the same question returns the same number.
Metric definitions with owners, versioned alongside the data.
SDAIA, PDPL
A user must not see through the assistant what they cannot see in the source.
Row-level security enforced at query time against the same identity.
Access test pack showing identical results with and without the assistant.
Where it runs
Hosted in-country against your warehouse, ERP and CRM. Questions, queries and results stay in Saudi infrastructure.
Arabic, as it is spoken here
Arabic is the primary interface: Saudi business terminology, Hijri and Gregorian dates side by side, and amounts in ر.س.
Built for the Saudi enterprise
Arabic-first NLU
MSA, Gulf, Egyptian, Levantine and code-switched Arabizi, trained on Saudi business vocabulary, not translated from English.
Governed semantic layer
Certified metrics from dbt, LookML, Cube, every answer maps to an approved metric. No metric drift across teams.
Row-level security honored
Sawal reads Postgres RLS, Snowflake row policies, Okta groups, and enforces them on every query. It never bypasses your RBAC.
Agentic actions
Approve a PO in SAP, refresh a forecast in Databricks, open a Jira ticket, with human-in-the-loop approval on every write.
In-country sovereignty
Hosted on HUMAIN by default. PDPL, SAMA, NDMO and Vision 2030 aligned. Air-gapped and Bilytica GPU Cloud deployment options.
Published evaluation harness
1,200+ Saudi-enterprise scenarios scored on execution, semantic accuracy and hallucination rate. Auditable, not just claimed.
Six capabilities you won't find combined anywhere else
Enterprise-grade differentiators, purpose-built for the Saudi context, no marketing spin.
True Arabic dialect coverage
MSA plus Gulf, Egyptian, Levantine and code-switched Arabizi, trained on Saudi enterprise vocabulary, not translated benchmarks.
Governed semantic layer, imported
Bring your certified dbt / LookML / Cube metrics as-is, every answer maps to an approved metric. No metric drift across teams.
Row-level security honored end-to-end
Postgres RLS, Snowflake row policies and Okta groups are enforced on every generated query. Your RBAC is never bypassed.
Agentic write-back with approvals
Approve a PO in SAP, refresh a forecast in Databricks, open a Jira ticket, human-in-the-loop on every write. Not just read-only Q&A.
In-country sovereignty by default
Hosted on HUMAIN inside Saudi Arabia, aligned to PDPL, SAMA, NDMO and Vision 2030, with air-gapped and Bilytica GPU Cloud options.
Auditable accuracy, not marketed accuracy
1,200+ Saudi-enterprise NL2SQL scenarios scored on execution, semantic accuracy and hallucination rate, results shared under NDA, not just claimed on a website.
From an Arabic question to a governed answer
Ask in Arabic or English
Web, Teams, WhatsApp or voice. Sawal detects dialect and normalizes intent using the Saudi business glossary.
Resolve against the semantic layer
Sawal maps entities to certified metrics, no ambiguous SUM(revenue) across five teams.
Generate governed SQL
SQL is generated, validated, RLS-enforced, and dry-run-costed before execution against your warehouse.
Answer with citation
Result + chart + link to the exact SQL and semantic metric that produced it. Trace, not trust.
Agents that take action, not just answer
Sawal understands, queries, then does, with human approval on every write.
Approve
'اعتمد طلب الشراء #4821 مع تسجيل السبب'، Sawal writes to SAP after human approval.
Refresh
'حدّث توقّع الطلب لـ Q3'، Sawal triggers a Databricks job and posts results back to the chat.
Ticket
'افتح تذكرة Jira على أعلى 3 انحرافات'، Sawal creates linked tickets with the source SQL attached.
Warehouses, ERPs, CRMs and the Saudi stack
Warehouses
ERP
CRM
Saudi stack
Semantic layer
Agent tools
6 places Sawal beats a dashboard
Executive queries in Arabic
The CEO asks بالعربية: 'كم مبيعات الفرع الشرقي هذا الربع مقارنة بالسابق؟'، a certified answer in 3 seconds, cited to the finance metric.
