Running multilingual AI support without silently drifting from your English source of truth

Shipping support in five languages sounds like a localization problem. For AI agents, it’s a consistency problem: the model will happily paraphrase policy in Spanish or Japanese in a way that no longer matches your English terms of service.

The fix isn’t “better translation.” It’s tying every localized answer to the same canonical source.

One authoritative doc per policy, then localize deliberately

Maintain English (or your primary language) as the version legal and product sign off on. Localized articles should be reviewed updates, not machine-translated dumps. The agent should retrieve from the locale-specific doc when it exists, and fall back to primary plus machine translation only when you’ve labeled that path as “best effort.”

Don’t mix free translation with retrieved facts

A reliable pattern: retrieve passages in the primary language, then ask the model to render the answer in the user’s language without adding new claims. The retrieval step stays auditable; the generation step only changes wording.

Surface the language boundary to users

When an answer is synthesized from English sources, a short line — “Based on our English policy” — sets correct expectations and reduces disputes. Transparency beats pretending every locale is equally maintained when it isn’t.

• Track which locales have human-reviewed docs vs. machine-assisted only.
• Re-review high-risk topics (refunds, data, compliance) per locale before AI-first rollout.
• Measure thumbs-down by language; a spike often means drift, not tone.