AI, Cantonese and English: what breaks in a Hong Kong office

Half the meetings in this city run in two languages at once. Someone makes a point in English, qualifies it in Cantonese, then switches back for the number. Every tool you point at that conversation handles it worse than it handles either language alone, and none of them say so. Running AI across Cantonese and English in a Hong Kong office fails quietly, which is the expensive way to fail.

The failure is not mistranslation. It is omission. The qualifying clause spoken in the other language does not come back wrong, it does not come back.

What actually breaks

Three things, and they are different problems.

The transcript mangles the switched span. Sometimes it drops it, sometimes it renders Cantonese as whatever English it sounded closest to, and the sentence still scans, so nothing looks broken.

The summary then rebalances toward the dominant language. If eighty per cent of a call ran in English and the objection was made in Cantonese, the objection is the single most likely thing to be missing from the summary you send round.

Register flattens last. A lot of meaning in a Cantonese negotiation sits in how softly something is said. Put through a model, a polite deferral and a firm refusal arrive on the page looking identical.

There is a fourth, smaller one: names. Mixed English and Chinese names, romanisation that varies from person to person, and a model that will quietly standardise them to whichever spelling it has seen most.

It does not come back wrong. It comes back missing, and a missing objection reads like agreement.

Where AI, Cantonese and English cost you in a Hong Kong week

The risk concentrates exactly where the stakes are. The switch into Cantonese usually happens on the delicate part of a conversation: the price, the objection, the thing nobody wants minuted in English. So the material you are most likely to lose is the material you most needed.

That makes four jobs worth treating carefully. Minutes going to people who were not in the room. Client call notes typed into a CRM, where the summary quietly becomes the record. WhatsApp threads with clients, which code-switch far more than meetings do. And anything that will be read by someone with no way to check it against the audio.

The ten-minute test

Do not judge this on a vendor demo, which will be in one language. Take a real recording where you already know what was said. Pick the three hardest minutes, the ones that switched most, and run only those. Read the output against your own memory of the conversation.

You are checking one thing: did anything said in Cantonese survive into the summary at all? Run it again whenever you change tool or model, because performance here moves.

What helps

• Say it once in one language. The decision and the number, out loud. "To confirm in English: 3.2 per cent, by the 30th." Four seconds, and the record now has an anchor.
• Hand the model a glossary before it starts. The names in the room, the firm names, the handful of terms your desk uses.
• Keep the audio. A summary you cannot check against a source is not a record.
• Split the job. Transcribe, correct the transcript yourself, then summarise the corrected text. The same logic applies as in how to check AI for errors: verify the input, not the polish.

The prompts that force a glossary and summarise from a corrected transcript rather than from memory are in the free prompt library.

The honest part

This is improving, and it is improving unevenly. Written Chinese is handled well. Spoken colloquial Cantonese, mid-sentence, against background noise, much less so. Any flat claim that AI cannot cope with Cantonese is already going out of date, this one included, which is why the advice above is a test you run rather than a table of which tool to buy.

None of it argues for doing less. Meeting notes are still the first thing I would automate for most people, for the reasons in AI note-taking for meetings, and a standing assistant still earns its place across a bilingual week, which I set out in what an AI chief of staff actually does. It argues for knowing which part of your week is bilingual and handling that part differently.

That map is specific to your desk, your clients and which conversations you cannot afford to lose. If you want to work out which parts can be handed over and which need you in the loop, start here.