Confidential — partner discussion

01 / 09

A sovereign AI service layer for Saudi digital government

Citizens describe what they are trying to accomplish. The system explains the requirements from verified government sources, names the agencies involved, and routes them into the right digital service.

GOV.SATawakkalnaNafathGovernment Service BusOpen the live prototype →

01 — Where the Kingdom already is

02 / 09

Saudi Arabia has already built the hard parts

This is not a market with missing digital government. It is a market with the ingredients in place and no shared intelligence layer across them.

1,100+

government services integrated into Tawakkalna

SDAIA reporting, July 2025 onward

817 → 550

government platforms after Whole of Government consolidation

DGA, H1 2025

100+

government entities involved in Tawakkalna integration

SDAIA, July 2025

200+

government representatives convened on agentic AI and MCP

SDAIA workshop, January 2026

GOV.SA national portalTawakkalna super-appNational Single Sign-On / NafathGovernment Service BusEmbedded ServicesNational Data Catalog

02 — The current state of AI

03 / 09

The intelligence exists. It is vertical.

Every assistant below is excellent inside one ministry, authority or destination. None of them can answer a question that crosses two.

NusukAI

Hajj & Umrah pilgrim guidance, AR/EN, text and voice

Balady Smart Assistant

Municipal services, regulations and requirements

Qiwa Smart Advisor

Labour platform guidance — 17% fewer human support requests

EtimadGPT

Government financial rules — RAG-like, semantic search

Nora — Visit Saudi

Tourism answers and personalised recommendations

Musaned chatbot

Domestic labour advice without human intervention

GOSI Brain

Sovereign KSA-hosted GenAI over proprietary knowledge

Seven assistants. Zero of them can answer a question that spans two of them.

03 — The gap

04 / 09

No layer does all four at once

As of August 2026, a search of public official material from GOV.SA, SDAIA, DGA, MCIT, SPA, NCA and Balady found no public evidence of a national cross-government assistant combining these four capabilities.

01

Open-ended questions

Any question, across the whole of government — not a menu of intents.

02

Grounded answers

Retrieval restricted to authoritative government sources only.

03

Explicit citation

Every factual claim attributable to an approved page.

04

Cross-agency routing

Straight into the correct digital service or transaction.

04 — What we build

05 / 09

Not another portal. An embeddable layer.

DGA reduced government platforms from 817 to 550 by H1 2025. Adding a 551st would be strategically wrong. The same engine should appear inside the surfaces citizens already use.

GOV.SA

National portal — discovery and information

Tawakkalna

Transaction surface via Embedded Services

Agency sites

Balady, Qiwa, Absher and others

Contact centres

Agent-assist over the same verified corpus

One retrieval, verification and orchestration engine. Many surfaces. We do not own the government service — we own the layer that makes government services conversational.

See it working →

05 — Why we start ahead

06 / 09

We do not spend year one rebuilding RAG plumbing

Canada open-sourced a production-oriented AI Answers architecture under MIT: government-only retrieval, citations, privacy protection, human evaluation and model independence. We use it as a risk reducer, not as the finished architecture.

Inherited — Canada AI Answers, MIT

  • —Government-only retrieval design
  • —Citation and attribution model
  • —Privacy protection pipeline
  • —Human evaluation methodology
  • —Model independence / gateway

Built here — the Saudi IP

  • +Sovereign Saudi source registry and approval workflow
  • +Arabic-first retrieval and bilingual evaluation set
  • +Cross-agency service registry and routing graph
  • +Two privacy modes for anonymous and authenticated use
  • +Agency tenant controls and content-owner tooling

06 — Evidence

07 / 09

The mechanism is already proven — including in the Kingdom

Every figure below is from published official or system-card reporting, not from our modelling.

96%

expert-rated factual accuracy

Canada AI Answers system card, 2025 trials

116,000+

inquiries handled by EtimadGPT

Saudi national portal reporting

2,256+

staff hours saved by EtimadGPT

Saudi national portal reporting

70%+

reduction in response time

EtimadGPT, national portal reporting

17%

fewer human support requests

Qiwa Smart Advisor, national portal reporting

07 — How we prove it here

08 / 09

A six-month pilot with contractual success criteria

Roughly 8-10 core FTE, plus one content owner and one subject-matter expert part-time from each pilot agency. That agency commitment matters more than another ML engineer.

Weeks 0-4

Sponsor alignment, agency selection, architecture, PDPL and security review, source inventory

Pilot charter, governance model, baseline evaluation set

Months 1-3

Adapt AI Answers, Arabic/English retrieval, Saudi source registry, model gateway, in-Kingdom hosting, service registry

Internal alpha, then SME-accessible alpha

Months 3-5

500-1,000 curated bilingual test questions, security testing, red team, limited real-user pilot

Closed government beta, pilot evidence package

Month 6

Evaluation, business case, accreditation planning, expansion design

Go / no-go for national integration

Success thresholds

Overall expert-rated accuracy
≥ 92%
Fees, deadlines, required documents
≥ 98% or abstain
Citation validity
≥ 99.5%
Arabic vs English accuracy gap
≤ 3 pts
Appropriate abstention
≥ 95%
Direct-service link correctness
≥ 99%
PII leakage into anonymous pipeline
0 incidents
P95 answer latency
< 12s

Who we pilot with

  • Strategic sponsor: DGA / GOV.SA

    The product is a Whole of Government experience layer; DGA runs a Proof-of-Concept Design process

  • Platform partner: SDAIA

    Tawakkalna, Embedded Services, data and AI governance, national standards

  • First content domain: Ministry of Commerce

    Business setup crosses agencies without entering sensitive contexts

  • Comparative pilot: Balady

    Already has a siloed assistant — the honest test of cross-agency value

08 — The ask

09 / 09

We are asking for a decision to explore this together

Not capital. Three things.

01

Confirm the thesis

Agreement that a sovereign cross-government AI service layer is worth pursuing together — as infrastructure, not as another portal.

02

Open one door

An introduction to a named sponsor at DGA or SDAIA who can host a proof-of-concept conversation.

03

Scope the pilot with us

A working session to turn this into a pilot charter: agency, sources, success thresholds, governance.