Playbooks · 5 min read
Running Automations for an Inbox in Several Languages
If your audience writes to you in more than one language, the thing to understand
first is mechanical rather than strategic: a keyword trigger compares strings. It
does not translate, and it does not know that price, kitna and கிமத் are
the same question.
That single fact determines almost every decision below, so it is worth being precise about what the matcher does before deciding what to do about it.
What folds away before the comparison
Three things are normalised, and knowing which three stops you over-planning for them.
- Case.
PRICE,Priceandpriceare one keyword. You never list capitalisation variants. - Punctuation and spacing.
price?,price!!!andpriceall matchprice. Emoji and invisible characters are stripped too, soprice 🔥matches. - Latin accents.
cafématches a keyword ofcafe;espanolmatchesespañol. This applies to accented Latin letters and nothing else.
What does not fold, and is therefore your job
Script, language, spelling and word endings. Each one is a separate entry in the list.
pricedoes not matchpricing. Matching is whole-word, so plurals and endings are their own entries.kitnadoes not matchkitneorkitni. Different strings.pricedoes not matchकीमत. Different scripts entirely.damdoes not matchdaam. Transliteration has no standard spelling, and your audience uses all of them.
The matcher is not multilingual. Your list is, or your automation is not.
Harvest the list, do not invent it
This is the whole technique, and it is unglamorous.
Open your notifications and read the last fifty comments and DMs as strings. Write down the exact words people typed, including the misspellings. That list is your trigger set. It will look nothing like the list you would have written from imagination, because the version in your head is the formal one and the version in your inbox is thumb-typed.
A real harvested list for one Indian clothing account looked roughly like:
price, pric, rate, kitna, kitne, kitne ka, kitna hai, daam,
dam, कीमत, कितने का. Eleven entries for one question, and every one of them
came from a real comment.
Trigger on the easy word in the sentence
When a question can be asked several ways, look for the word that survives all of
them. In "is this available in blue", the load-bearing word is available, not
the sentence. In a mixed-language inbox, the surviving word is often the English
loanword — size, delivery, book, price — because those get typed in Latin
script even by people writing the rest in Hindi or Tamil.
That is not a rule you can apply blindly. It is a thing to check against the fifty comments you just read.
Decide which language the reply is in
Two separate decisions, and people conflate them.
The public reply
The private reply
You can run separate automations for this: one keyword set with a Hindi reply, another with an English reply, both scoped to the same post. Templates are free and unlimited, so having four language variants of the same answer costs nothing.
Where a generated reply helps, and where it does not
A model can compose a reply in the language of the message, which is genuinely useful for the long tail of questions you have not templated. What it cannot do is fire an automation — matching happens before generation, so a message that matched no keyword never reaches the model at all.
That ordering is the thing to internalise. Improving the AI does not improve your coverage. Only the keyword list does.
A workable setup
- One automation per intent, not per language. All eleven price spellings go in one keyword list, on the any-of setting.
- Separate automations only when the reply differs. Same question, two languages, two replies, two automations scoped to the same post.
- Negative keywords in every language too.
not interested,nahi chahiye,rehne do. A refusal in the wrong language is still a refusal. - Re-harvest monthly. New spellings arrive constantly, especially after a post travels beyond your usual audience.
What this does not do
There is no translation feature, no language detection setting, and no auto-localisation of your templates. Nothing here decides what language to reply in on your behalf — you decide, by which automation you built.
If you want the underlying mechanics of triggers in more depth, the keyword guide covers them. If your audience is primarily Indian, the India post covers what else is different.
