Google Cloud's Translation pricing page, two DeepL help-center articles and its pricing page, four Intercom help-center articles (Set up Fin AI Agent's multilingual support; Automatic language detection in conversations, 13 May 2026; Multilingual workflows, 30 July 2026; Understanding translation behavior in Intercom) and Tidio's Lyro Multilanguage support article (31 October 2025) were read on 5 September 2026, and every price, count and rule below carries that date; they change without notice. Every count of our own reviews is a tally of the fifteen files matched by sample-reviews/*-review.md, with the search string printed in Sources.

Machine translation· AI and NLU

Machine translation (MT) is software that converts text from one language into another without a human writing the target text. The systems in production today are neural: either a dedicated neural machine translation model trained on paired sentences, or a general-purpose large language model prompted to translate. In a chatbot, MT can run at four points: on the bot's content when it is built, on the customer's message as it arrives, on the bot's reply as it leaves, and on the conversation as a human agent reads it. Which of those a platform does, and whether it meters them, decides both the bill and the failure modes.

By Chatbotscape Editorial· Methodology· Published 6 September 2026· Updated 6 September 2026

Machine Translation — What It Is, Where It Runs in a Chatbot, and What the Meters Say

Quick answer: Machine translation converts text between languages by software. The engines behind chatbots are neural, and many of them are now ordinary large language models told to translate; Intercom's help center says plainly that it "uses AI and large language model (LLM) providers for translations" and that two tools producing different wording for the same sentence is "normal," not an error. The question for a buyer is not whether the engine is good (the WMT24 evaluation of eight LLMs and four online providers is titled "The LLM Era Is Here but MT Is Not Solved Yet") but where in the conversation it runs and what it costs there.

Read on 5 September 2026, Google Cloud charges $20 per million characters for its neural model after a free 500,000 a month; DeepL sells a Growth plan with one million characters included and meters the rest by the character; Intercom shows no separate meter for real-time translation in its help articles, bills the answer as a Fin outcome, and lets a workflow auto-translate its messages and buttons; Tidio's Lyro does not translate at all, because its model answers in the customer's language directly. A 1,000-conversation month costs $40 to $50 on Google's meters, is invisible on Intercom's, and does not exist on Tidio's. The rule that survives all four: translate the reply, never the knowledge it was drawn from, and decide the language once per conversation, not once per message.

Two ways a machine translates, and why the difference now shows up on invoices

The first generation still in production is neural machine translation (NMT): a model trained on millions of sentence pairs to map a source sentence to a target sentence, one language pair at a time or many pairs in one multilingual model. It is fast, cheap and deterministic enough that the same input returns the same output. Google's Cloud Translation pricing page calls this "the default NMT model," charges it on input characters only, and prices it at $20 per million characters after the first 500,000 in a month, which the page describes as "applied as $10 credit every month."

The second is translation by a large language model, either a general model prompted to translate or a model tuned for it. The same Google page prices its "Translation LLM" at $10 per million characters of input plus $10 per million of output, notes that it "meters the number of input and output characters independently, so the number of input and output characters are likely to be different," and calls the result "cost equivalent with NMT." Its Adaptive Translation, listed on the same page as an LLM method, is $25 in and $25 out. A custom-trained model is $80 per million on the first 250 million characters, plus $45 an hour to train.

The practical difference is not accuracy, which the WMT24 shared task found to be close at the top and unsolved everywhere ("LLM era" in the title, "not solved yet" after it; the paper evaluated eight LLMs and four online providers with professional annotators marking error spans across eleven language pairs). The differences that matter to a chatbot are three. An LLM translator can be told things ("use informal German," "keep product names in English") in the same prompt, which is how Tidio's Lyro Guidance handles register; NMT needs a glossary feature for that. An LLM translator's output varies between runs, which Intercom's help center documents as expected ("these variations cannot be adjusted because they're inherent to how LLMs work"); NMT's does not. And an LLM can skip translation altogether: given a knowledge base in English and a question in Portuguese, it can answer in Portuguese in one step, which is not translation in the engineering sense and is the reason one of the four vendors below has no translation meter.

The four places translation runs inside a chatbot

Where the engine sits decides what can go wrong and who pays.

