# Multilingual customer support: AI coverage and limits > Multilingual customer support with AI: how it extends language coverage, where quality breaks down, and when to keep native-speaker humans. - **Published:** July 21, 2026 - **Category:** Support - **Author:** Udit Goenka - **URL:** https://communicate.so/blog/multilingual-customer-support --- > **TL;DR:** Multilingual customer support is answering your customers in their own language instead of forcing everyone through one. For most of history that meant hiring native speakers for every market, which is why only large companies could afford real coverage. An AI agent grounded in your own content changes the economics, because the same model can read a question in one language and answer in another without a separate team behind each one. This guide is the honest explainer: what multilingual support means now, how AI extends language coverage, where the quality genuinely holds up, and where it breaks down on idiom, tone, and low-resource languages that the model has barely seen. It covers when you should keep native-speaker humans on a language, how escalation works when the customer and the agent do not share a first language, and how to roll all of this out without promising coverage you cannot deliver. The short version is that AI widens the door to many more languages than you could ever staff, but the width of that door varies sharply by language, and treating every language as equally solved is the fastest way to embarrass yourself in a market you do not understand. --- Multilingual customer support used to be a hiring problem before it was a software problem. If you wanted to answer customers in Portuguese, Japanese, and Arabic, you hired people who spoke Portuguese, Japanese, and Arabic, and if you could not afford that, those customers got English and a shrug. The cost of coverage scaled with the number of languages, so coverage stayed narrow for almost everyone. An [AI agent](/ai-agents) grounded in your own content rewrites that math. One model can detect the language a customer wrote in, retrieve the relevant answer from your knowledge base, and reply in that same language, without a dedicated team standing behind each one. That does not make language free, but it moves the bottleneck from headcount to quality, which is a very different and more manageable problem. This guide is written for the person deciding how far to stretch their languages: a founder selling into new markets, a support lead staring at tickets in languages nobody on the team reads, or an ops owner weighing [customer support automation](/blog/customer-support-automation) against the risk of getting a translation subtly wrong. It is honest about both sides. AI extends your reach dramatically, and it also fails in ways that are easy to miss until a customer points them out, so the goal is to use the reach without trusting it blindly. ## What multilingual customer support means today Multilingual customer support is the practice of helping customers in the language they choose, rather than the language your team happens to speak. At its simplest it is a translated help center. At its most capable it is a live agent that converses fluently across many languages from a single knowledge base, and everything in between is a point on that range. **The useful distinction is between translated content and live conversation.** Translating your documentation once is a static, one-time act you can review carefully before it ships. Answering a live question in another language is dynamic, unrehearsed, and impossible to fully proofread in advance, which is why the live case is where AI both shines and stumbles. There is a second distinction worth naming early, between translation and native answering. A system can take an English answer and translate it into the customer's language, or it can generate the answer directly in that language. Modern models mostly do the latter, and the quality gap between the two approaches is one of the things that makes multilingual quality so uneven across languages. Language is not just words, it is expectation. A German customer often expects formal address and precision, a Brazilian customer may expect warmth, and a Japanese customer may expect a level of politeness that has grammatical consequences. Groups like [CSA Research](https://csa-research.com) have documented for years that customers strongly prefer buying and getting support in their own language, so getting this right is a revenue question, not just a courtesy. It also matters that multilingual support is not one feature but a chain of them. Detecting the language, retrieving the right content, generating a fluent reply, and escalating cleanly when the model is unsure are separate steps, and each one can succeed or fail on its own. A system can detect the language perfectly and still answer it badly, which is why you evaluate the whole chain rather than any single link. The rest of this guide treats multilingual support as that full chain, from detection to escalation, and it is opinionated about where humans still belong. If you want the broader context of what an always-on agent does across time zones as well as languages, the [24/7 AI support](/blog/24-7-customer-support-ai) guide is a useful companion, because global coverage is a language problem and a clock problem at the same time. ## How AI extends language coverage ![Line-art diagram showing one AI agent and one knowledge base fanning out into replies across many languages](https://communicate.so/blog/multilingual-customer-support-coverage.png) **The core shift is that language coverage decouples from headcount.