Field ops on WhatsApp
A branch manager: 'كم مخزون المنتج X عندي؟'، Sawal answers in the branch's local WhatsApp, respecting branch-level RLS.
Finance close acceleration
Controllers ask 'variance vs budget by cost center' and Sawal returns cited numbers plus a Jira ticket for outliers > 5%.
Supply chain what-if
'لو تأخّرت شحنة X أسبوعاً، وين ينكسر المخزون الأمان؟'، Sawal simulates and posts back to Databricks with approval.
Call-center conversational BI
Agents ask about a customer's LTV and last 3 tickets in a single Arabic sentence, no dashboard-switching.
Government sector KPIs
Ministry analysts query Vision 2030 program KPIs by region, Sawal enforces classification and residency by policy.
Five production deployments, with a published harness and adversarial testing
Each case documents the research hypothesis, the evaluation harness, accuracy and hallucination metrics, the red-team suite, and the business outcome, no marketing spin.
From a 30,000-table Snowflake warehouse to Arabic questions answered under 4 seconds, with SAMA-grade auditability
Tier-1 Saudi commercial bank (Riyadh HQ, 4,300 seats) · 14 weeks kickoff → production
The bank's Treasury and Risk desks were blocked on BI backlog: every liquidity, FX, or credit-exposure question took 3–9 working days through a shared analyst pool. Bilytica delivered a Sawal deployment against the bank's 30k-table Snowflake warehouse with a certified semantic layer, honored Snowflake row-access policies end-to-end, and shipped a published evaluation harness of 1,842 Arabic + English NL2SQL scenarios reviewed by the bank's own Treasury quants.
Challenge
- •30,412 tables across 47 Snowflake schemas, no single analyst had a mental model of the warehouse.
- •Treasury vocabulary is Arabic-first (سيولة، انكشاف، مرابحة، سواب) with heavy code-switching to English at column level.
- •SAMA required every automated query on customer-facing exposure to carry a full audit trace including the resolved semantic metric.
- •Any Row-Access Policy bypass, even by an authorised model, was a hard release blocker.
R&D protocol
- •Written R&D hypothesis: hierarchical schema linking on a 30k-table warehouse outperforms flat retrieval by ≥8 points of execution accuracy on Arabic questions.
- •Built a two-stage schema linker: BM25 + Arabic-aware bi-encoder over table+column descriptions, then a listwise reranker fine-tuned on 12,400 bank-authored (question → tables) pairs.
- •Constructed a certified semantic layer of 214 Treasury/Risk metrics from the bank's existing dbt project, every generated query resolves through it, never against raw columns.
- •Held-out evaluation set of 1,842 scenarios split 70/30 across Arabic (MSA + Gulf) and English, gold SQL written by the bank's Treasury quants and never shown during training or few-shot selection.
- •Row-Access Policy enforcement was tested with a 380-case adversarial harness that attempts implicit filter removal, cross-schema joins, and prompt-injected role escalation, release blocker on any successful bypass.
- •Hallucination guards: (a) refuse if schema-linker top-1 confidence < 0.62, (b) dry-run every candidate SQL against Snowflake's EXPLAIN and reject queries scanning > 250 GB, (c) require the answer to cite the exact certified metric ID.
Evaluation harness
Gold set: 1,842 scenarios: 620 liquidity & funding, 540 credit exposure, 410 FX & rates, 272 regulatory (SAMA returns). Gold SQL written by 6 Treasury quants; every scenario reviewed by a second quant and a data-governance officer.
