H&H Technology Insights | Enterprise AI and CX in Saudi Arabia

Custom AI Engines vs Off-the-Shelf LLMs: What Saudi Enterprises Need to Know

Eajaz Ali

August 1, 2026

A custom AI engine is a language model system trained or tuned on your organisation's own data and deployed inside your environment, rather than a general-purpose chatbot accessed over a public API. Enterprises choose custom engines when accuracy on internal knowledge, Arabic language performance, data residency or auditability matter more than getting started quickly.

The gap between a demo and a deployment

Most enterprises in Saudi Arabia have now run the same experiment. Someone connects a general-purpose model to a business question, the answer looks impressive, and a pilot gets approved. Then the model is asked something that depends on internal reality — a pricing exception, a contract clause, a regulatory position, a customer's service history — and it produces an answer that is fluent, confident and wrong.

This is not a defect. A general model was trained on public text. It has no access to your CRM, your policy documents or the last eight years of your operational decisions. Asking it to reason about your business is asking it to guess.

Saudi Arabia's AI market was estimated at around USD 9.3 billion in 2025 and is projected to keep compounding through the decade, supported by a sovereign AI fund and SDAIA's national strategy. A significant share of that spend will be wasted on pilots that never cross this gap.

What a custom AI engine actually is

A custom engine is not a model trained from scratch — almost no enterprise should do that, and the cost rarely justifies it. In practice it is a layered system:

A capable foundation model as the reasoning base
A retrieval layer (RAG) that grounds every answer in your documents, records and databases
Domain tuning so the model uses your terminology, formats and business rules
Guardrails that define what the system must refuse, escalate or flag
A deployment boundary that keeps sensitive data inside your infrastructure

The retrieval layer is where most of the accuracy comes from, and it is the part most often skipped.

Arabic is not a translation problem

For organisations serving Saudi and Gulf customers, Arabic performance is usually the point at which the off-the-shelf option fails.

General models handle Modern Standard Arabic reasonably and Gulf dialect poorly. They mishandle code-switching between Arabic and English in the same sentence, which is how a great many customers actually write. They lose accuracy on right-to-left formatting in generated documents, and they have limited exposure to Saudi regulatory and commercial vocabulary.

A custom engine addresses this at the data layer rather than by bolting translation onto the front. Your own Arabic correspondence, tickets and documents become the grounding material, so the system learns how your customers write and how your organisation responds — not how Arabic appears in a general web corpus.

When you should not build one

An honest assessment matters more here than a sales pitch. A custom engine is the wrong choice when your use case is genuinely generic — summarising public documents, drafting routine copy, general research. An off-the-shelf assistant does that well and costs a fraction as much.

It is also the wrong choice when your underlying data is a mess. If your knowledge lives in undocumented spreadsheets, contradictory policy PDFs and individual inboxes, a model built on it will faithfully reproduce the contradictions. Data remediation has to come first, and any vendor who tells you otherwise is selling you an expensive disappointment.

Frequently asked questions

How long does a custom AI engine take to build?
A grounded, production-ready first version typically takes eight to twelve weeks, depending on the state of the source data. Data preparation is usually the longest phase, not model work.

Do we need to send our data to an external provider?
No. Where data residency or regulatory constraints apply, the engine can be deployed inside your own cloud tenancy or on-premise, so proprietary data never leaves your control.

Will it hallucinate?
Retrieval grounding substantially reduces fabrication because answers are tied to retrieved source passages, and the system can be configured to cite those sources and to refuse when it has no basis to answer. No system is perfectly reliable, which is why high-stakes workflows keep a human approval step.

How is a custom engine different from fine-tuning?
Fine-tuning adjusts a model's style and behaviour. Retrieval gives it access to facts. Most enterprise problems are factual, which is why retrieval usually delivers more value per riyal than tuning alone.

What does it cost to run?
Ongoing cost is driven by inference volume and hosting, not licence seats. Most clients find it predictable once usage patterns settle after the first quarter.

A sensible sequence

Start with one workflow where wrong answers have a measurable cost — contract review, technical support triage, tender response, internal policy lookup. Ground it properly, measure accuracy against a human baseline, then expand. Organisations that begin with a narrow, well-instrumented use case tend to reach production; those that begin with "an AI assistant for the whole company" tend not to.

You can read more about how we build custom AI engines, how they fit into your existing ERP and CRM stack, or start with a consulting engagement to work out whether you need one at all.

Success is the result of perfection, hard work, learning from failure, loyalty and persistence”

Eajaz Ali

The question is never whether the model is clever enough. It is whether it can see the information it needs to answer correctly — and most of the time, it cannot.