The buried document
It exists, in a shared drive, an inbox or a DMS folder nobody remembers.
«الملف موجود… لكن أين؟»
AI portal for government and enterprise · Built for the Gulf
معرفة مؤسستك… على بُعد سؤالٍ واحد
Tawi connects to every document store, database and legacy system you already run, speaks Arabic natively, respects who may see what, and acts through agents your own team configures without code. On your servers.
تحتاج إلى السجل التجاري، وبطاقة الهوية، وشهادة عدم ممانعة من البلدية. تستغرق المعاملة يومَي عمل عند تقديمها إلكترونياً.
المصدر: دليل الإجراءات، الفصل الثالث، صفحة 14A real question, a cited answer, and permissions applied before retrieval.
The problem
Five everyday tasks that quietly consume the working week.
It exists, in a shared drive, an inbox or a DMS folder nobody remembers.
«الملف موجود… لكن أين؟»
The answer is spread over forty PDFs. Keyword search returns four hundred hits and no answer.
«نبحث عن الكلمة… لا عن المعنى»
A simple figure needs a SQL query. A ticket, a queue, and days of waiting.
«اطلبها من قسم تقنية المعلومات»
You remember what the photo shows, never its file name or its folder.
«أذكر الصورة… لا اسم الملف»
Only one specialist can operate the legacy system. Hours or days until they are free.
«انتظر حتى يفرغ الموظف المختص»
of the working week goes to searching for information
Atlassian, State of Teams 2025
a day looking for information in large organisations
Coveo, EX Relevance Report 2025
of digital workers can’t find the information they need
Gartner, 2023 (4,861 workers)
a week waiting on colleagues or redoing existing work
Panopto / YouGov
Roughly one working day a week, for every employee. In a ministry of 5,000 people, that is the time of about 1,250 full-time staff spent looking for information instead of serving the public.
Illustrative: a global survey figure (Atlassian, 2025) applied to a hypothetical headcount. The pilot replaces it with your own numbers.
Why now
Every Gulf state now has a national AI strategy, sovereign Arabic models are being built in Riyadh, Abu Dhabi, Doha and Muscat, and data-protection laws require governed, in-country data. Governments and enterprises across the region are moving from pilots to platforms.
GCC states with a national AI strategy or roadmap in place
sovereign Arabic language models built in the Gulf: ALLaM, Jais, Fanar and Mu’een
AI spending across the Middle East, Türkiye and Africa by 2028, up from $4.5B in 2024
UAE, Saudi Arabia and Qatar lead the Arab world in the 2024 Government AI Readiness Index
data-protection laws across the Gulf: AI must run on governed, in-country data
Sources: BCG, The GCC AI Pulse (2025) · IDC, AI spending in the Middle East, Türkiye and Africa (Jan 2025) · Oxford Insights, Government AI Readiness Index 2024 · national data-protection laws of Saudi Arabia, the UAE, Qatar, Bahrain, Kuwait and Oman
The solution
An AI portal that connects to any knowledge base, speaks Arabic natively, runs on your servers, and acts through agents your own team configures, without code.
SQL, NoSQL, file systems, DMS, email, SaaS
Text, tables, images, graphs, scanned Arabic
MCP builder for apps and legacy systems
On-prem or via API, chosen per task
Portal, WhatsApp, Teams, Telegram, Meta, email
Architecture
Everything in the diagram is configured by your own team from Tawi Studio, without code.
Web portal · WhatsApp · Teams · Telegram · Meta · Email · Voice · API / A2A
Router · planner / executor · validator · memory · guardrails · full trace
On-prem or API, routed per task
Vector + keyword + graph · ACL-filtered · cited
Schema discovery · glossary · governed SQL
MCP · OpenAPI · A2A · auth · audit
DMS, SharePoint, Drive, file shares, email, CMS
Oracle, SQL Server, PostgreSQL, data lakes
ERP, HR, GIS, ticketing
Capabilities
Knowledge
Connect every repository
«ابحث عن صورة الوزير مع الوفد في الميناء — تلك التي تظهر فيها الرافعات الزرقاء»
Tawi finds the photo by what it shows, not by its file name or folder.
