It's 10:47 p.m. A customer opens your chat widget and types, in Spanish, that her deposit went through twice and she needs one charge reversed before her card statement closes. Your bot answers in fluent Spanish within two seconds. It sounds warm. It sounds competent. It also tells her, because one phrase was mistranslated, that the duplicate charge "will be kept" instead of "will be held for review."
That is the real problem with bilingual AI customer service. Most machine translation is good. The errors that slip through, though, are delivered with the same confidence as everything else, and the customer has no way to spot them. This guide is for founders, operators, and technology teams deciding where Spanish–English AI translation belongs in their customer communication, where it doesn't, and how to tell the difference before a customer finds out for you.

Why this is not a niche problem
Nationally, more than one in five U.S. residents age 5 and older spoke a language other than English at home between 2017 and 2021, and about 61% of them spoke Spanish, according to the U.S. Census Bureau's American Community Survey detailed language tables. The same release notes that 62% of people who speak another language at home also speak English "very well." That number matters: a large share of your Spanish-speaking customers can read your English. They choose Spanish because it is easier, faster, or more comfortable. A clumsy translation does not just confuse them. It tells them you didn't think it was worth doing properly.
In South Florida the numbers look different again. ACS 2020–2024 estimates put the share of Miami-Dade County residents age 5+ who speak a language other than English at home at 75.3%, compared with 30.6% statewide (Census QuickFacts). For many local businesses, Spanish is not an extra language. It is a primary channel.
What the research says about AI translation accuracy
The most useful evidence for operators comes from fields where a translation error has consequences. In medicine, researchers have tested machine translation on real patient instructions and counted the errors that could cause harm.
- Spanish does well, but not perfectly. A UCSF study in JAMA Internal Medicine ran 647 sentences from 100 sets of emergency department discharge instructions through Google Translate. It found 92% of sentences translated accurately into Spanish and 81% into Chinese (Khoong et al., 2019). One cited example: "hold the kidney medicine" came out in Spanish as "keep the medication" (Physician's Weekly interview with the lead author). That is the same kind of error as the deposit example above: one verb, opposite meaning.
- A small error rate still includes serious errors. A follow-up study summarizing that work reports that inaccurate translations with potential for significant harm made up about 2% of Spanish sentences and 8% of Chinese sentences (Taira et al., Journal of General Internal Medicine, 2021).
- Language pair changes everything. In that same study's test of simple discharge statements, accuracy ranged from 94% for Spanish down to 55% for Armenian, with Tagalog, Korean, Chinese, and Farsi in between (Taira et al.). A vendor's claim about "accuracy" means very little until you know which language pair and which content it was measured on.
- Writing style drives errors. The UCSF team found most errors appeared when the English source was long and full of jargon (ScienceDaily summary of the study). This is the most controllable variable you have.
Machine translation research outside medicine points the same way. A large study in Transactions of the Association for Computational Linguistics had professional translators annotate errors with full document context using the Multidimensional Quality Metrics (MQM) framework. It found that crowd-worker ratings correlated poorly with expert error annotation, and that professional human translation still clearly outperformed top machine systems (Freitag et al., 2021). The practical takeaway: if your QA process is "a bilingual employee skimmed it and it looked fine," you are measuring the wrong thing.
One caveat. These studies tested engines as they existed several years ago, and translation quality, especially with large language models, keeps improving. That makes the method more useful than the specific percentages: test your own content, in your own language pairs, with qualified reviewers, and test again when you switch tools.
Where machine translation helps
Machine translation earns its place where messages are frequent, predictable, and easy to fix if something goes slightly wrong.
- First response and triage. Acknowledging a message, collecting an order number, or routing a question to the right queue in the customer's language beats making them wait until morning.
- Low-stakes FAQs. Hours, locations, parking, general turnaround times. Errors here are annoying, not damaging.
- Understanding inbound messages. Translating a Spanish email into English so an English-speaking agent can grasp the issue is a lower-risk use than sending machine output to the customer. The agent can still ask a bilingual colleague before replying.
- Drafting, not sending. A machine first draft that a bilingual staff member edits is often faster than writing from scratch, and the human stays accountable for what goes out.
Where machine translation hurts
It hurts where a single word changes money, commitments, safety, or trust.
- Money and policy. Refunds, deposits, cancellation terms, late fees, warranty conditions. "Held," "kept," "refundable," and "credit" are not interchangeable, and a customer will hold you to the word you sent.
- Anything regulated or safety-related. Health, allergens, legal notices, insurance, contracts. The clinical studies above exist precisely because these errors matter.
- Complaints and emotional conversations. An upset customer reads tone closely. Spanish forces a choice between tú and usted that English never makes. A bot that drops into casual tú with a furious long-time customer can make things worse before any facts are discussed.
- Your own vocabulary. Product names, plan tiers, internal service names, and local terms tend to get "translated" into something generic or wrong.
