Playbooks · 4 min read
Do You Actually Have a Multilingual Inbox? Check Before Building
The mechanics of running automations across languages are already written down twice on this site. What the matcher folds away and what it does not is the main one, and multi-language keyword lists covers the same ground from the list-building side. Neither needs restating and this post is not going to try.
What is missing is the step before them, which almost nobody does: finding out whether you have a multilingual inbox at all. A large share of accounts that go looking for multilingual automation turn out to have a different, smaller, much more fixable problem.
The audit, which takes about ten minutes
Open your notifications. Read the last fifty comments and DMs — actually read them, as strings, not as a general impression of your audience. Tally each one into three columns: which script it was written in, which language, and what it was asking for.
That is the whole exercise. The result lands in one of three shapes.
- One script, one language, plus English loanwords.
price,size,delivery,linktyped in Latin script by people whose sentences are otherwise Hindi or Tamil. This is by far the most common outcome and it is not a multilingual problem. - Two scripts, one language.
daam,damandदामall arriving for the same question. Also extremely common in India, also not really multilingual — it is one intent wearing several spellings. - Genuinely different languages needing different answers. Your audience splits across languages you would reply to differently, and the reply itself changes rather than just the trigger.
Most inboxes described as multilingual are one language with eleven spellings, which is a list problem rather than a strategy problem.
What each result means you should do
Shape one and shape two are the same fix. One automation, one keyword list on the any-of setting, containing every spelling you saw. Not a language strategy — just a longer list, harvested rather than invented, because the version in your head is the formal one and the version in your inbox is thumb-typed at a bus stop.
The thing that makes this work is that matching folds case and punctuation but nothing else. price does not match pricing. kitna does not match kitne. दाम does not match दम, because a matra is a vowel rather than an accent, and folding it would collapse the word for price onto the word for breath. So every spelling is its own entry, and the ten minutes you just spent reading comments is where the entries come from.
Shape three is the only one that needs more than one automation. Same post, two automations, two keyword sets, two replies in two languages. Templated replies are free and unlimited, so having four language variants of the same answer costs nothing at all — which is worth knowing, because people assume the multiplication has a price and design around it.
What the audit tells you that a strategy cannot
Two things, both worth more than the language question itself.
Which questions you actually get. Reading fifty comments as strings is the single highest-return maintenance task in this product, and the language finding is a by-product. The main finding is that four questions cover most of the inbox, whichever languages they arrive in.
Which spellings you are currently missing. Every one is somebody who asked and got nothing, and adding it takes ten seconds. Do this once a month, especially after a post travels beyond your usual audience, because those people write differently — they are not your regulars.
The thing the audit usually finds instead
Almost every time somebody runs this, the headline finding is not about language. It is that one question they had never templated accounts for a large share of the inbox, arriving in four spellings, and nothing was set up to catch any of them.
That is worth knowing before you spend a week designing a language strategy, because the fix is one automation and an afternoon. The language work, if you genuinely need it, is easier to do afterwards on a setup that already answers the common case.
What no amount of this fixes
There is no translation feature, no language detection setting, and no auto-localisation of templates. Nothing decides what language to answer in on your behalf; you decide, by which automation you built.
And matching runs before generation, so a message that matched no keyword never reaches a model at all. Improving the AI does nothing for coverage in any language. The matching mechanics in full is the reference if you want to know exactly what the comparison does before you write the list.
If your audience is primarily Indian, what else is different is the post that goes with this one. If your audience is spread across timezones rather than languages, that is a genuinely separate problem and it is an operational one rather than a configuration one.