** With human-only support, adding a language means adding people who speak it, and each new market is a hiring cycle. With a grounded AI agent, adding a language is closer to a configuration change, because the model already carries broad multilingual ability and your content is the same underneath. Detection is the first thing AI buys you. A customer writes in Spanish, the system recognizes the language, and it replies in Spanish without the customer selecting anything from a menu. That automatic matching removes a small but real friction, because a customer who has to hunt for a language switcher often just gives up and writes in broken English instead. Grounding is what keeps multilingual answers honest. Instead of translating from general knowledge, the agent retrieves the relevant passage from your connected [data sources](/data-sources) and answers from it in the customer's language. That means the same facts drive the Spanish answer and the English one, so you are not maintaining a separate, drifting truth for every market you serve. The reach here is genuinely wide because modern models were trained on enormous multilingual corpora. Research efforts like Meta AI's No Language Left Behind have pushed machine translation toward hundreds of languages, and general-purpose models inherit a broad, if uneven, version of that ability. The result is that an agent can attempt far more languages than any support team you could realistically hire. This is also where the widely quoted automation ceiling lives. Gartner has projected that by 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention ([Gartner](https://www.gartner.com)). That projection is not language-specific, but it depends on the same repetitive, documented questions that translate well, so the multilingual version of it is realistic for your strongest languages and shakier for your weakest. The practical payoff is that a small team can suddenly serve customers it could never have staffed for. A three-person support team in one country can answer the overnight German, Korean, and French questions from the same knowledge base, at least for the documented majority. That is a real expansion of who you can serve, and it is the honest case for using AI on language at all. ## Where AI-driven multilingual support holds up **AI multilingual support is strong exactly where it is grounded, repetitive, and factual.** A documented how-to, an account status lookup, a policy explanation, or a setup question translates cleanly because the answer is the same fact in a different language. When the underlying answer is unambiguous, the language layer has an easy job, and quality stays high across most major languages. High-resource languages are where you get near-parity with English. Spanish, French, German, Portuguese, Italian, and the other widely spoken European languages, along with major Asian languages, have so much training data behind them that fluent, accurate replies are the norm for standard support questions. For these, a grounded [AI agent](/ai-agents) often needs no special handling beyond the same guardrails you would use in English. Speed is a quiet advantage that applies across every language equally. An AI agent answers a Thai question at the same instant it answers an English one, with no queue behind the less-common language. Under a human-only model, minority-language customers usually wait longest, because there are fewer agents for them, so AI actually flattens a fairness problem most teams never measured. The table below maps where AI multilingual support tends to hold up versus where you should keep a human in the loop. Read it as a starting bias you refine with evidence from your own analytics, not a fixed ruling, because your language mix and your content quality both move the lines. | Support scenario | AI handles it well | Keep a human | | --- | --- | --- | | Documented FAQ in a high-resource language | ✓ | ✗ | | Account or order status lookup, any major language | ✓ | ✗ | | How-to and setup steps, high-resource language | ✓ | ✗ | | Language detection and routing | ✓ | ✗ | | Idiom-heavy or slang-heavy message | ✗ | ✓ | | Low-resource or minority language, high stakes | ✗ | ✓ | | Legal, medical, or financial wording | ✗ | ✓ | | Emotional or cultural sensitivity in reply tone | ✗ | ✓ | Read the table as a bias, not an absolute: the left column is where automation earns its keep, and the right column is where a person protects the relationship. The pattern is the same one that governs automation generally, sharpened by the fact that language adds its own failure surface on top of the usual ones. Consistency is the last strength worth naming. A grounded agent gives the same correct policy answer in every language, while a distributed team of contractors across markets tends to drift, each interpreting the policy slightly differently. Centralizing the source of truth in your [data sources](/data-sources) and letting one agent express it many ways is how you keep a global answer from fragmenting into ten local versions. ## Where multilingual quality breaks down ![Line-art illustration of multilingual failure modes: a mistranslated idiom, a tone mismatch, and a thinly covered low-resource language](https://communicate.so/blog/multilingual-customer-support-quality-limits.png) **Multilingual AI fails in ways that are invisible to the team that cannot read the language.