Everything below is written for decision-makers who need open source vs closed llm to produce commercial results, not for people collecting best practices.
The reliable returns from AI right now are unglamorous — support deflection, document search, drafting, extraction. The ambitious autonomous systems work best once those foundations exist and the data underneath them is clean.
Framing the problem properly
Strip away the jargon and open source vs closed llm is a resource-allocation decision: where to put attention, budget and technical effort so that the return is visible within a defined period. In Saudi Arabia that decision is shaped by three constraints — a bilingual audience, a mobile-first population, and a regulatory floor that has risen sharply since 2024. Any recommendation that ignores those three is imported advice, and imported advice under-performs here consistently.
Data sovereignty and where the model runs
For regulated Saudi sectors, in-Kingdom processing is increasingly expected and sometimes required. Options range from local hyperscaler regions with contractual guarantees, through sovereign cloud arrangements, to self-hosted open-weight models on your own infrastructure. Each trades capability against control and cost. Decide based on data classification, not on general anxiety.
Governance, in one page
Which tools are approved. What data may never be pasted into an external model. When AI assistance must be disclosed. Who reviews AI output before it reaches a customer. How incidents are reported. One page people actually read beats a policy document that lives unopened on the intranet — and given SDAIA's active role in AI and data governance, having something written is now table stakes.
The Saudi market rewards specificity — local prices, local proof, local language — and punishes generic content faster than most.
Choose the cheapest architecture that solves the problem
Prompt engineering with a capable general model handles more than most teams expect. Retrieval-augmented generation adds your own documents and is the right answer for the majority of business use cases. Fine-tuning is for consistent format, tone or a narrow specialised task — rarely for adding knowledge. Work upward through that ladder and stop at the first rung that meets the requirement; each step up multiplies cost and maintenance.
Evaluation before deployment
Build a test set of a hundred real questions with known good answers before launch. Score accuracy, refusal behaviour on out-of-scope questions, and tone. Re-run it whenever you change the prompt, the model or the corpus. Without this you are shipping on anecdote, and quality regressions arrive silently after routine changes.
Where the return actually shows up
The reliable wins are unglamorous: first-line support deflection, document search across years of accumulated files, drafting and summarising routine correspondence, extracting structured data from invoices and forms, and translation quality assurance. Each is measurable, contained and pays back inside a year. The ambitious autonomous agent projects usually work best after these foundations exist.
Chunking: write in liftable units
Retrieval systems break pages into passages. A paragraph that depends on the three before it to make sense will be discarded or, worse, quoted misleadingly. Write self-contained units: each section names its subject explicitly, avoids unresolved pronouns, and includes enough context to stand alone. This single habit does more for AI visibility than any technical file you can add to your root directory.
Where agentic systems beat fixed rules
Rule-based automation excels at deterministic, stable processes. Agentic approaches earn their keep where inputs vary — unstructured documents, free-text enquiries in mixed Arabic and English, exception handling that previously required judgement. The practical pattern is a hybrid: rules for the deterministic path, an agent for the exceptions, and a human reviewing anything above a defined risk threshold.
Typical pilot shape
| Stage | Typical window | What you should see |
|---|---|---|
| Use case selection and baseline | 1–2 weeks | Must be measurable or the pilot cannot be judged |
| Data preparation and retrieval build | 2–4 weeks | Usually the largest share of effort |
| Evaluation and tuning | 2–3 weeks | Against a hundred-question test set |
| Controlled production rollout | 4–8 weeks | With human review on defined risk thresholds |
Windows assume consistent execution and a market of ordinary competitiveness. Treat them as planning ranges, not commitments.
Human in the loop, positioned deliberately
Decide in advance which decisions the system may take alone, which need approval, and which it must never take. Log every action for audit. Set confidence thresholds that escalate rather than guess. This is what makes automation defensible to auditors, regulators and the team whose work it touches — and it is what keeps a small error from becoming a systemic one.
The short audit
- Plan the role change for affected staff explicitly rather than leaving it to rumour
- Cache repeated queries and route simple requests to smaller models
- Log every interaction for audit and quality review
- Test candidate models on your own Arabic content, not published English benchmarks
- Measure retrieval quality separately from generation quality
- Choose one contained use case with an existing measurable baseline
- Agree what data may never be pasted into an external model
- Re-run the evaluation set after every prompt, model or corpus change
Arabic-language visibility is a separate project
Assistants answering in Arabic draw on a thinner corpus than they do in English, which means less competition and a genuine first-mover advantage. Publishing authoritative Arabic content — properly written, structurally clean, factually consistent — is currently one of the highest-leverage moves available to a Saudi business, and it will not stay uncontested for long.
The national context, briefly
Saudi Arabia designated 2026 its Year of Artificial Intelligence, with substantial state-backed investment channelled through SDAIA, sovereign AI vehicles including HUMAIN, and Arabic-language model development such as ALLaM. For an ordinary business the significance is less about the headline figures and more about what they produce downstream: local compute capacity, in-Kingdom cloud regions, Arabic models that work properly, a talent pipeline, and procurement expectations that increasingly assume digital maturity.
Where to start this week
Choose one contained use case with a measurable baseline — support deflection, document search, invoice extraction. Build a hundred-question evaluation set from real examples before you build anything else. Test your shortlisted models on your own Arabic content rather than published benchmarks. Write the one-page usage policy while the pilot runs.
The competitive advantage in this market is still consistency. Most competitors will read something like this, agree with it, and change nothing. The gap that creates is the opportunity.