- →Execution accuracy: 94.8% overall (Arabic 93.9%, English 96.1%)
- →Semantic exact-match (against certified metric IDs): 91.2%
- →Hallucination rate (unsupported claim in answer text): 0.6%
- →Refusal rate on low-confidence (calibrated target 4–7%): 5.3%
Red team: 380-case adversarial harness targeting RLS bypass, cross-tenant leakage, prompt-injected role escalation and metric spoofing. 0 successful bypasses at go-live; 3 pre-release regressions caught and remediated during hardening.
Outcomes
- Treasury question TAT collapsed from 3–9 working days to a median of 3.4 seconds (−96%).
- The BI analyst pool was redirected from ad-hoc queries to metric certification and model monitoring.
- Two consecutive SAMA IT-risk audits passed with zero findings on the NL2SQL surface.
- The bank adopted the harness as an internal standard, new dbt metrics ship only after a scenario is added and passes.
Frontline stores and call centres asking Gulf-Arabic Customer 360 questions on WhatsApp, with in-country PDPL residency
Top-3 Saudi telecom operator (retail, enterprise & fibre) · 18 weeks kickoff → production
1,900 store agents and 3,400 call-centre agents needed a single Arabic surface into a Customer 360 that spans a BigQuery lakehouse (billing, network) and an on-prem Postgres CRM. Bilytica shipped Sawal on WhatsApp and a web console, with a dialect-first NLU pack tuned to Gulf Arabic and MSA-Gulf code-switching, and a differential-privacy layer for aggregate queries hitting sensitive network KPIs.
Challenge
- •Store agents were switching between 5 systems per customer visit, averaging 6m 40s of handling time.
- •Existing text-to-SQL prototypes trained on English benchmarks scored 61% execution accuracy on Gulf-Arabic questions with heavy code-switching (وش، أبغى، كذا).
- •Network KPI queries risked re-identifying individual subscribers from aggregate ARPU/traffic slices.
- •PDPL and the operator's own data-classification policy required all inference and logs to remain inside Saudi Arabia.
R&D protocol
- •Assembled a 22,600-utterance Saudi telecom corpus (store dialogues, call-centre transcripts, WhatsApp logs), hand-labelled for dialect, intent, and gold entity spans.
- •Trained a dialect-aware intent classifier (MSA / Gulf / code-switched Arabizi) with contrastive loss; evaluated on a stratified held-out slice of 3,200 utterances the model never saw.
- •Extended schema-linking with a subscriber-lifecycle prior: 34 canonical entities (customer, MSISDN, contract, complaint, ticket, network cell) hard-coded as retrieval anchors rather than free-text columns.
- •Applied ε=1.4 differential-privacy noise on aggregate KPI queries with cohort size < 40, with sponsor-approved release-blocker if noise would flip a category ranking.
- •Built a 480-case red-team suite for re-identification (small-cohort filters, join-based inference, temporal narrowing), all findings triaged with the operator's DPO.
- •Deployed model, retrieval indexes, and logs on HUMAIN in-country; every prompt/response pair is hashed and pinned to a WORM store for 24-month PDPL retention.
Evaluation harness
Gold set: 3,200 Gulf-Arabic + 900 English utterances with gold SQL against a masked warehouse snapshot; every scenario dual-annotated by a linguist and a network analyst.
- →Execution accuracy: 93.4% Gulf, 95.7% MSA, 96.3% English
- →Dialect classification F1: 0.889 (baseline 0.71)
- →Re-identification attempts blocked: 100% on the 480-case suite
- →Toxic / unsafe output rate on 5,000-turn stress test: 0.02%
Red team: 480 adversarial cases across small-cohort re-identification, prompt injection through customer complaint text, and locale-spoofing (fake dialect switch to bypass filters). 0 successful re-identifications; 2 pre-release prompt-injection paths remediated.
Outcomes
- Average handling time in stores dropped from 6m 40s to 1m 55s (−71%).
- First-call resolution in the call centre rose 18 points on billing-dispute intents.
- The DPO signed off on the aggregate-KPI surface after the differential-privacy layer removed all findings from the 480-case suite.