** This is the core danger. An English-speaking founder can spot a wrong English answer instantly, but a subtly wrong Korean answer sails past unnoticed until a Korean customer complains, which means the failures accumulate silently in exactly the languages you can least audit. The first failure is idiom and slang. Customers do not write in the clean, textbook language that models translate best, they write with regional expressions, abbreviations, and inside jokes. A phrase that means one thing literally and another thing colloquially can flip an answer's meaning, and the model often takes the literal reading because it has no context for the local usage. The second failure is tone and register. Many languages encode formality grammatically, so choosing the wrong level of politeness is not a style choice, it is a small insult. A model that addresses a formal Japanese customer too casually, or a German customer with the familiar form, sounds subtly wrong in a way that erodes trust, and this is precisely the kind of error a [shared inbox](/shared-inbox) with a native speaker catches and a monolingual dashboard never will. The third and largest failure is low-resource languages. The model has seen enormous amounts of Spanish and very little of many smaller languages, so its fluency drops off steeply as you move away from the well-represented ones. For a low-resource language, the agent may produce text that looks confident and reads as broken or wrong to a native speaker, which is worse than an honest fallback. This is not a fringe risk, it is the general pattern of AI projects that overreach. RAND's 2025 review of more than 2,400 enterprise AI initiatives found roughly 80% failed to deliver measurable value, mostly on operational discipline rather than raw model quality ([RAND](https://www.rand.org)). In a multilingual context, that discipline means knowing which languages you have actually validated and refusing to fake the ones you have not. There is also a quieter class of failure around formatting and locale. Dates, currencies, number formats, addresses, and units differ by region, and a fluent reply that quotes the wrong date format or currency symbol still misleads the customer. Language correctness and locale correctness are separate problems, and a system can pass one while failing the other. Standards bodies exist precisely because this is hard. The Unicode CLDR project maintains locale data for how dates, numbers, and names are formatted around the world ([Unicode CLDR](https://cldr.unicode.org)), and the fact that it takes a dedicated standard to track this should tell you that a general model will not always get it right on its own. Treat locale formatting as something to verify, not assume. ## When to keep native-speaker humans ![Line-art diagram showing high-stakes and low-resource-language conversations routing to a native-speaker human while routine ones stay automated](https://communicate.so/blog/multilingual-customer-support-when-humans.png) **The decision to keep a native speaker on a language is about stakes and confidence, not pride.** You keep a human where a wrong answer is expensive and where you cannot verify the AI's quality yourself. Those two conditions, high stakes and low visibility, are exactly the ones that make unattended multilingual automation risky. Keep native speakers on your highest-value markets during the conversations that matter most. A routine password reset in French can run on the agent, but a French enterprise customer negotiating a contract, disputing a charge, or threatening to churn deserves a person who reads the room in their own language. The rule mirrors general escalation: automate the documented majority, keep the judgment calls human. Keep humans on low-resource languages until you have evidence the agent performs. If you cannot audit the quality and the training data behind the language is thin, the safe default is to let the AI assist a human rather than answer alone. A native speaker reviewing AI-drafted replies inside a [shared inbox](/shared-inbox) gives you the speed of automation with a human backstop on the fluency you cannot check. Cultural sensitivity is its own reason to keep a person. Complaints, apologies, bereavement, and anything touching local law or custom carry expectations that vary by culture, not just by language. A literally correct translation can still land as cold or tone-deaf, and only someone inside that culture reliably catches the difference between correct and appropriate. There is a measurement reason too. You cannot improve what you cannot read, so early on you need native speakers not just to answer but to grade the agent's output and feed corrections back into your [data sources](/data-sources). Treat your first native-speaker hires in a language partly as evaluators, because their judgment is how you learn whether the AI is safe to widen there. The honest framing is that AI does not remove the need for native speakers, it changes what they do. Instead of typing every reply, they supervise the languages you trust least, handle the high-stakes conversations, and teach the system by correcting it. That is a smaller, higher-leverage role than staffing every language end to end, which is the real efficiency multilingual AI delivers. ## Escalation across languages ![Line-art diagram of a multilingual conversation escalating from an AI agent to a human, preserving language and full context](https://communicate.so/blog/multilingual-customer-support-escalation.png) **Escalation is the hardest part of multilingual support, because the handoff can lose the language.