Insertion point	What is translated	When	Who sees the result	Typical failure
1. Content, at build time	Flows, buttons, FAQ answers, templates	Once, when you save	Every customer in that language, forever	A wrong translation is served to thousands before anyone reads it; sources drift from the copy
2. Inbound, at run time	The customer's message, into the bot's working language	Every message	The bot's NLU or retrieval	Meaning lost before understanding starts; short messages and mixed languages mistranslated
3. Outbound, at run time	The bot's reply, into the customer's language	Every reply	The customer	Register and terminology wrong on every turn; the answer was right, the wording foreign
4. Inbox, for the agent	The whole conversation, into the agent's language, and the agent's reply back	On open, and on send	The human agent and the customer	Agent and customer read different texts; what the agent saw is not what was sent

Intercom is the useful example because it runs three of the four as separately configured features and says so: "AI Inbox Translation (what teammates see), Fin's multilingual support (what customers receive from Fin), and Multilingual workflows (what customers see in automated workflow messages) are separate features. They must be configured independently, and disabling one has no effect on the others." Point 1 is its Multilingual workflows: switch Auto-translate on and "chat messages and reply buttons" are translated into every supported language in the background when you save; attribute-collector steps and Custom Answers are not. Point 3 is Fin's Real-time translation: "when Fin can't find relevant support content in the customer's language, it will translate existing support content in your chosen fallback language," with the note that "Fin translates its answer to the customer's language. The original source content that Fin references or pulls from will remain in its original language." Point 4 is AI Inbox Translation, with the caveat its Understanding translation behavior article states flatly: "What you see in the Inbox is not always what the customer received." Its multilingual setup article adds that English, French and German regional variants can be Fin and Help Center languages but not inbox-translation languages, so point 4 runs on the base language.

Tidio's Lyro is the counter-example. Its help center says Lyro "works natively in the customer's language instead of translating messages to English (internally) first," and that data sources "are processed directly in their original language as well, and do not require any translation to English first." That is a design choice to skip points 2 and 3 entirely by letting the model answer across languages, and the article lists 48 languages it will do that in.

What decides whether it works: three rules taken from the vendors' own notes

Translate the reply, not the source. Intercom's real-time translation and Lyro's native answering agree on this: the knowledge base stays in the language it was written in, and only the customer-facing text crosses the boundary. The alternative, machine-translating the whole knowledge base into six languages at build time, multiplies every drift and every error by six and gives retrieval six chances to pick the wrong passage. Build-time translation belongs to short, stable, high-traffic copy (buttons, greetings, menus), which is what Intercom's workflow feature limits itself to.

Decide the language once. Intercom's multilingual setup article says Fin "only detects the language once per conversation," and its detection article (13 May 2026) adds that it needs "at least 10 characters from customer messages to determine the language with confidence," and ignores button clicks. The full order is the customer's language override, then the AI detection, then the first browser language that is enabled in the workspace, then the workspace default. Translation that re-decides on every message flips languages on a borrowed word. The language detection entry covers why short first messages are the hard case; the point here is that the translation step inherits that decision and should not make its own.

Know which text is the original. Intercom's Understanding translation behavior article describes the failure exactly: a workflow written in Spanish while the workspace default is English is treated as English and the system "translates from that, producing incorrect output in all other languages," and a French workflow with Auto-translate left on is re-translated "treating it as the workspace Default language." Both are the same defect, an engine that does not know the source language, and both are fixed by configuration, not by a better model: match the default to the language you write in and re-save, or switch Auto-translate off on a flow already written in its delivery language. The same trap exists on Google's API, which detects the source for you at no extra charge when you omit it, and on any platform that assumes everything you type is in one language.

The meters, with the same 1,000 conversations on each

Assume 1,000 conversations in a month, each with eight customer messages of about 60 characters and eight bot replies of about 250 characters, which is 2,480 characters crossing the language boundary per conversation and 2.48 million in the month. The assumptions are ours and are stated so you can swap in your own transcript averages.

