Arabic AI agent: what it is and how to build one for your business

What is an AI agent, and how does it work?

An AI agent combines a language model with knowledge, tools and instructions. The model interprets the request; retrieved documents or live system results provide evidence; connected tools let the agent perform an allowed action. An Arabic agent needs this same setup plus testing for dialects and mixed Arabic–English messages.

A chatbot can explain how to open a support ticket. An agent connected to the ticketing system can also prepare or create it, subject to its permissions. The label alone proves little: ask the provider to demonstrate a complete task and show what happens when permission is missing or approval is required.

Example: checking an Odoo order and preparing a reply

Consider an illustrative workflow, not a customer case study: an employee asks, “Check the order for this customer and draft a reply.” The agent must identify the customer, read only records the employee can access, retrieve the current status and prepare a response based on that result.

Sending the response is a separate action. If company policy requires approval, the agent presents the draft and waits. It should not treat a failed lookup as evidence that an order does not exist, or retry a write blindly when the result is uncertain. Record the lookup, proposed action and final outcome for review.

• Confirm the customer or order identifier when several records match.
• Fetch order status from Odoo at request time rather than from an old document.
• Test denied access and rejected approval alongside the successful workflow.

What makes an Arabic assistant different?

An Arabic assistant reads requests in Arabic, finds the right information or takes the right action, and replies in Arabic. The same definition fits an English assistant, but Arabic adds problems that sink projects which treat it as “English with a different alphabet”.

People write Modern Standard Arabic in formal documents and their own dialect everywhere else. A Saudi employee may write «ابغى أعرف حالة الطلب», a Jordanian colleague «بدي أعرف شو صار بالطلب», and both mean the same thing. Many users also switch languages mid-sentence, especially for technical terms: «السيرفر رجع down بعد الـ deployment» is a perfectly normal Arabic sentence in an IT team.

Arabic morphology adds a second layer. Prefixes and suffixes attach to the word itself, so «وسيكتبونها» is one token that carries a whole English clause. Some users even type Arabic in Latin letters with digits for missing sounds (3 for ع, 7 for ح), usually called Arabizi. An assistant that relies on exact text matching fails on all of this.

Start from your data and systems, not the model

The projects that stall are the ones that begin by choosing a model. Models are replaceable and improve every few months; what persists is the information your assistant answers from and the systems it can act on. Decide first what it must know and do, then pick the model that handles it well in Arabic.

For a customer-facing assistant the core sources are policies, product information and past support conversations. For an internal IT or operations assistant they are ticket history, knowledge articles, runbooks and live monitoring data. In both cases, past conversations and tickets are the most valuable and most ignored source, because they are written exactly the way your users write: in their dialect, with their typos and their abbreviations.

• List the 30 to 50 questions or tasks users actually bring, in their own words, before writing any prompt.
• Map each one to the document or system that holds the true answer.
• Mark which answers change daily (status, stock, incidents) and must be fetched live rather than copied.
• Keep a separate set of real Arabic questions, across dialects, as your evaluation set.

How do you make Arabic text searchable?

Retrieval quality decides Arabic assistant quality far more than model size. The same word can be spelled several ways, and a search that treats those spellings as different words returns nothing, or returns the wrong passage. Normalise both the stored text and the incoming question with the same rules, and keep the original text for display.

Combine two search methods. Semantic search with multilingual embeddings puts an Arabic question and an English document in the same vector space, which handles paraphrase and dialect. Keyword search still wins on exact identifiers such as ticket numbers, error codes and product names. Running both and merging the results is more reliable than either alone.

• Split documents on logical boundaries (headings, paragraphs, ticket fields), not every fixed number of characters.
• Set a minimum similarity score: returning “I don’t know” beats returning the nearest irrelevant passage.
• Test retrieval separately from generation, with your real dialect questions, before judging the model.

Ground every answer: no source, no answer

The line between a useful assistant and a dangerous one is grounding. An ungrounded model will invent a refund policy or a root cause you never wrote, confidently and in flawless Arabic. Fluency makes the error more convincing, not less.

Make the rule explicit in the instructions and enforce it in the application: if nothing retrieved or fetched supports the answer, the assistant says it does not know and offers the next step. Show the source with every answer, whether that is a document title, a ticket number or the name of the system that was queried, so users can check it.

For information that changes by the minute, such as incident status or monitoring data, do not copy it into a document index at all. Let the assistant call the system of record at question time through a tool, and answer from what comes back.

Arabic on voice and phone channels

Voice multiplies the dialect problem. Speech recognition that works on news-style Arabic can struggle with fast Gulf or Levantine speech, background noise and English technical terms spoken with an Arabic accent. Test with recordings of your actual users, not with scripted readings.

Numbers, names and codes are where voice fails most. Have the assistant read back anything it will act on, such as an account number, a ticket reference or a date, and confirm before continuing. Let callers interrupt: finishing a long sentence while the caller is correcting it feels broken in any language.

Botify runs the same governed setup in Arabic and English across web chat, browser voice, phone calls, email, Telegram, WhatsApp, Microsoft Teams and Slack, with live transcription and interruption in browser voice. Whatever platform you choose, check that channel choice does not change what the assistant is allowed to do.

Right-to-left output: reports, tables and mixed text

Arabic output is not finished when the words are right. Text that mixes Arabic with English names, numbers and codes is laid out by the Unicode Bidirectional Algorithm (UAX #9), and small mistakes in direction handling reorder punctuation, flip ranges like “10–20” or push an English product name to the wrong end of the line.

Set the base direction explicitly for Arabic content in the interface, and test the cases that break: a sentence that ends with an English word, a list of error codes, a table with numbers and currency. Generated files matter too; a PDF or spreadsheet report must render right-to-left with its columns in the order an Arabic reader expects.

Plan the handoff and measure what matters

Handoff is a feature, not an admission of failure. Sharp complaints, anything touching money or contracts, and any request the assistant has misunderstood twice should reach a human with the full conversation attached, so nobody asks the user to start again.

Measure three things from the start: how often the assistant resolves the request without a human, how often it hands off and why, and how often it answers without a supporting source. The third is the most important and the one teams usually forget to track. Review failed Arabic conversations weekly and feed the missing spellings, dialect phrases and sources back into the system.