One setting decides what the agent writes in and what the widget's own buttons say.
Open the agent's Settings, then the Language card.
Assistant language
Pick one language from the list. It does two things:
- The agent answers in it by default.
- The widget's own chrome is translated into it — buttons, labels, placeholders, the AI disclosure. Twenty-eight languages are supported.
Follow the visitor
A checkbox under the picker. On by default.
- On — the agent answers in your language by default, and switches when a visitor clearly writes in another one. The widget's buttons and labels stay in your language either way.
- Off — the agent answers in your language throughout, even when a visitor writes in something else.
The chrome deliberately never follows the visitor. It has to paint before anyone has typed, so there is nothing to detect yet, and re-labelling buttons mid-conversation under someone's cursor is worse than one consistent language.
How switching actually works
There is no language detector. The model decides, reading the whole conversation rather than one message in isolation — which makes it far better at this than a statistical guess on a short string.
It is told to ignore the things that used to cause false switches: a name, a place, a currency symbol, a single foreign word quoted inside an otherwise English sentence. When in doubt it stays in your language.
The visitor's browser language is used as a hint to disambiguate very short messages, never as a trigger on its own. Their country is deliberately not used at all — where someone is says very little about what they want to read.
If you previously set a "forced language", it was carried over as your language with Follow the visitor switched off — the same behaviour you had. Nothing to redo.