Meter	Vendor and product (read 5 September 2026)	What the 1,000-conversation month costs	What is and is not included
Per character, input only	Google Cloud Translation, NMT	(2.48M − 0.5M free) × $20 per million = $39.60	Language detection free when bundled with a translation call; whitespace and untranslated characters count; glossaries free
Per character, input and output	Google Cloud Translation, Translation LLM	2.48M × $10 in + ~2.48M × $10 out = ~$49.60; ~$39.60 if the up-to-$10 monthly credit, which the page lists on the NMT row and says applies to both editions, also covers LLM usage, which the page does not state	Output characters metered separately and "likely to be different" from input; Adaptive Translation at $25 + $25 would be ~$124
Folded into another meter	Intercom Fin, Real-time translation	$0 on a translation line; the answer itself is a Fin outcome at $0.99 (the meter our Intercom review records), priced whether or not it was translated	No separate charge for real-time translation or workflow Auto-translate appears in the four help articles read (Intercom's pricing page was not re-read for this entry); AI Inbox Translation free for 10 conversations per teammate per month, unlimited with a Copilot unlimited-usage seat
No meter, because no translation	Tidio Lyro	$0, and no line to look for; Lyro is metered by conversation allowance on Tidio's plans, in any of 48 languages	The model answers natively; "supported languages" can be restricted to those your agents speak; the default language is the fallback

DeepL is the fourth vendor and sits outside the table because its overage rate is quoted per plan on deepl.com/pro rather than as a single figure, so we show the plan structure rather than a fourth row: the API Growth plan includes one million characters on the monthly term, meters the remaining 1.48 million (if the plan is otherwise unused) by the source character at the rate on deepl.com/pro, caps the plan at 50 million characters a month, and counts source characters only, with HTML and XML tags excluded when tag_handling is set. Its API Developer plan is one million characters in total, not per month, and its help center says the older API Free and API Pro plans "can no longer be purchased."

The arithmetic is ours. The two lessons are that a per-character engine at this volume costs about what one helpdesk seat does (Intercom's Essential seat is $29 a month in our review), so the price of MT is rarely the reason to avoid it, and that the platforms have moved translation off the invoice entirely, either by bundling it into an answer fee or by not doing it. The cost that remains is the one no meter shows: a native speaker's afternoon per language checking what the engine did to your top flows and your twenty highest-traffic messages, which the 7 truths guide calls the cheapest localization test available and the localization guide schedules as a step.

Five rules before switching it on

Name the source language in the tool, not in your head. Every mistranslation case in Intercom's Understanding translation behavior article is a source-language mismatch. Set the workspace default, then write in it.
Keep the knowledge base monolingual; translate the reply. One source of truth, one place to fix an error, and retrieval that is not guessing between six copies.
Machine-translate only what a person will re-read. Buttons, greetings, menus and WhatsApp templates are short, stable and served thousands of times: translate them once by machine, then have a native speaker edit them once. Meta approves a template per language, so the translation is also a submission.
Do not translate the inbound message before understanding it if the platform can understand it directly. Point 2 in the table exists for platforms whose NLU is monolingual. If yours is an LLM that reads Portuguese, the translation step only adds a place to lose meaning.
Carry the language to the human. Inbox translation lets an English-speaking agent read Portuguese, and the warning in Intercom's article is the reason to be careful: the agent's screen and the customer's screen show different texts. A handoff to an agent who speaks the language beats a translated one when the conversation is about money.

What our fifteen reviews record about machine translation

None of them uses the phrase. Searched 5 September 2026, machine translation matches zero of the fifteen review files and auto-translate or real-time translation matches zero, while localization matches all fifteen, because our rubric's localization dimension scores per-language NLU, admin-interface languages and documentation rather than a translation feature, and multilingual matches seven. What the reviews do record is the outcome MT is supposed to produce, measured per language without a translation step: Tidio's review ran a 15-question fact-extraction set through Lyro in four languages against the same single-page index and recorded 87 percent (13/15) in English and Polish and 80 percent (12/15) in Mexican Spanish and Brazilian Portuguese, with the failures "clustering on idiomatic phrasings rather than semantic content," which is the shape you expect from a model answering natively rather than translating. Intercom's review scores multilingual NLU at 4 out of 5 on the strength of a 12-language interface recorded in that review, and notes that per-language benchmarks are not published; the Fin language list read for this entry has 65 language and locale entries. SendPulse ranks first on our multilingual list for a different reason, a five-language admin interface and support staff, which no translation engine substitutes for. No review measured a translation engine's output, and this entry does not claim one did.

