Skip to content

Illustrative: a government entity

A cited Arabic assistant for circulars, policies and procedures

An illustrative scenario: staff at a government entity spend hours searching thousands of Arabic circulars, policies and procedures, many of them scanned, to answer routine questions. This is how we would deliver an on-premises assistant that answers in Arabic, cites the exact document and clause, and shows each person only what they are already allowed to see.

Illustrative scenario

Illustrative scenario: a composite of common requirements, not a description of a specific client project. Figures are targets we would agree and measure against — not results achieved.

The challenge

In this scenario, the entity has issued circulars and policies over many years. They sit in a document management system, shared drives and e-mail archives; some are scanned, some supersede others, and few people know which version applies. Staff ask colleagues or the legal team, and the same questions are answered again and again.

A public AI service is not an option. The documents are internal, some are restricted, and the entity’s security requirements — including the NCA Essential Cybersecurity Controls — shape where data may be processed and how access is controlled.

Our approach

We would run the work in three stages: a discovery to agree the document scope, the questions that matter and the evaluation set; a build inside the entity’s environment; and a phased rollout, department by department, with training for users and for the team that will run it.

An answer with its source, or no answer

Every answer cites the circular or policy, its number and date, and the passage it relies on. When the documents do not support an answer, the assistant says so and suggests who to ask. Superseded documents are marked, so an answer never rests on a withdrawn circular without saying so.

Permissions carried through

The assistant searches as the user. Access controls from the document management system are carried into the index, so a person only receives answers drawn from documents they could already open.

What we would build

What we would build

  • 01

    Document ingestion

    Connectors to the document management system and shared drives, Arabic OCR for scans, and metadata for number, date, issuing department and status — in force or superseded.

  • 02

    Arabic retrieval

    Search that combines keywords and meaning, tuned for Arabic word forms and the entity’s own terms, returning passages rather than whole documents.

  • 03

    Language model, on-premises

    An Arabic-capable model — such as ALLaM, Jais or Qwen — hosted on the entity’s own servers. No question, document or answer leaves the network.

  • 04

    Permission-aware access

    Single sign-on with the entity’s directory, document permissions checked on every query, and every question and answer logged for audit.

  • 05

    Evaluation and feedback

    A test set of real questions written with policy owners and run before each release. Users can flag wrong or incomplete answers for review.

  • 06

    Where staff already work

    A web interface, with an optional entry point in Microsoft Teams or the intranet, so nobody has to change tools to use it.

Target measures

Answers that cite their source
100%
Target agreed at kick-off; unsupported questions get “not found”, not a guess
Correct answers on the evaluation set
90% or more
Target, measured on questions written with policy owners
Time to a cited answer
Under 2 minutes
Target, compared with a timed baseline
Answers from documents the user cannot open
Zero
Target for acceptance testing, with test accounts for each role

Expected outcome

The expected outcome is that staff get a cited answer to a routine policy question in minutes rather than waiting for a colleague, and policy owners spend less time answering the same questions. The legal and policy teams remain the authority; the assistant points to what they have issued.

These are targets, not results. They are agreed with the business owner at kick-off, checked against a timed baseline and an evaluation set, and reported against after each rollout stage.

Have a process that should run itself?

Tell us about it. We will help you work out whether AI is the right answer, what it would take, and what it should return.

Newsletter

Practical AI, in your inbox.

Real use cases, Saudi regulations and what works in Arabic — once or twice a month. No spam; unsubscribe any time.

We’ll e-mail you to confirm. Unsubscribe any time — see our privacy policy.