The capability is real, the readiness usually isn’t
Models handle major languages well enough that enabling another looks like a toggle. What doesn’t come with the toggle is the knowledge, the review capability, and the escalation path in that language.
1. Your knowledge is monolingual
The system will answer in French from documents written in English, translating policy on the fly. Nuance in terms, obligations and exceptions is exactly where translation drifts — and contradictions in the corpus become harder to spot across languages.
2. Nobody can review it
If no one on the team reads the language, output quality is unverifiable, which fails the basic test that an unverifiable feature shouldn’t ship. Sampling requires someone who can actually read the sample.
Answering in a language nobody on your team can check is not a feature. It’s an unmonitored liability.
3. Escalation lands in the wrong place
A conversation that escalates arrives with a human who may not share the language, converting a hand-off into a dead end. Escalation is only as good as the person on the other end, and that person needs the language too.
4. Obligations differ by market
Serving another language often means serving another jurisdiction, with different disclosure, data and consumer rules. Disclosure requirements in particular vary, and defaults chosen for one market may not transfer.
The workable version
Support the languages you can review and escalate in, say plainly which those are, and offer a clear path for everyone else. Narrow and honest beats broad and unmonitored.