AI chatbot for your website, in Arabic and English
Written conversation is still where most customer contact begins. It is the cheapest channel to run, the easiest to audit, and the hardest to get right in two languages at once.
Written conversation is still where most customer contact begins. It is the cheapest channel to run, the easiest to audit, and the hardest to get right in two languages at once.
What this actually means
A chat assistant is judged on four things, and only four: whether it understood the question, whether the answer was correct, whether it arrived fast enough that the person did not give up, and whether it knew its own limits. Everything else — the widget design, the personality, the emoji — is decoration on top of those four.
Elbi is built around the assumption that the fourth is the most important. An assistant that answers ninety questions well and confidently invents an answer to the tenth is worse than one that answers eighty and escalates twenty, because the invented answer is the one that ends up screenshotted, shared, and quoted back to you by a regulator or a journalist.
What happens, step by step
- The language is settled firstScript is decisive — Arabic characters mean an Arabic conversation. Short replies like "ok" or "yes" are deliberately not treated as evidence of a language change, because one word is not enough to switch on and a mis-switch mid-conversation is jarring.
- The question is matched against your materialCompared against your own approved content — policies, product pages, price lists, procedures. The difference between a useful assistant and a liability is almost entirely whether answers come from a source you control.
- An answer is drafted from what was foundComposed from the retrieved material, in the language of the question, in the tone you configured. Where your documents are silent, the reply says so.
- The reply is checked before it is sentAnswers pass a check for claims unsupported by the retrieved material, and for personal data that should not be repeated back.
- It streams, so the wait feels shorterText appears as it is produced rather than landing in one block. Total time is similar; perceived wait is much shorter, and abandonment falls accordingly.
- Escalation is one sentence away"I want to speak to someone" is understood as an instruction, not a query. The conversation moves to a live agent with its history intact.
Where it earns its place
Before the sale
Availability, specification, price, delivery window. Questions people will not phone about but will abandon a purchase over.
After the sale
Order status, returns, warranty, branch details. High volume, low complexity, and the largest category of avoidable contact.
Eligibility questions
Whether someone qualifies, what documents they need, what the process looks like. Long answers people would rather read than be told.
Out of hours
The majority of chat traffic for many organisations arrives when the contact centre is closed. This is not replacing a person; it is covering hours nobody was covering.
Internally
Staff asking about policy, leave, expenses and procedure — the same questions, asked repeatedly, that quietly consume HR and IT time.
As a filter
Even where a person will handle the case, chat can gather details first so the agent starts with the account number and the problem rather than "how can I help".
Why two languages makes this harder
Most assistants sold into this region are English systems with Arabic bolted on, and you can usually tell within three messages. The giveaways are consistent: the reply is grammatically correct but reads as though it was written in English first; numbers and dates flip to the wrong side of the sentence; a mixed Arabic-English question gets answered in whichever language dominated the training data rather than the one the customer used.
Arabic also carries a problem English does not. It is diglossic — the formal written language and the language people actually type are meaningfully different, and Gulf, Levantine and Egyptian usage diverge further. A customer typing in Saudi dialect is not writing Modern Standard Arabic, and a system tuned only to formal text will miss them. That is a retrieval and matching problem, and it has to be solved on the Arabic side directly rather than by translating into English and back.
The interface has to follow. Arabic is written right to left, which changes far more than alignment: the reading order of the layout inverts, the panel sits on the opposite side, arrows point the other way, and a question mark belongs at the left end of the sentence.
What good looks like
| Time to first token | Under one second is the point at which a reply feels immediate. Above three seconds, abandonment rises sharply — the person has already decided nothing is happening. |
|---|---|
| Containment rate | The share resolved without a person. 60–80% is realistic for a well-fed assistant on common questions. Anyone quoting 95%+ is either counting differently or answering questions nobody asked. |
| Escalation quality | Measured not by how rarely it happens but by whether the person receiving the conversation has to ask the customer to repeat themselves. |
| Answer grounding | The share of answers traceable to a source document. The number that matters in a regulated sector, and the one most vendors do not publish. |
| Language accuracy | Whether the reply came back in the language of the question. Sounds trivial; is the most common complaint about bilingual assistants in this market. |
| Abandonment | The share who leave mid-conversation. Rises with latency, with menu depth, and with any answer that reads as evasive. |
The failure mode worth asking about
The specific risk with any generative assistant is that it answers confidently when it should not. The industry term is hallucination; in practice it looks like a plausible, well-written, entirely invented answer — an opening hour that was never true, a policy that does not exist, a fee that was made up because the question implied there should be one.
This is not solved by instructing the model to be careful. An instruction is a suggestion, and a sufficiently leading question will override it. It is reduced by grounding every answer in your own material, and caught by checking the finished reply against that material before it is sent — a deterministic check that runs after generation rather than a hope expressed before it.
When you evaluate any assistant, including this one, the question worth asking is not "how accurate is it" but "show me what it does when it does not know". Ask it something adjacent to your material but not in it. A good system declines. A weak one improvises, and it improvises most convincingly on exactly the questions where being wrong costs you most.
What to ask any vendor about this
Ask it something you have not told it
Adjacent to your material but not in it. A good system declines; a weak one improvises fluently.
Ask in dialect
Not formal Arabic. Type the way your customers type, including a mixed Arabic-English sentence.
