Answer Intercom chats in any language, with a meaning check

Build me a multilingual support desk my English-speaking agents open at the start of every shift, so they can work foreign-language Intercom conversations without switching tabs or pasting customer messages into a translation site. The whole point is that a monolingual agent can read the thread, write a reply, verify it means what they intended, and send it, all on one screen.

The surface is a two-pane inbox. The left pane is a list of open conversations pulled with Intercom Search Conversations, filtered to open state. Each row shows the customer, a snippet of their latest message, a badge for the detected language, and how long the conversation has been waiting since the customer last wrote. Detect the language by sending the latest customer message through DeepL Translate Text without specifying a source language and reading the detected source language DeepL returns. Batch the detection: Translate Text accepts up to 50 strings in one call, so detect the whole visible page of conversations in a single request rather than one call per row. Sort so the longest-waiting non-English conversations sit at the top, and let the agent filter to non-English only.

Clicking a row opens the full thread in the right pane using Intercom Retrieve a Conversation, which returns the conversation parts. Render it oldest first. Every customer message shows as the original text with its English translation stacked directly underneath in a visually quieter style, so the agent reads the English but can always see what was actually written. Translate the whole thread in a single Translate Text call by passing the message bodies as an array of strings. Do not translate one message per call. Label each part with its author and timestamp, and mark internal notes so the agent can tell them apart from what the customer saw.

Below the thread is a composer where the agent types their reply in English. Before anything sends, a preview panel shows two things: the DeepL translation into the customer's language, and a back-translation of that target text returned to English. The back-translation is a second Translate Text call that takes the translated output as input and targets English, so the agent sees the round trip rather than a claim that the translation is fine. Show the agent's English, the outbound translation, and the back-translation together so they can compare meaning at a glance, and highlight the back-translation as the thing to read. Debounce so it does not re-translate on every keystroke, and never let Send be reachable without a rendered preview.

Send posts the translated text to the customer using Intercom Reply to a Conversation as an admin comment, so the customer reads their own language. Immediately after, make a second Reply to a Conversation call with the note message type carrying the agent's English original, so teammates picking the thread up later read what was actually meant rather than reverse-engineering it from machine translation. These are two separate calls: if the customer reply lands but the note fails, say so plainly in the UI instead of silently dropping half the record. Get the sending admin from Identify the Logged-In Admin.

Add a draft a reply button that kicks off a background agent. The agent reads the whole translated thread, works out what the customer is actually asking, and drafts a response in English grounded in the specifics of the conversation rather than generic filler. It then runs that draft through DeepL Rephrase Text (Write) to tighten the tone, and drops the result into the composer for the agent to edit. It never sends and never calls Reply to a Conversation itself. Show that it is running and let the agent keep reading the thread while it works. Once the draft lands, the translation and back-translation preview render from it like any other draft.

Bake in three pieces of language nuance. First, the detected language is a suggestion, not a verdict: let the agent override it from the full list returned by DeepL List Supported Languages, and re-render both the thread translation and the reply preview against the override. Second, prefer regional target variants where the customer's locale suggests one, so a Brazilian customer gets PT-BR rather than PT-PT and a UK customer gets EN-GB. Pull the contact's locale with Intercom Get a Contact to inform that choice, and remember the agent's per-conversation override. Note that DeepL source languages use base codes only (EN, PT, ZH) while target languages accept the regional variants, so only apply variants on the outbound side.

Third, let the team pick a DeepL glossary from List Glossaries in a settings view so product and brand names are never translated. DeepL glossaries are monolingual, meaning each one covers a single source to target pair, so apply the glossary whose pair matches the direction being translated and quietly skip it when no glossary matches. Show which glossary is in effect on the preview panel so the agent knows whether their terms are protected.

Cache inbound thread translations per conversation so reopening a chat does not re-translate and re-bill the same messages. Keep the conversation list visible while a thread is open so an agent can work through several chats in one sitting without navigating back and forth.