How to build an Arabic AI agent for customer service
A practical guide to an agent that understands Modern Standard Arabic and dialects, answers only from your knowledge, and hands off to a human when needed.
· 9 min read
Start from the knowledge, not the model
The projects that stall are the ones that begin by picking a model. Models are a swapping commodity; what persists is your knowledge base. Gather first what actually answers your customers: FAQ pages, return policies, product descriptions, and transcripts of past support conversations.
Past transcripts specifically are the most valuable source and the most consistently ignored, because they are written in the exact language your customers ask in: their dialect, their typos, their abbreviations.
Why Arabic is harder to retrieve
An Arabic word carries its prefixes and suffixes inside its own structure, so a single token can correspond to a full English clause. Keyword search alone therefore fails often, and semantic vector search is not a luxury here but a requirement.
On top of that, one customer may write "ابغى ارجع الطلب", "أريد إرجاع الطلب", and "i want to return my order" inside the same conversation. An agent relying on text matching treats those as three unrelated questions.
- Use multilingual embeddings so an Arabic question and an English document land in the same vector space.
- Set a similarity floor: returning "I do not know" beats returning the nearest irrelevant chunk.
- Chunk on logical boundaries (headings, paragraphs), not blindly every 500 characters.
Ground the agent in your knowledge
The line between a useful agent and a dangerous one is grounding. An ungrounded model will invent a returns policy you never wrote, confidently, and in flawless Arabic. The rule: with no support in the retrieved context, do not answer.
Plan the human handoff from day one
Handoff is a feature, not an admission of failure. Sharp complaints, anything touching money, and any question the agent has misunderstood twice should reach a human, carrying the full transcript, not restarting from zero.
Frequently asked
- Does an Arabic agent need a model trained specifically on Arabic?
- Usually no. Modern multilingual models answer in fluent Arabic. The higher-return investment is retrieval and embedding quality, not training a bespoke model.
- How many documents do I need to start?
- Twenty to thirty pages covering your most repeated questions is enough for a useful launch. Precise coverage of common questions beats raw volume.
- How do I measure whether the agent is working?
- Three metrics: first-contact resolution, handoff rate, and the rate of ungrounded answers. The third matters most and is usually the one nobody tracks.
Read next
RAG or fine-tuning? The practical difference for business agents
When retrieval-augmented generation is the answer, when fine-tuning is, and how to decide based on how fast your data changes.
Connecting an AI agent to Odoo without breaking user permissions
Most Odoo integrations run as one account that sees everything, bypassing Odoo access rules entirely. Here is the alternative.
A security checklist before launching an AI agent
What to verify before your agent talks to a real customer: prompt injection, tenant isolation, data leakage, and audit.