Chatbot language detection — the step that decides which language the customer is writing in; translation inherits that decision.
Large language model — the engine that translates on request, or answers across languages without translating.
Natural language understanding — the layer that either reads the customer's language directly or reads a translation of it.
Chatbot knowledge base — the content that should stay in one language while replies cross the boundary.
Human handoff — where an inbox translation is a convenience and a same-language agent is the safer route.
WhatsApp message template — the one piece of bot copy that is translated once, submitted per language and approved by Meta.

FAQ

What is machine translation?

Software that converts text from one language to another without a human writing the target text. Production systems are neural: a dedicated neural machine translation model, or a large language model prompted to translate. In a chatbot it can be applied to the bot's content when it is built, to the customer's message as it arrives, to the reply as it leaves, or to the conversation as an agent reads it.

Is machine translation the same as a multilingual chatbot?

No. A multilingual chatbot is one that serves customers in several languages; machine translation is one way to get there. Tidio's Lyro is multilingual in 48 languages without translating anything, because its model answers in the customer's language. Intercom's Fin answers from content in the customer's language when it has some and translates its fallback-language content when it does not. The 7 truths guide covers the other layers, understanding, detection, templates and human agents, that translation does not touch.

How much does machine translation cost for a chatbot?

On a per-character API, less than you would guess: $40 to $50 for a 1,000-conversation month on Google Cloud's neural or LLM model at our assumptions of 2,480 translated characters per conversation. Intercom shows no separate charge for real-time translation in the help articles we read, only Fin's $0.99 per outcome from our Intercom review; Tidio has no translation charge because Lyro does not translate. The unpriced cost is native-speaker review of the copy the engine produced.

Should I translate my knowledge base or the bot's answers?

The answers. Both Intercom's real-time translation and Tidio's native answering leave the source content in its original language and put only the customer-facing text through the boundary. Translating the whole knowledge base multiplies drift and retrieval errors by the number of languages. Machine-translate short, stable copy (buttons, greetings, templates) at build time and have a native speaker edit it once.

Why does my chatbot translate the wrong way or into mixed languages?

Almost always because the engine does not know the source language. Intercom's Understanding translation behavior article gives both cases: content written in Spanish while the workspace default says English is treated as English and mistranslated everywhere; content already in French with auto-translate left on is re-translated as if it were the workspace default language. Set the default language to the language you write in, and turn auto-translation off on any flow already authored in its target language.

