Can AI speak your customers' language yet?

Most AI demos are in English, and most of our clients' customers do not only speak English. In Dubai a conversation can start in English, switch to Gulf Arabic when it gets specific, and end with a voice note. In Kerala, people often type Malayalam in English letters, mixed with English words, in the same sentence. Any assistant that talks to customers here has to cope with that.

The good news is that the past six months have brought a burst of serious work on our region's languages.

What has arrived

In the UAE, the Technology Innovation Institute released Falcon-H1 Arabic in January, in three sizes up to 34 billion parameters, with long memory for documents. In December, Inception and MBZUAI released Jais 2, in 8 and 70 billion parameter versions. In Saudi Arabia, Humain launched Humain Chat last August, an Arabic-first assistant built on its ALLaM model, with voice input across dialects.

India has moved too. At the AI Impact Summit in New Delhi last month, Sarvam unveiled two models trained from scratch, at 30 and 105 billion parameters, and on 6 March it released them as open source under the Apache 2.0 licence. They cover all 22 scheduled Indian languages. The government-backed BharatGen project showed Param2, a 17 billion parameter model covering the same 22 languages.

Meanwhile, the big general models from OpenAI, Google and Anthropic have kept improving at Arabic and Hindi without anyone making much noise about it.

Where AI handles our languages well

For written Modern Standard Arabic and formal Hindi, the leading models are now good enough for most business text: answering questions from your documents, drafting replies, summarising, translating your own content with review. That was not true two years ago.

They also cope better with mixing than people expect. A question that is half English and half Arabic, or Hindi typed in Latin letters, usually gets understood.

Where it still slips

The gaps are more specific than "Arabic is hard".

• Dialects. Gulf, Levantine and Egyptian Arabic differ a lot from the formal written language. Understanding is usually fine; replying naturally in a customer's dialect is weaker, and a reply in stiff formal Arabic can feel cold in a WhatsApp chat.
• Arabizi and romanised Indian languages. Arabic written in Latin letters and numbers, or Malayalam typed in English letters, varies by person. Models understand a lot of it, but quality drops, and spelling varies so much that search over your own content can miss.
• Smaller Indian languages. Hindi gets the most training data. Malayalam, Kannada, Odia and others get much less, and it shows in fluency.
• Voice. Speech recognition for dialects and accents still lags text. If your customers prefer voice notes, test that path separately.
• Your own vocabulary. Product names, local place names, legal and medical terms. A model can translate "deductible" correctly in general and still use a different word from the one your policy documents use.

Big general model or regional model?

For most customer-facing assistants today, we start with a leading general model and test it hard in the customer's language. It is usually the strongest at following instructions, using tools and admitting when it does not know.

Regional models earn a place in three situations. When data has to stay in the country or on your own servers, an open model like Falcon-H1 Arabic, Jais 2 or Sarvam can run where you choose. When volume is high and the task is simple, a smaller model can be much cheaper. And when a regional model clearly beats the big ones on your own test, which Sarvam claims for Indian languages on its benchmarks. Claims made by any model's own builder should be checked against your own examples before you rely on them.

How to test before your customers do

The only test that counts is one built from your customers' real messages.

1. Collect a hundred real customer questions from WhatsApp, email and calls, in the languages and scripts people actually use, typos included.
2. Write down what a good answer looks like for each, with help from a native speaker on your team.
3. Run two or three models against them and have that native speaker grade the answers blind.
4. Pay special attention to the wrong answers. A fluent wrong answer in Arabic is more dangerous than a clumsy right one.
5. Keep a glossary of your own terms in each language and give it to the model as part of its instructions.

A few rules we would not bend

Let customers choose their language, and answer in the one they wrote in. Do not machine-translate your whole knowledge base and publish it without a person reading it. Keep anything legal, medical or financial under human review in every language. And always offer a person who speaks the customer's language when the AI cannot help.

Done this way, an assistant that handles Arabic, Hindi or Malayalam properly is a real advantage. Plenty of businesses here still answer only in English.