- The dialect corpus and eval harness were transferred to the operator under full IP assignment.
Bilingual NL2SQL over classified program data, with SDAIA-aligned governance and data classification honored on every generated query
Saudi ministry program office (anonymized, 3 sector programs) · 16 weeks kickoff → production
The program office needed executives, sector leads and ministry analysts to query Vision 2030 KPIs by region and initiative in either language, while respecting a four-level data classification (Open, Internal, Confidential, Restricted). Bilytica delivered Sawal against a classified SQL Server warehouse with a policy engine that inspects every generated query at parse time and refuses execution, with an explanation, when the requesting principal is not cleared for the referenced columns.
Challenge
- •Program KPIs were spread across 8 sectoral marts with inconsistent naming (initiative_code vs init_id vs مبادرة_رمز).
- •Sector leads needed answers in MSA; ministry executives preferred English; both had to land on the same certified numbers.
- •SDAIA guidance and the ministry's own classification policy forbid any AI system from surfacing Restricted columns to a non-cleared principal, even indirectly through a join.
- •Executive briefings needed to cite the exact program document (KPI definition, target, baseline year) driving the number.
R&D protocol
- •Built a bilingual metric ontology: 186 program KPIs, each with an Arabic canonical name, English canonical name, unit, dimension set, owning entity, and classification level.
- •Trained a bilingual entity linker on 9,800 (question, KPI) pairs authored by ministry analysts; hard negatives sampled from within the same program to force fine-grained discrimination.
- •Implemented a query-time policy engine: every candidate SQL is parsed into an AST, each column node checked against the principal's clearance, and joins pruned if they would leak a Restricted column into a lower-tier result.
- •Held-out evaluation: 2,140 scenarios (1,190 AR / 950 EN), gold SQL by 4 sector analysts, reviewed by the ministry's data-governance office.
- •Red-team suite of 260 cases specifically targeting classification leakage via aggregation, join laundering, and Arabic/English switch attacks, all rejected by the policy engine.
- •Every answer carries a citation block linking to the KPI definition document and the target/baseline row that produced the number.
Evaluation harness
Gold set: 2,140 scenarios across 3 sector programs; every question paired with an Arabic canonical form, an English canonical form, expected classification level, and gold SQL.
- →Execution accuracy: 92.1% overall (AR 91.4%, EN 93.0%)
- →Semantic exact-match on program KPIs: 89.4%
- →Cross-language consistency (same answer, either language): 97.8%
- →Classification enforcement precision/recall: 1.000 / 1.000
Red team: 260 adversarial cases: 100% blocked. Every attempted leakage path was recorded, remediated in the policy grammar, and locked with a regression test.
Outcomes
- Executive briefings moved from a 4-day cycle to same-day, with the source KPI document cited on every line.
- The ministry adopted the bilingual metric ontology as a cross-program standard.
- Zero classification violations were recorded in the first six months of production use.
640 store managers asking inventory, staffing and promo questions on WhatsApp, with brand-level RLS honored end-to-end
Saudi retail group (12 brands, 640 stores across Saudi Arabia) · 12 weeks kickoff → production
Store managers had no real-time visibility into inventory, staffing gaps, or promo lift; they filed BI tickets that took 2–3 days. Bilytica shipped Sawal on WhatsApp against an Oracle ERP + Databricks lake with row-level security that scopes every generated query to the requesting manager's brand and store, enforced through a signed session claim rather than a prompt instruction.
Challenge
- •Oracle ERP and Databricks lake had overlapping but non-identical inventory tables, the same SKU could show different on-hand counts in each.
- •Managers wrote in casual Arabic (‘كم موجود من X في مستودعي؟’)، off-the-shelf models mapped questions to English column names and missed the branch scope.
- •Any query that returned another brand's data, even accidentally, would violate the group's franchise agreements.
- •WhatsApp latency budget was strict: managers abandoned answers slower than ~2 seconds.
