An assistant that answers and acts in Arabic and English

A rule-based chatbot follows a script. If the customer steps off the script, it fails. A language-model assistant understands free text, but on its own it can only talk. An agent adds tools: defined actions it is allowed to take, such as looking up an order by number, checking available slots, creating a ticket or sending a payment link. Each tool is a narrow function with its own permissions, and the model decides which one to call.

That design is also what keeps it safe. The assistant cannot do anything we have not wired up, and risky actions, such as refunds or changes to a contract, can require a human approval step before they run. We write down what the agent may do alone, what needs approval and what it must never touch. That document is reviewed with you before any build begins.

Saudi customers do not write the way textbooks do. Messages arrive in Najdi, Hijazi or Gulf Arabic, in Arabizi, with missing diacritics and English product names in the middle. We build a test set from your real chat logs or support emails, label what the customer meant, and measure the assistant on it. Where a general model struggles, we add examples, glossaries or a separate classification step until it reads your customers correctly.

WhatsApp is usually where the volume is. It runs through the official business platform, which has its own rules on templates, opt-in and the twenty-four hour customer service window, and we design flows inside them. Website chat and in-app chat share the same assistant, so a customer who moves between channels does not start over. Voice is possible, but we advise starting with text.

Conversations contain names, phone numbers, order details and sometimes worse. Under the Personal Data Protection Law you need a lawful basis, clear retention periods and a way to answer access and deletion requests. We set retention on transcripts, mask sensitive fields before they reach the model where we can, and keep a log you can audit. SDAIA's guidance on AI ethics also asks for transparency, so the assistant tells people it is automated and how to reach a person.

Hosting follows the same logic as any AI project. Low-risk questions about opening hours can run on a hosted model API. Conversations touching health, finance or identity documents may belong on infrastructure inside the Kingdom. A standard assistant takes four to eight weeks to launch. An agent with several system integrations and approval steps takes eight to fourteen.