Retrieval & data

AI Assistants in Saudi Arabia: From Basic Chatbots to Intelligent Business Agents

The word "chatbot" now covers three genuinely different kinds of software, and the gap between the cheapest and the most capable is far wider than the gap in what they cost. Knowing which one you are being sold is most of the buying decision.

8 min read Jul 14, 2026

A decade ago "chatbot" meant one thing: a decision tree with a text box in front of it. Today the same word is used for a scripted menu, for a language model answering from your documents, and for software that can complete a task on your behalf. They share an interface and almost nothing else. Procurement documents still treat them as one category, which is how organisations end up paying for the third and receiving the first.

This piece is about the actual differences — not the marketing ones. What changes technically at each step, what each step is genuinely good for, and where the honest ceiling of each sits.

Three kinds of software wearing the same interface

The clearest way to tell them apart is to ask what the system does when it meets a question nobody anticipated. That single question separates the categories more reliably than any feature list.

What each generation can and cannot do

	Rule-based bot	LLM assistant	Agentic assistant
Answers a scripted question	Yes	Yes	Yes
Handles unanticipated phrasing	No	Yes	Yes
Answers from your own documents	No	Yes	Yes
Remembers earlier turns	No	Yes	Yes
Reads a live system for an answer	No	No	Yes
Completes an action for the user	No	No	Yes
Behaves predictably by construction	Yes	No	No

The last row is the one buyers forget. Predictability is a real property and the rule-based column is the only one that gets it for free — everything to the right has to earn it back with grounding, scoping and review.

A capability comparison, not a quality one. A well-built rule-based bot beats a badly-built assistant on the narrow set of things it was designed for.

A rule-based bot is a menu. It is completely predictable, it costs very little, and it fails the moment a customer types something its author did not imagine — which is most of the time, because customers do not phrase questions the way process documentation does.

An LLM assistant understands the question rather than matching it. That is a genuine step change, and it introduces a genuine new risk: a language model asked something it does not know will often produce a confident, fluent, wrong answer. Everything serious in this field is about closing that gap.

An agentic assistant goes one step further and does something — checks an order status, books a slot, opens a ticket. This is where the value gets large and where the engineering discipline has to get much stricter, because a wrong answer is embarrassing while a wrong action has consequences.

3kinds of software sold under one word

1question that separates them: what happens when nobody scripted it

2languages that must each be measured on their own

0answers a grounded assistant should invent

What "context" and "memory" actually mean

Two words appear in every assistant datasheet and are almost never defined. They are worth pinning down, because they are the difference between a conversation and a sequence of unrelated questions.

Context is what the system can see while answering the current message: the conversation so far, the documents retrieved for this question, the identity of the person asking, and the instructions it operates under. It is finite. Every assistant has a limit, and what gets dropped when that limit is reached is a design decision with visible consequences.

Memory is what survives between conversations. It is genuinely useful and genuinely sensitive — a system that remembers a customer’s previous complaint is helpful, and a system that remembers their national ID number has become a data-protection question. Under the Personal Data Protection Law, whatever the assistant retains is processing like any other, with the same basis, retention limit and erasure duty.

Worth remembering

Ask any vendor two specific questions: what exactly is kept between sessions, and how long. "The assistant remembers your customers" is a feature claim; the answer to those two questions is a compliance position.

Where the reliability comes from

The reason a well-built assistant does not invent answers is not that its model is larger. It is that it is not being asked to remember anything. The question goes to a retrieval step first, the relevant passages come back, and the model is asked to answer using those passages and nothing else. Where nothing relevant comes back, the correct output is that it does not know.

What happens between the question and the answer

QuestionIn Arabic or English, as typed

RetrieveFind the passages that actually bear on it

GroundAnswer from those passages only

CheckRefuse rather than guess when nothing fits

AnswerWith the source it came from

Remove the retrieve step and you have a general chatbot with your logo on it.

The retrieval step is what makes the difference. A model answering from its own training alone is guessing about your business, however fluently.

This is also why the quality of an assistant tracks the quality of the material it was given far more closely than it tracks the model behind it. Two organisations deploying identical software get very different results, and the difference is almost always in whether their documentation contradicts itself.

The step most organisations skip

There is a stage between "we should look at this" and "it is live", and it is the stage that determines whether the project works. It is not a technical stage.

A realistic path from interest to production

WEEK 1Classify a month of real contactNot estimated from memory — actually categorised. This tells you what fraction is automatable and sets the target you will later be measured against.
WEEK 2Fix the source material firstEvery contradiction between two documents becomes an inconsistent assistant. This stage produces value even if the project stops here.
WEEKS 3–4Ground, then test in both languagesAnswers checked against the source, in Arabic and English separately. English passing tells you nothing about Arabic.
WEEK 5+Narrow live scope, then widenOne channel, one topic area, escalation on by default. Widen on evidence.

Nothing here requires the next stage to justify itself. Each is useful alone.

Weeks are indicative and vary with how much material exists and what state it is in. The middle two stages are where projects succeed or quietly fail.

Which one does a Saudi organisation actually need?

For a small, genuinely fixed set of questions with no personal data involved, a rule-based bot is not a bad answer. It is cheap, it is predictable, and it will not surprise you. The failure mode is that the fixed set is almost never as fixed as it looks in the requirements document.

For the far more common case — a large body of documentation, questions arriving in two languages and several registers, and a real cost to answering wrongly — a grounded assistant is the appropriate tool. Elbi sits here by design: bilingual, answering from the material you provide rather than from a general model’s recollection, and built to hand over to a person when the question needs one.

Agentic capability is worth adding after that, not before. It is the highest-value step and the one with the least tolerance for a shaky foundation. An assistant that cannot yet answer reliably should not be given the ability to act.

Common questions

What is the difference between a chatbot and an AI assistant?

A rule-based chatbot matches your message against a script and fails on anything its author did not anticipate. An AI assistant interprets the question, works from your own documents, and carries context across a conversation. The practical test is what happens when someone asks something unexpected — a chatbot offers a menu, an assistant answers or says it does not know.

Do AI assistants remember previous conversations?

They can, and whether they should is a decision rather than a default. Anything retained between sessions is personal-data processing under the PDPL, with a lawful basis, a retention limit and an erasure duty attached. Ask a vendor exactly what is kept and for how long before treating memory as a straightforward feature.

Are AI assistants more expensive than chatbots?

The licence usually is. The total often is not, because the cost of a rule-based bot sits in the ongoing work of maintaining its script and in the contact it fails to deflect. Compare the two on volume actually resolved rather than on list price.

What is agentic AI?

An assistant that completes a task rather than only answering — checking a live status, booking a slot, opening a ticket. It is the highest-value capability and the one that needs the strictest scoping, because actions have consequences that wrong answers do not.

Can one assistant handle both Arabic and English?

Yes, and it should be evaluated as two systems rather than one. Retrieval quality, phrasing coverage and tone all differ by language, and strong English results are not evidence of anything about Arabic. Test each separately on real customer phrasing.