• Representative example
  • Case B-02, Saudi Arabia

A cited Arabic GenAI assistant for a Saudi insurer

In this representative example, 6 engineers would build a cited Arabic and English assistant for a Saudi insurer’s policy servicing, live in 14 weeks. Average handling time would fall 52%, with 100% of outputs logged.

Build my squad

No card needed · 72 h free hold

–52%

average handling time

In short

Case B-02 is a representative example the bilingual retrieval-augmented assistant a Remote Pro Ai squad would build for a Saudi insurer. 6 engineers would deliver policy-servicing and internal-knowledge answers in Arabic and English, with an evaluation harness, human review, monitoring and one-click rollback. It would go live in 14 weeks, would log 100% of outputs and cut average handling time by 52%.

In this representative example, a Saudi insurer’s agents search dozens of documents per call. A 6-engineer squad would put a cited assistant live in 14 weeks.

Answers with sources, in Arabic and English

  • Client (representative)

    A Saudi insurer

  • Team

    6 engineers · 7 months

  • Model

    Managed Delivery

  • Data pattern

    Pattern B, the insurer’s in-Kingdom cloud region

  • Stack

    Python · LlamaIndex · Arabic and multilingual embedding models · pgvector · LLM gateway with policy filters · MLflow · evaluation harness · OpenTelemetry · React agent console

  • Assembled in

    4 working days

The right answer exists, buried in bilingual paperwork

In this example, service agents answer policy questions by searching product guides, circulars and internal procedures, some in Arabic, some in English, many in both with slightly different wording. Calls run long, answers vary by agent, and supervisors spend hours correcting them.

The insurer wants generative AI, but not a black box. Every answer has to cite its source, personal data has to stay masked in logs, and the business needs to roll back any model change in one step.

  • Arabic and English questions, often mixed in one sentence
  • No answer without a cited source document
  • PDPL-compliant logging with personal data masked
  • One-click rollback for any model or prompt change

6 AI engineers, assembled in 4 working days

RoleLevelCountLocation
AI solutions architect and Squad LeadLead1Riyadh
NLP engineer, ArabicSenior1Cairo
AI engineer, retrieval and orchestrationSenior1Cairo
Data engineerSenior1Cairo
MLOps engineerSenior1Cairo
Full-stack engineer, agent consoleMid1Cairo

The Squad Lead would sit with the insurer’s operations team in Riyadh to run the golden-question workshops with supervisors.

Evaluation set first, model second

  1. Weeks 1–⁠2

    Top call intents mapped, documents inventoried, and a bilingual golden question set written with supervisors.

  2. Weeks 3–⁠8

    Retrieval pipeline, LLM gateway, citation-first answers and an evaluation harness scoring every build.

  3. Weeks 9–⁠12

    Pilot with one agent group; low-confidence answers routed to human review; adversarial prompt testing.

  4. Weeks 13–⁠14

    Production go-live with monitoring dashboards and rollback rehearsed.

  5. Months 4–⁠7

    Internal-knowledge use cases added, drift monitoring, rollback drills and hand-over.

Operating rhythm
2-week sprints · weekly executive report to the Head of Customer Operations every Friday · monthly steering · continuity target: zero unplanned turnover across 7 months.

Results

  • –52%

    average handling time

  • 100%

    of AI outputs logged

  • 14 weeks

    to production

Check it Every figure on this page has a definition, a source, a period and a verification date in the metrics book. Open the metrics book

These figures are representative
They show the kind of result this squad model is built to deliver, not a verifiable named-client record. Handling time compares the pilot agent group’s average call duration for covered intents with the 3 months before launch. Logging coverage is checked daily by reconciling gateway requests against stored records. Time to production runs from squad start to the first production answer.

About this case

Representative engagement (anonymised)

No client quote: this is a representative example

This is a representative example. No client is named, and the figures show the kind of result this squad model delivers, not a verifiable named-client record. On a scoping call you can ask for a reference under NDA from a comparable engagement, subject to that client’s consent.

Book a scoping call

You receive the reference's name and the call date within 3 working days, or a written reason if a call is not possible.

Request a reference call

Fields marked * are required.

References speak under NDA, so the name is shared with you by email, not published on the site.

Compliance this squad would handle

  • PDPL-compliant logging, with personal data masked before storage
  • Human-in-the-loop review for every low-confidence answer
  • Source citation required on every answer; uncited answers blocked
  • One-click rollback for models, prompts and retrieval settings
  • Hosting in the insurer’s in-Kingdom cloud region (Pattern B)

Related roles & practices

Practices and industries
AI & Data practice · banking and fintech · Saudi Arabia

Frequently asked questions

Book a scoping call

20 minutes with an Account Director. No sales script.

Answers are generated only from retrieved documents, and each answer must cite the passage it used. If retrieval confidence is low or no source supports an answer, the assistant routes the question to a human reviewer instead of answering. The evaluation harness checks these rules on every build before release.

The squad would test multilingual embedding models against a golden question set written by supervisors, including questions that mix Arabic and English. Documents are indexed in both languages, and answers follow the language of the question. Arabic quality is scored separately so it cannot hide behind strong English results.

Every model, prompt and retrieval configuration is versioned in MLflow. If monitoring shows a drop in evaluation scores or a rise in human escalations, the operations team rolls back to the previous version in one step. The squad would rehearse rollback before go-live and run drills each quarter.

Personal data is masked before any prompt, answer or retrieval record is stored, so logs keep full traceability without exposing customer details. Logs, documents and the vector store stay in the insurer’s in-Kingdom cloud region under Pattern B. Logging coverage is reconciled daily against gateway requests, which is how the squad would confirm that 100% of AI outputs are recorded.

The figure compares the pilot agent group’s average call duration for intents the assistant covers with the same group’s average in the 3 months before launch. Only covered intents are included, so unrelated call types do not distort the result. The insurer’s operations team would supply the call data, and the squad would verify the calculation before any publication.

All questions, by topic

Need an Arabic AI assistant that runs in the Kingdom? Bring us your use case

Arabic-capable AI engineers named within 72 hours, working in your cloud within 4 working days.