R&D protocol
- •Reconciled Oracle and Databricks inventory into a single certified 'inventory_position' metric with a written source-of-truth policy per SKU class.
- •Fine-tuned a compact retrieval reranker on 7,200 hand-labelled (Arabic question, table set) pairs; ablation showed a +6.4-point execution-accuracy gain over BM25 alone.
- •Session-claim RLS: the WhatsApp identity is exchanged for a signed JWT carrying brand_id + store_id; every SQL is rewritten server-side to inject a mandatory WHERE clause; the LLM never sees or influences the scope filter.
- •Held-out eval: 1,560 scenarios in Gulf Arabic, gold SQL by 5 regional operations managers.
- •Latency co-design: retrieval index sharded per brand (avg 380 tables/shard), query plan cached on the semantic-layer key, warm connections pinned per region.
- •Red-team suite of 210 cases attempting to spoof brand_id via prompt injection, unicode look-alikes, and multi-turn manipulation, 100% blocked by the signed-claim design.
Evaluation harness
Gold set: 1,560 Gulf-Arabic scenarios covering inventory position, staffing gaps, promo lift, and shrinkage; gold SQL and expected brand_id/store_id scope on every case.
- →Execution accuracy: 94.2%
- →Correct brand/store scoping: 97.1%
- →Cross-brand data leakage attempts: 210 tested, 0 successful
- →Latency SLO (< 2 s on WhatsApp): met on 96.4% of production traffic
Red team: 210 spoofing / injection attempts, 100% blocked by the signed-claim scope design; findings documented in a threat model shared with the group's CISO.
Outcomes
- Stock-out incidents per 100 SKUs fell 63% in the first quarter after go-live.
- Promo sell-through rose 11 points in the 84 pilot stores that used Sawal for daily replenishment questions.
- The BI ticket backlog for store operations shrank to zero within 8 weeks of full rollout.
Bed-management, OR utilization and ED throughput questions answered in Arabic, with a de-identification layer engineered before the model sees a row
Saudi private hospital group (9 hospitals, 2,400 beds) · 15 weeks kickoff → production
Operations directors needed live answers on bed occupancy, OR turnover, ED length-of-stay and staffing ratios, cross-cutting an HIS on SQL Server, an OR management system, and an ED tracker. Bilytica shipped Sawal with a mandatory de-identification proxy that hashes MRNs and suppresses direct identifiers before rows reach the model, plus a clinical semantic layer authored jointly with the group's medical informatics team.
Challenge
- •Ops directors had no unified Arabic surface across three source systems with different codings of the same event (admission, transfer, discharge).
- •Any prototype that exposed raw MRNs or diagnosis codes to a general-purpose LLM was a non-starter with the group's ethics committee.
- •Clinical vocabulary is highly context-sensitive, ‘occupancy’ means different things in ICU, ward and day-surgery contexts.
- •The evaluation had to be signed off by clinicians, not just data engineers.
R&D protocol
- •Wrote a data-minimization spec: the model receives event types, timestamps, department codes, and hashed patient tokens, never MRN, name, phone, or free-text diagnosis.
- •Built a de-identification proxy in front of the warehouse; a formal test proved every generated query's projection list is a subset of the allow-listed columns, or the query is rejected.
- •Authored a clinical semantic layer of 96 operations metrics with a medical informatics lead; each metric has a written definition, exclusions, and a worked example.
- •Held-out evaluation of 1,320 scenarios; every scenario dual-reviewed by an operations analyst and a physician; disagreements adjudicated by the CMIO.
- •Bias & safety review: measured performance by department, shift, and hospital site, no group performed more than 3.1 points below the median.
- •Red-team suite of 190 cases attempting to reconstruct patient identity from small-cohort ED filters or timestamp joins; the de-identification proxy and cohort-size floor of 20 blocked all attempts.