Structured data
How it stays governed
Works with PostgreSQL, SQL Server, Oracle, MySQL, Snowflake, BigQuery, Databricks and Fabric.
كم معاملة أُنجزت هذا الشهر، موزّعة حسب الفرع؟
SELECT branch, COUNT(*) AS completed
FROM transactions
WHERE status = 'completed'
AND month = '2026-10'
GROUP BY branch
ORDER BY completed DESC;
Results by branch, with the SQL and the source tables shown, ready to export or chart.
Tools and legacy systems
At an API specification, a database’s stored procedures, or a legacy system’s screens and exports.
An MCP server: tool schemas, authentication, confirmation rules and Arabic descriptions the agent understands.
One gateway: least privilege, rate limits, an audit record of every call, and a confirmation before any write.
Examples
The specialist’s know-how becomes a procedure any employee can run, in Arabic, from WhatsApp or Teams.
Agent harness · Tawi Studio
Name, persona, tone, Arabic and English, employee or citizen audience
Step-by-step flows for key requests, written in plain language
Primary and fallback model per step: router, answerer, SQL writer
Which collections and data models the agent may use, and how it retrieves
Search, SQL, MCP servers, OpenAPI actions, other agents, with confirmation rules
Session and long-term memory per person, with a retention period
Allowed topics, PII redaction, injection filter, groundedness check, spend limits
When and where to escalate to a person, with the full conversation
Where the agent is published and how answers are formatted per channel
Every published version is tested against an evaluation set, can be A/B-tested against the previous one, and rolled back in one click.
Arabic
Omnichannel
A WhatsApp number, a Teams login and an email resolve to the same person, so context follows them. Handoff to a human carries the full conversation.
Hosting and models
Inside the ministry’s own data centre. No data leaves the network.
Hosted by a national provider in Saudi Arabia, the UAE, Qatar, Oman or elsewhere in the Gulf. Data residency by default, aligned with your national data-protection law.
Control plane hosted; connectors, index and database access stay inside your network.
One model gateway, routed per task
| Task | On-prem, sovereign | Via API |
|---|---|---|
| Intent routing | Qwen, Llama (small models) | — |
| Arabic answers | Mu’een*, Jais, ALLaM, Fanar | Claude, GPT, Gemini |
| SQL writing | Qwen-Coder, Llama | Claude |
| Embeddings | bge-m3 (multilingual) | — |
*Where made available by the national provider. Bring your own keys; swap a model without rebuilding an agent.
Use cases
Leave, allowances, circulars and procedures answered from HR policies and the HR system.
Channel · Teams
Decrees, case files and scanned Arabic archives searched by meaning, with citations to the page.
Channel · Portal
Permit requirements, application status and appointments, answered from the service catalogue and case system.
Channel · WhatsApp
KPIs, budget execution and transaction volumes asked in plain Arabic over the data warehouse.
Channel · Portal and Teams
Tender status and supplier documents pulled from a legacy procurement system through a generated MCP tool.
Channel · Portal
One question across ministries’ documents and databases, with a cited summary for the morning meeting.
Channel · Email and portal
Each starts with one agent, one set of sources and one channel, then grows from there.
Next steps
Weeks 1–2
Three document repositories and one database, installed on your servers.
Weeks 3–5
Two agents, their procedures, an Arabic evaluation set and guardrails.
Weeks 6–8
Portal plus WhatsApp or Teams for about 200 staff in one ministry.
Weeks 9–10
Review the numbers together and decide on the scale-out.
What we measure
One ministry, real documents, real users, your servers, and a decision based on your own numbers rather than a demo.
لنجعل معرفة مؤسستكم على بُعد سؤالٍ واحد
Book a 30-minute introduction, or go straight to the ten-week pilot in one ministry.