** An AI agent might handle a conversation in Italian and then need a human, but the human on shift may not read Italian. That mismatch is unique to multilingual support, and it is where a clumsy handoff does the most damage. The first requirement is that context survives the handoff. When the agent escalates, the human must inherit the full conversation, not a cold restart, because asking a customer to repeat themselves is already a top frustration, and asking them to repeat themselves in a second language compounds it. The whole thread, in the original language, has to travel with the escalation. This is why the handoff mechanism matters more than the model. In Communicate, the [Shared Inbox](/shared-inbox) uses presence-based human takeover, so when a person opens a conversation the AI steps back automatically, and a per-turn backstop stops the AI from talking over a human mid-reply. The [AI to human handoff](/blog/ai-human-handoff-support) guide covers this mechanism in depth, and in a multilingual setting it is what keeps two responders from colliding in a language one of them cannot read. The second requirement is honest routing when no native speaker is available. If a customer writes in a language nobody on shift reads, the honest move is to tell them, offer an asynchronous reply from someone who does, or let the AI assist under supervision, rather than pretending the coverage exists. A transparent delay beats a confident wrong answer in a language you cannot check. Repetition is the friction to design against, and the data is blunt about it. Zendesk's 2024 CX Trends research found 74% of customers rank having to repeat themselves among their biggest annoyances ([Zendesk](https://www.zendesk.com/blog/customer-service-statistics/)). Across a language boundary that annoyance is amplified, so the escalation has to carry everything the AI already learned, in the language it was learned in. The last piece is deciding when to escalate at all. The agent should hand off when its confidence is low, and confidence should be treated more conservatively in weaker languages. A model that would answer an English edge case might reasonably escalate the same edge case in a low-resource language, because the cost of a fluent-looking mistake is higher exactly where you can verify it least. ## How to roll out multilingual support without overpromising Rolling out multilingual support well is mostly about sequence and honesty, not about flipping every language on at once. The failure pattern is to advertise support in twenty languages on day one and quietly deliver good answers in three. The steps below keep your promises matched to your actual coverage. **Start from your traffic, not from a wish list.** Pull the languages your customers actually write in and rank them by volume and value, because there is no point perfecting a language nobody contacts you in. Your top two or three non-English languages are almost always where the return is, and that is where you focus first. Get your content in order before you translate anything, because grounding decides multilingual quality. Connect clean, current [data sources](/data-sources) and fix contradictions first, since a grounded agent will confidently repeat whatever wrong thing it finds, in every language. The companion guide on how to [train AI on your help center](/blog/train-ai-on-help-center) covers this groundwork, and it matters more across languages, not less. Validate each language with a native speaker before you advertise it. Assemble real, messy tickets in that language, add idiomatic and adversarial ones, and have someone who speaks it grade the agent on correctness, tone, and locale formatting. Set a go/no-go bar per language, and route anything below it into a [shared inbox](/shared-inbox) for human handling until the evidence improves. Advertise only what you have validated. If you have proven Spanish, French, and German, say you support Spanish, French, and German, and let the agent attempt the rest with a clear, honest fallback and a fast escalation. Under-promising and over-delivering builds trust in a new market, while the reverse teaches a market to distrust you in a language you barely serve. Then watch the numbers per language, not blended. Point your analytics at resolution rate, escalation rate, and re-contact rate broken out by language, because a blended average will hide a language that is quietly failing. Widen coverage one language at a time as the evidence earns it, and be willing to pull a language back if its numbers turn out worse than the dashboard first suggested. ## Where Communicate fits, honestly Communicate is a grounded AI support agent that answers in the customer's language and hands off when it is unsure, built for the automate-the-repetitive, keep-the-judgment split this guide describes. The agent detects the language, retrieves from your connected [data sources](/data-sources), and replies in that language, while the [Shared Inbox](/shared-inbox) uses presence-based human takeover with a per-turn backstop so a native speaker can step in without the AI talking over them. Here is what it does without embellishment. The live channels are a web widget, live chat, and email, with in-app messages, analytics, and scoped actions running from the same agent and knowledge base, so behavior stays consistent across every surface and every language. There is no WhatsApp, Messenger, SMS, or voice, so if any of those is a hard requirement it is not your best fit today. On the model, Communicate runs a single model, gpt-4o-mini through [OpenRouter](https://openrouter.ai), with response and prompt caching to keep cost and latency down. That model handles many languages, but quality varies by language, and this guide is deliberate about that: it is strongest on high-resource languages and weaker on low-resource ones, so validate the languages you care about rather than assuming uniform coverage. We do not claim a fixed number of supported languages, and we do not staff a bench of human translators. On pricing, there is no free tier. Entry is a one-time $1 activation that confirms you are a real person and includes 100 test credits, then