Evaluation harness
Gold set: 1,320 scenarios across bed-management, OR utilization, ED throughput and staffing ratios; dual-reviewed by an operations analyst and a physician; CMIO-signed evaluation report.
- →Execution accuracy: 93.6% overall (ICU 92.2%, ED 94.1%, OR 94.3%)
- →Hallucination rate (unsupported clinical claim): 0.4%
- →Direct-identifier suppression at proxy: 100%
- →Sub-group parity: max delta from median = 3.1 pts
Red team: 190 re-identification attempts, all blocked. Every attempted vector was added as a permanent regression case in the CI pipeline.
Outcomes
- Emergency Department length-of-stay for admitted patients dropped 34% after Sawal was wired into the flow huddles.
- OR utilization rose 9 points on average across the 9 hospitals in the first six months.
- The de-identification proxy and evaluation report were adopted by the ethics committee as the group's template for future AI systems.
Three deployment tiers, scoped to every stage of maturity
Departmental
One warehouse · up to 25 seats
- 1 warehouse + 1 ERP or CRM connector
- Up to 25 named users
- Arabic + English channels: web, Teams
- Certified semantic layer (managed)
- PDPL-aligned logging
- Business-hours support
Enterprise
Unlimited seats · agentic actions
- Unlimited warehouses, ERPs, CRMs
- Unlimited seats + SSO/SCIM
- All channels + WhatsApp + voice
- Agentic actions with human approval
- Row-level security honored end-to-end
- Published NL2SQL harness on your data
- 24/7 support with 4h SLA
Sovereign / Air-gap
AI-in-a-Box or HUMAIN dedicated
- Deploy on AI-in-a-Box or HUMAIN dedicated
- Air-gapped operation supported
- Dedicated fine-tuning on your corpus
- Custom Arabic domain glossary
- Defense-grade audit trail
- White-glove onboarding
- Named engineering pod
Frequently asked questions
What is Sawal?+
Sawal (سؤال، Arabic for 'question') is Bilytica's Arabic-first conversational data platform. It lets business users ask questions in Arabic or English and returns governed, row-level-secure answers by writing SQL against your warehouse, ERP and CRM.
Why is Sawal built differently for Saudi enterprises?+
Sawal is engineered for Arabic dialects (not only MSA), enforces a governed semantic layer with certified metrics, honors row-level security end-to-end, and executes agentic write-back actions with human-in-the-loop approval, all hosted in-country on HUMAIN with a published Arabic NL2SQL evaluation harness so accuracy is auditable, not just claimed.
Is Sawal hosted in Saudi Arabia?+
Yes. Sawal runs on HUMAIN's Saudi-based sovereign infrastructure by default, with air-gapped on-prem and Bilytica GPU Cloud deployment options. All customer data, prompts and query logs stay inside Saudi Arabia, aligned to PDPL and NDMO.
Which systems does Sawal connect to?+
Warehouses (Snowflake, BigQuery, Databricks, Redshift, Postgres, Oracle, SQL Server), ERPs (SAP S/4HANA, Oracle Fusion, Microsoft Dynamics, Odoo), CRMs (Salesforce, HubSpot, Zoho), and Saudi-common systems (Ojoor, PayPeople, Qoyod, ZATCA e-invoicing feeds).
How is accuracy measured?+
Sawal ships with an Arabic NL2SQL evaluation harness covering 1,200+ Saudi-enterprise scenarios (finance close, supply chain, HR analytics, sector KPIs). Every release is scored on execution accuracy, semantic accuracy and hallucination rate, and results are shared with customers under NDA.
How is pricing structured?+
Sawal is offered in three deployment tiers, Departmental, Enterprise, and Sovereign / Air-gap, sized to warehouse count, seats, agentic scope and residency requirements. Pricing is scoped per engagement; contact the Bilytica team for a tailored proposal.
Ready to move your data into conversation?
Book an executive demo, we connect to your warehouse in 15 minutes and answer your real questions in Arabic.