AI Training for Teams Who Work in Sinhala and Tamil

By Ammar Ahamed, Founder, Cocoon · 21 October 2026 · Part 4 of 8 in the Corporate AI in Sri Lanka series

Two rooms, one company

Picture two training rooms in the same company on the same day. In the head office in Colombo, a marketing team is building a content calendar with an AI assistant, typing quick English prompts and arguing about the results. Two hours away, at a factory, a hotel or a regional branch, a team of supervisors is listening politely to the same session, translated in their heads, and leaving with very little they can use.

Both rooms are the same company's investment. Only one of them got its money's worth.

Most AI training in Sri Lanka is built for the first room. If we want AI to change how a whole company works, and not just its head office, we have to design for the second room too.

What the Census says about language at work

The Census of Population and Housing 2024, published by the Department of Census and Statistics in April 2026, asked people aged ten and over whether they could read and write in each language. The answers include people who rated themselves as poor, so they describe literacy rather than fluency, but they are still telling.

English literacy has nearly doubled since the 2012 Census, when it was 30.8%. But the spread is wide. It reaches 74.2% in Colombo district and falls to 31.7% in Mullaitivu. Among people aged sixty and over it is 35.0%. The same census puts digital literacy at 67.6% for people aged five and over, and computer literacy at 34.7%.

For an employer, the message is simple. The further a team is from head office, and the more it works with its hands or directly with customers, the less safe it is to assume an English-only AI session will land.

Why AI answers thinner in Sinhala and Tamil

Large language models learn from enormous amounts of text, and the internet is mostly English. Sinhala and Tamil have far less written text online, which shows up in the answers: thinner vocabulary, awkward phrasing, more mistakes on local terms, and sometimes a reply that quietly switches to English.

There is a technical reason too. Models break text into small pieces called tokens before they read it, and scripts that the model saw less of during training are often split into many more pieces. More pieces means the model works harder for the same meaning. Our free Micro LLM lab shows this happening, step by step, with Sinhala and Tamil examples.

None of this means Sinhala- and Tamil-speaking teams should be left out. It means the training has to handle the gap deliberately instead of pretending it is not there.

The bridge method

The approach that works best is to treat English as the working layer for the model, not as a requirement for the person.

People describe what they need in the language they think in, Sinhala, Tamil or a mix. That request is turned into a clear, structured English prompt. The model answers in English, where it is strongest. The answer comes back in both languages, so the person can understand it fully and check it against what they asked.

We built a free tool that does exactly this, Prompt Bridge. Type a request in Sinhala, Tamil or English, and it builds a proper English prompt, runs it, and shows you the answer in both languages. It is a useful way to show a mixed-language team that AI is for them too, and a good first exercise in any session.

How to design training for mixed-language teams

Ask before you design. Before booking anything, ask people which language they plan their work in. It takes one question in a survey. The answers often surprise head office.

Teach in the room's language, show in English. Explanations, discussion and questions can happen in Sinhala or Tamil. The screen can show English prompts, with the bridge method as the bridge. What matters is that nobody has to translate in their head while also learning something new.

Keep groups small. People are much more willing to say "I did not follow that" in a group of eight than in a hall of eighty, especially when the question is about language. Pairs work even better for hands-on time.

Pick champions who share the language. The person who helps a supervisor at a factory or a front office team at a hotel should be someone they talk to every day, in the language they use every day. That matters more than how technical the champion is.

Use voice. Many people speak more comfortably than they type, especially in Sinhala or Tamil script on a phone keyboard. Most AI assistants now accept voice input on phones. Showing people they can simply talk to the tool removes a surprising amount of friction.

Use their real documents. A shift handover note, a guest message, a supplier email, a safety checklist. The closer the exercise is to their actual day, the less language gets in the way.

Frontline tasks worth starting with

Every one of these must follow your company's data rules. Customer names, phone numbers and NIC numbers stay out of AI tools unless a use has been approved. With Sri Lanka's data protection duties for companies due on 1 January 2027, that rule matters as much at a regional branch as in head office. We cover it in the PDPA and AI at work.

Teach checking before speed

When an AI answer comes back in a language you read less comfortably, it is harder to spot when it is wrong. That makes checking the most important skill to teach mixed-language teams, and it should come before anything about saving time.

The habits are simple. Ask for the answer in both languages and compare. Check names, numbers and dates against the source, every time. Ask the tool what it is unsure about. And when something will reach a customer, have a fluent reader look at it first.

Teams that learn this early end up more careful with AI than many head office teams, not less.

Where to start

If your company has teams outside Colombo, or teams who mostly work in Sinhala or Tamil, include them in the first phase of your AI rollout, not the last. They are often where the time savings are largest, because nobody has given them tools before. Our 90-day rollout plan shows where they fit.

When you talk to training providers, ask directly: how will you work with people who are less confident in English? Tell us which languages your teams work in, and we design the programme around it.

One thing to try this week: sit with a frontline supervisor for ten minutes and open Prompt Bridge together. Ask it to help with one real task from their day, in their language. Watch their face when it works.

Frequently asked questions

Sources

1. Department of Census and Statistics, Census of Population and Housing 2024, Final Report
2. Cocoon, Prompt Bridge (free tool)