credit-based usage from there, which keeps support cost tracking usage rather than headcount. Spend those test credits on the per-language validation this guide describes, using your ugliest real messages in each language, before you trust the agent live in that market. Now the honest limits. Communicate is GDPR-ready but not certified, holds no SOC 2, HIPAA, or ISO 27001, runs in a single region, and does not offer SSO. It supports TOTP two-factor authentication, encryption at rest, and workspace isolation, a posture stated plainly on the product. Questions go to communicate@support.communicate.so. ## Key takeaways - Multilingual customer support used to scale with headcount, but a grounded AI agent decouples language coverage from hiring, so a small team can serve many more languages than it could ever staff. - AI holds up best on documented, repetitive questions in high-resource languages, where the answer is a fact expressed in another language, and it flattens the wait that minority-language customers usually suffer. - Quality breaks down on idiom and slang, on tone and formality that some languages encode grammatically, and most of all on low-resource languages the model has barely seen, and these failures are invisible to a team that cannot read the language. - Keep native-speaker humans on high-stakes conversations, on low-resource languages until they are proven, and as early evaluators who grade and correct the agent so you can improve what you cannot personally read. - Roll out from your real traffic, validate each language with a native speaker against a go/no-go bar, advertise only what you have proven, and watch resolution and re-contact rates per language rather than blended. Ready to reach customers in their own language without overpromising coverage? [Start with a one-dollar account activation](/ai-agents) that includes 100 test credits, connect your [data sources](/data-sources), and validate your top languages with a native speaker before you go live. When you are ready to prepare the content that grounds every language, the guide on how to [train AI on your help center](/blog/train-ai-on-help-center) is the right next read. ## Frequently asked questions ### What is multilingual customer support? Multilingual customer support is helping customers in the language they choose rather than the one your team speaks, across your help content and your live conversations. It ranges from a translated help center to a live [AI agent](/ai-agents) that converses fluently across many languages from one knowledge base. The goal is to meet customers in their own language, which research consistently ties to higher satisfaction and conversion. ### Can AI handle customer support in multiple languages? Yes, and this is one of the clearest wins of a grounded agent. A single model can detect the customer's language, retrieve the answer from your connected data sources, and reply in that language without a separate team behind each one. The reach is wide, but quality varies by language, so treat it as a capability to validate rather than assume. ### How many languages can AI customer support cover? More than you could staff, but the honest answer is that coverage is uneven rather than a single number. Modern models carry broad multilingual ability from their training data, so they can attempt many languages, yet fluency drops off sharply for less-represented ones. Rather than claiming a fixed count, validate the specific languages your customers actually use and advertise only those. ### Does AI translate or answer natively in each language? Modern models mostly generate the answer directly in the target language rather than translating a finished English reply, and the two approaches produce noticeably different quality. Native generation tends to read more naturally, while literal translation can carry English phrasing into a language where it sounds off. Either way, the underlying facts should come from your grounded content, not the model's general knowledge. ### Which languages does AI support best? High-resource languages with large amounts of training data, such as Spanish, French, German, Portuguese, Italian, and the major Asian languages, get close to English-level quality on standard support questions. Less common and low-resource languages are weaker and riskier, because the model has seen far less of them. The practical rule is to trust the well-represented languages and supervise the rest. ### What are low-resource languages in customer support? Low-resource languages are those with relatively little training data behind them, so a model's fluency and accuracy in them drop off compared to widely spoken languages. In support, the danger is that the agent may produce confident-looking text that a native speaker reads as broken or wrong. For these, keep a human in the loop until you have real evidence the agent performs. ### Why does AI struggle with idioms and slang? Customers write with regional expressions, abbreviations, and colloquialisms rather than textbook language, and a model often takes the literal reading because it lacks the local context. A phrase that means one thing literally and another thing in everyday use can flip an answer's meaning. This is why idiom-heavy or slang-heavy messages are a common place to route to a person who understands the local usage. ### Can AI match the right tone and formality in each language? Sometimes, but tone is a frequent failure point because many languages encode formality grammatically. Addressing a formal customer too casually, or using the familiar form where the polite one is expected, reads as a small insult rather than a style slip. A native speaker reviewing replies in a