Sources

Google Cloud, Cloud Translation pricing (cloud.google.com/products/translate/pricing), read 5 September 2026: the Advanced (v3) table (NMT text translation "First 500,000 characters per month — Free (applied as $10 credit every month)," "Over 500,000 characters per month — $20.00"; custom models $80.00 to 250 million characters; Translation LLM $10.00 input and $10.00 output; Adaptive translation $25.00 and $25.00; document translation $0.08 per page NMT); the sentences "the number of input and output characters are likely to be different" and "it is charged $10 per million characters input and $10 per million characters output, making it cost equivalent with NMT"; the language-detection sentence ("there's no additional charge for the language detection in addition to the translation"); the custom-model training rate ("$45" per hour, "$300" cap). The Charged characters section: per code point, whitespace counted, and the footnote "Credit usage applies collectively to both Cloud Translation - Basic and Cloud Translation - Advanced. The credit is up to $10," which names the two editions and not the LLM method, hence the range in the body. The $39.60 and $49.60 figures are (2,480,000 − 500,000) × $20 per million and 2,480,000 × $10 × 2.
DeepL Help Center, DeepL API plans (support.deepl.com, article 360021200939) and Usage count and billing in DeepL API (article 360020685720), read 5 September 2026: "DeepL API Developer allows developers to translate up to 1,000,000 characters in total. Once this limit is reached, it is not reset"; "With a monthly billing term, the DeepL API Growth plan includes 1 million characters and 10 hours of speech to text"; "The DeepL API Growth plan has a usage limit of 50 million characters and 300 hours speech to text per month"; "The DeepL API Free plan can no longer be purchased" and "The DeepL API Pro plan can no longer be purchased"; "DeepL counts the characters in the source text"; the tag_handling exclusion. DeepL, Pricing (deepl.com/pro), opened the same day: the per-character overage rate is published there and is not quoted in this entry.
Intercom Help, Set up Fin AI Agent's multilingual support (article 8322387, written by Beth-Ann Sher), read 5 September 2026: the supported-languages list (65 entries as counted by us, including regional variants); the three-step setup; "By enabling real-time translation, Fin can work in all your supported languages without needing content in each language"; "Fin will first search for the answer using content that is available in the language a question was written in"; "Fin translates its answer to the customer's language. The original source content that Fin references or pulls from will remain in its original language"; "Fin only detects the language once per conversation"; the note that English, French and German regional variants "are not available as Workspace languages or AI Inbox translation languages."
Intercom Help, Automatic language detection in conversations (article 9423767, written by Paula Lopez, dated 13 May 2026), read 5 September 2026: the 10-character minimum, the once-per-conversation rule, the four-step order (language override, AI detection, first enabled browser language from the Accept-Language list, workspace default), and the note that button clicks and attribute collectors are ignored.
Intercom Help, Multilingual workflows (article 10522253, written by Beth-Ann Sher, dated 30 July 2026), read 5 September 2026: the Auto-translate toggle; "Currently supported: Chat messages and reply buttons. Translations are available on all channels"; attribute collector and Custom Answers not supported; the instruction to turn Auto-translate off on a workflow "already authored in its intended delivery language."
Intercom Help, Understanding translation behavior in Intercom (article 16296591, written by Beth-Ann Sher, "Updated over 3 weeks ago" on the read date), read 5 September 2026: the three-feature table and the independence note quoted in the body; "What you see in the Inbox is not always what the customer received"; the two mistranslation cases quoted in the body ("translates from that, producing incorrect output in all other languages"; "treating it as the workspace Default language"); "Each teammate gets 10 free translated conversations per month" and the Copilot seat; "Intercom uses AI and large language model (LLM) providers for translations" and "These variations cannot be adjusted because they're inherent to how LLMs work."
Chatbotscape, Intercom review (/reviews/intercom-review): the $0.99 per Fin outcome meter and the 12-language interface count cited in the body.
Tidio Help Center, Lyro Multilanguage support (help.tidio.com, article 12643043611036, dated 31 October 2025), read 5 September 2026: "There are currently forty-eight (48) supported languages" and the list; "Lyro works natively in the customer's language instead of translating messages to English (internally) first"; the data-sources sentence; the default-language fallback and Specific languages option; the Lyro Guidance formal/informal German example.
Kocmi et al., Findings of the WMT24 General Machine Translation Shared Task: The LLM Era Is Here but MT Is Not Solved Yet (Proceedings of the Ninth Conference on Machine Translation, November 2024, pages 1–46, aclanthology.org/2024.wmt-1.1), abstract read 5 September 2026: 11 language pairs, "translations from 8 different large language models (LLMs) and 4 online translation providers," professional annotators and the Error Span Annotations protocol. The body of the paper was not re-read for this entry; only the title and abstract are cited.
Chatbotscape review corpus (the fifteen platform reviews listed at /reviews), searched 5 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. grep -liE 'machine translation|machine-translat' sample-reviews/*-review.md returns 0; grep -liE 'auto-translat|auto translat|real-time translation|realtime translation' sample-reviews/*-review.md returns 0; grep -liE 'localization|localisation' sample-reviews/*-review.md returns 15; grep -liE 'multilingual' sample-reviews/*-review.md returns 7. Passages cited: tidio-review.md line 692 (per-language results and the "idiomatic phrasings" sentence); intercom-review.md lines 69–70 and 228 (12-language UI, multilingual NLU 4/5, "per-language benchmarks not published").
Ahrefs Keywords Explorer, US overview, queried 5 September 2026 — the demand, difficulty, CPC, global-volume and parent-topic figures in this entry's keyword note.
Chatbotscape evaluation methodology. /methodology (continuously updated).