Ask for the grounding
What share of answers can be traced to a source document, and whether that can be shown per answer.
Ask about escalation
How a customer reaches a person, how fast, and what the agent sees when they arrive.
Ask about retention
How long conversation content is kept, and whether that is a policy you set.
Ask what happens under load
Behaviour when many people are talking at once — a clean refusal is far better than a degraded answer.
How this fits industry practice
The direction of travel across the industry is consistent and worth understanding before you evaluate anyone. Assistants have moved from scripted decision trees, to retrieval over a knowledge base, to systems that can take actions against connected systems. Each step added capability and added a category of risk, and the risks did not replace each other — they accumulated.
Scripted systems were safe and useless: they could only answer what someone had anticipated, so they frustrated everyone and were quietly abandoned. Retrieval systems answer far more but introduce the possibility of answering wrongly with confidence. Action-taking systems answer and do, which raises the stakes again — an incorrect answer is embarrassing, an incorrect action is a liability.
The mature position, and the one worth insisting on, is that capability should be earned rather than assumed. Ground answers in material the organisation controls. Check the finished output rather than trusting an instruction given beforehand. Keep actions narrow, reversible and reviewable. Escalate rather than improvise. None of that is exotic; it is simply what separates a deployment that survives its first year from one that gets switched off after an incident.
The specific thing this market adds is language. Nearly every published benchmark, best practice and vendor claim was developed against English. A capability that performs well in English and adequately in Arabic is not bilingual — it is an English product with Arabic support, and the difference shows up precisely where it costs most: in the questions your customers actually ask, in the words they actually use.
Measuring whether it is actually working
Most assistant deployments are measured on the wrong number. Containment — the share of conversations resolved without a person — is easy to report and easy to improve for the wrong reasons. An assistant that makes escalation difficult will show excellent containment and a deteriorating relationship with customers, and the second effect takes months to surface while the first appears on a dashboard immediately.
The numbers worth watching are the ones that expose deferred problems. Repeat contact within forty-eight hours tells you whether a resolved conversation actually resolved anything. Abandonment mid-conversation tells you where people gave up. The share of escalations in which the customer had to repeat information already given tells you whether handover is working or merely happening. None of these flatter the system, which is precisely why they are useful.
Two more are worth instrumenting from the start. The first is grounding — what share of answers can be traced to a source document. In a regulated sector this is the number an auditor will eventually ask for, and retrofitting the ability to answer it is painful. The second is language accuracy: whether the reply came back in the language of the question. It sounds trivial and it is the single most common complaint about bilingual assistants in this market.
Set a baseline before launch. Without one, every improvement is an assertion. Take a month of existing contact, categorise it, count it, and time it. That baseline is what makes the business case defensible six months later when somebody senior asks what the assistant actually changed.
Misconceptions worth clearing up early
"It will replace the contact centre"
It will not, and deployments sold on that promise tend to fail. It removes the repetitive share of contact and leaves the judgement, the complaints and the complex cases — which is where trained people were always most valuable.
"More training data means better answers"
For a grounded assistant, answer quality depends on the quality and organisation of your own approved material, not on volume. A hundred well-maintained documents outperform a thousand contradictory ones.
"It needs to know everything on day one"
The opposite. Narrow deployments that answer a handful of questions extremely well build trust; broad ones that answer everything adequately lose it at the first confident mistake.
"Arabic support means Arabic works"
Nearly every vendor claims Arabic. The question is whether it handles the Arabic your customers actually type and speak — dialect, mixed Arabic-English sentences, and right-to-left interface behaviour — or only the formal written form.
"Accuracy is a single number"
Accuracy against what, asked by whom, in which language? A figure quoted without a described test set and a stated language is marketing, not measurement.
"We can add governance later"
Consent, retention and audit are far cheaper designed in than retrofitted. The most common reason a working pilot never reaches production is that it cannot pass legal review.
The numbers that matter
Getting this right in practice
Feed it your real material, not a curated subset. The instinct is to give an assistant polished marketing content; the questions people actually ask are answered in policies, procedures and price lists. Quality of answers tracks quality of source material more closely than anything else you control.
Read the transcripts weekly for the first month. Not the metrics — the actual conversations. Every deployment discovers within two weeks that customers ask something nobody anticipated, phrased in a way nobody predicted, and that is worth more than any amount of pre-launch planning.
Decide what it must never say before launch. Competitors, legal advice, medical guidance, anything about individual accounts it cannot verify. Writing that list down is a thirty-minute exercise that prevents the class of incident that gets a project cancelled.
Common questions
Both. Customers do not type Modern Standard Arabic, and a system that only handles formal text will miss most real questions. Matching is done against how people actually phrase things, in both languages.
The reply follows the most recent message. Script is decisive, so an Arabic message gets an Arabic answer even if the conversation began in English. Very short messages are excluded from that decision to avoid flipping on a one-word reply.
No, and that is intended. Answers are assembled from your material. Where nothing covers the question it says so and offers a person.
For the length of the conversation, so follow-ups work naturally. After that, message content is destroyed within 24 hours — usually as soon as the conversation ends.
Yes. Tone, formality, length and what it should never say are configurable per agent, and can differ between Arabic and English personas.
Written chat is the cheapest channel to operate. Cost scales with conversation volume rather than seat count, and the economics improve as containment rises.