The agent that worked in English but completely fell apart in production multilingual support
Lesson from a real deployment.
We built a customer support agent for a logistics company. Worked beautifully in English testing. Shipped to production across SEA, where customers write in English, Vietnamese, Thai, and Bahasa Indonesia, often mixed in one message.
What broke:
- The agent confidently answered nonsense for mixed-language inputs
- It treated transliterated English ('lah,' 'sis') as code-switching it didn't recognise
- Region-specific phrasing for shipping ('parcel lor' vs 'package') broke intent detection
The fix wasn't a better model. It was switching to Gemini 1.5 Pro specifically for the multilingual cases, while keeping GPT-4o for English-only flows.
Lesson: production language coverage is harder than the benchmarks suggest. Multilingual isn't a checkbox, it's a routing decision.
Replies
The "multilingual isn't a checkbox, it's a routing decision" line is the real takeaway here, and it generalizes past language. We see a version of this in coding agent tooling: a model that's great on a clean, well-typed codebase falls apart on a legacy repo with inconsistent conventions, and the naive fix people reach for is "swap in a smarter model" when the actual fix is detecting which situation you're in and routing accordingly. The hard part in your case is that transliterated English and code-switching are genuinely ambiguous signals, not a clean language tag. Did you end up doing that detection with a cheap classifier pass before the main call, or did you just look at metadata like region/locale and use that as a proxy instead of trying to classify the text itself?
@galdayan Thanks for pushing on this, honestly the language routing failure is one of the most underrated production issues in AI agents right now. Everyone tests in English and calls it done, then gets blindsided when real users switch mid-conversation or use mixed-language queries.
The example about "language-specific tool calls" is spot on. We saw the same thing with an agent that worked perfectly in English but couldn't parse a Vietnamese date format properly, so it kept calling the wrong CRM field. Not a model problem, a context and normalization problem. Adding a language detection layer before tool routing fixed 80% of our issues, but the last 20% (mixed-language inputs) is still where we lose accuracy.
The bigger unlock for us was treating language as a first-class variable in the agent design, not an afterthought. Prompts, tool schemas, and even error messages all need locale-aware handling. Curious how you're handling regional dialects too, because that's where things get really messy fast.
@nolan_vu honestly we haven't cracked dialects either, we punted on it. what we do instead is skip trying to classify dialect at all and just widen the "acceptable answer" surface - the agent asks a clarifying question instead of guessing when confidence is low, rather than trying to silently resolve ambiguity upstream. feels less elegant than a real dialect model but it fails a lot safer in prod. the 80/20 split you described matches what we see too, the long tail of mixed-language stuff is where a confident wrong answer does more damage than a "sorry, can you rephrase that."
@galdayan Thanks Gal, honestly this reply just saved me a few weeks of over-engineering. The idea of skipping dialect classification entirely and just widening the "acceptable answer" surface is such a smart pragmatic move. We've been trying to solve dialect detection at the model layer and it's been a rabbit hole with diminishing returns.
The "confident wrong answer does more damage than sorry-can-you-rephrase" framing is exactly right, and it's honestly the thing most AI teams underweight. In production, humility from the agent actually builds trust way faster than trying to look smart on every response. Going to steal the clarifying-question-when-confidence-is-low pattern for our next iteration. Really appreciate you sharing what worked (and what you punted on), that kind of honesty is rare 🙌
NINA
This feels like a broader lesson for AI products.
We spend a lot of time choosing the "best" model, but in production the bigger advantage often comes from knowing when not to use that model.
Good routing, confidence thresholds, and graceful fallbacks usually matter more than squeezing another 2% out of a benchmark. A humble agent that asks for clarification will earn more trust than one that confidently guesses wrong.
@varun1jan I agreed with your point. AI product is already a large topic to explore so it is crucial to focus on specific niche