shared inbox catches this, while a monolingual dashboard never will. ### When should I use native-speaker human agents? Keep native speakers on high-stakes conversations, on your highest-value markets, and on low-resource languages until the agent is proven there. A routine question can run on automation, but disputes, cancellations, legal wording, and culturally sensitive moments deserve a person who reads the room in that language. Early on, native speakers also act as evaluators, grading the agent and feeding corrections into your knowledge base. ### How does escalation work across languages? The agent hands off when its confidence is low, and confidence should be treated more conservatively in weaker languages. The critical requirement is that the full conversation travels with the escalation, in the original language, so the customer never repeats themselves. Communicate handles this through presence-based takeover in the [Shared Inbox](/shared-inbox), and the [AI to human handoff](/blog/ai-human-handoff-support) guide covers the mechanism in detail. ### Does multilingual support need separate knowledge bases per language? Usually not, and a single grounded source is often better. When one agent retrieves from one set of data sources and expresses the same facts in many languages, you avoid maintaining parallel, drifting versions of the truth. You may still add language-specific notes for locale details, but the core answer should stay centralized so a global policy does not fragment into ten local ones. ### Can one AI agent detect the customer language automatically? Yes, automatic language detection is one of the first things AI buys you. The customer writes in their language, the system recognizes it, and it replies in kind without the customer selecting anything from a menu. That removes a small friction that otherwise pushes customers to write in broken English, and it makes the whole experience feel native rather than bolted on. ### How do I measure multilingual support quality? Break your metrics out by language rather than blending them, because an average hides a failing language. Watch resolution rate, escalation rate, and re-contact rate per language in your analytics, and have native speakers grade a sample of real replies for correctness, tone, and locale formatting. The blended number will look fine long after a specific language has quietly started failing. ### Is machine translation good enough for customer support? For documented, factual questions in well-represented languages, it is often good enough, especially when the answer is grounded in your own content. For idiom, tone, low-resource languages, and high-stakes wording, it is not reliable on its own and needs a human check. The honest position is that translation quality is uneven, so you validate per use case rather than trusting it uniformly. ### How does multilingual support affect response time? It usually improves it, and it improves it most for minority-language customers. An AI agent answers a less-common language at the same instant as a common one, with no thin queue behind it, so the customers who waited longest under human-only support now get an immediate first response. A [24/7 AI support](/blog/24-7-customer-support-ai) layer extends that speed across time zones as well as languages. ### Does multilingual AI support work 24/7? Yes, and global support is a clock problem and a language problem at once. An always-on agent resolves documented questions in many languages around the clock and escalates the rest, so overnight customers in every market get an answer rather than silence. The [24/7 AI support](/blog/24-7-customer-support-ai) guide covers how always-on coverage works and where it still needs a human backstop. ### What channels support multilingual customer support in Communicate? Communicate handles multilingual conversations across a web widget, live chat, and email, plus in-app messages, all from one agent and one knowledge base, so answers stay consistent across surfaces and languages. It does not cover WhatsApp, Messenger, SMS, or voice. If a messaging app or phone line is a hard requirement for your markets, confirm channel support before committing, because those channels are not available today. ### Is multilingual customer support GDPR compliant? Communicate is GDPR-ready but not certified, and language does not change that posture. It supports encryption at rest, TOTP two-factor authentication, and workspace isolation, runs in a single region, and holds no SOC 2, HIPAA, or ISO 27001, nor does it offer SSO. If a specific compliance certification is a requirement for a market you serve, treat that as a gating question and reach out to communicate@support.communicate.so before you rely on it. ### How much does multilingual AI customer support cost? With Communicate there is no free tier, and language coverage does not carry a separate price. Entry is a one-time $1 activation that includes 100 test credits, then credit-based usage from there, which tracks usage rather than headcount. That matters for multilingual support specifically, because you are covering many languages without paying for a team in each one, so cost follows volume rather than the number of markets. ### How do I start offering multilingual support safely? Start from the languages your customers actually write in, clean and connect your [data sources](/data-sources), and validate each language with a native speaker against a go/no-go bar before you advertise it. Route anything below the bar into a [shared inbox](/shared-inbox) for human handling, advertise only the languages you have proven, and widen coverage one language at a time as your per-language analytics earn it.