- Real South Florida Spanish. Customers write in Cuban, Venezuelan, Colombian, Puerto Rican, and other varieties, often mixed with English mid-sentence. Code-switched messages and regional slang are where a system is most likely to misread the input before it even attempts a reply.
A practical decision framework
Instead of asking "should we use AI translation?", sort every message type your team sends into three lanes.
- Green: machine translation can send alone. Greetings, confirmations of receipt, hours, directions, generic status updates. Keep these on approved templates where possible.
- Yellow: machine drafts, a human approves. Quotes, scheduling changes, account-specific answers, anything that mentions a price or a date. A bilingual reviewer checks meaning, not just fluency.
- Red: human-written or professionally translated only. Contracts, refund and cancellation policies, legal and health-related content, escalated complaints. Translate these once, properly, and reuse the approved versions.
Then decide two things explicitly: what triggers a handoff from bot to human (keywords like refund, cancel, charge, lawyer, allergy; low confidence; a customer asking for a person), and who on your team is responsible for Spanish quality. If nobody owns it, nobody fixes it.
Make the machine better before you trust it
Write English that translates well
Given the UCSF finding about long, jargon-heavy sentences, the cheapest accuracy gain is in your source text. Use short sentences. One instruction per sentence. Avoid idioms ("we'll circle back," "you're all set") and ambiguous verbs ("hold," "keep," "charge"). This also makes your English clearer for everyone.
Use a glossary for your terms
Major translation APIs let you lock down terminology. Google describes its Cloud Translation glossary as a custom dictionary for translating domain-specific terms consistently; entries are single words or short phrases, usually under five words (Google Cloud Translation documentation). Google's tutorial also notes that the API is case-sensitive and sensitive to accents, so you need to list spelling variants explicitly (Google Cloud glossary tutorial). DeepL supports glossaries through a glossary_id parameter on translation requests (DeepL API reference). Start with your product names, plan names, policy terms, and any word that has burned you before.
Control formality on purpose
For Spanish, tú versus usted is a brand decision, not a translation detail. DeepL exposes a formality setting, but support depends on the target language, which you can check through its languages endpoint (DeepL Languages API documentation). Its official client libraries describe "prefer" options that fall back to the default when a language doesn't support formality control (DeepL .NET library). If you're using an LLM-based chatbot instead of a translation API, put the register rule in the system instructions and test it, because it can drift over a long conversation.
Test like an expert, not like a crowd
Following the MQM approach from Freitag et al., have a qualified bilingual reviewer read a sample of real conversations in full context and log each error by type (mistranslation, terminology, omission, register) and severity (minor, major, critical). Weight your sample toward the yellow and red lanes. Ask the question that matters: would a customer act differently because of this error? Re-run the review whenever you change engines, prompts, or glossaries.
Let customers flag problems
Offer a visible "talk to a person" option in both languages and a quick way to say a reply didn't make sense. Feed those flags back into your glossary and your lane assignments.
Choosing a multilingual chatbot or translation tool
When you evaluate vendors, skip the headline accuracy claim and ask:
- Which language pairs and which kinds of content was quality measured on?
- Can we enforce a glossary and a formality level?
- Can we set hard handoff rules for specific topics, regardless of model confidence?
- Can we see and export the original message, the translation, and the reply for every conversation, so we can audit them?
- Where does customer data go when it's sent for translation, and is it retained?
Frequently asked questions
Is AI translation accurate enough for Spanish customer service?
For routine, predictable messages, often yes: Spanish has performed better than many other languages in published evaluations (Taira et al.). "Mostly accurate" still includes occasional errors that reverse meaning, so money, policy, safety, and complaint conversations need human review.
Can bilingual staff just spot-check translations?
Spot-checking helps, but research shows that casual ratings miss errors that expert, context-aware review catches (Freitag et al.). Give reviewers full conversations, a simple error checklist, and a focus on high-stakes message types.
Should our bot use tú or usted?
Pick one deliberately and enforce it. Many service businesses default to usted for first contact and complaints, and allow tú only where the brand voice is clearly casual. Whatever you choose, test that the system holds the choice for the whole conversation.
What about languages beyond Spanish?
Don't assume Spanish results carry over. Published accuracy varied widely by language, down to 55% in one study (Taira et al.). Test each language separately and keep more message types in the yellow and red lanes for less-supported languages.
The bottom line
Machine translation is a strong tool for speed and reach, and a weak tool for accountability. Use it to answer faster, triage better, and draft more efficiently. Keep humans responsible for anything involving money, safety, commitments, or an upset customer. Write cleaner source text, lock down your terminology, decide your register, and measure errors the way experts do. The goal is not a bot that speaks Spanish. It's a customer who trusts what you told them in Spanish.
Sources
- U.S. Census Bureau: New Data on Detailed Languages Spoken at Home and the Ability to Speak English (2017–2021 ACS)
- U.S. Census Bureau QuickFacts: Miami-Dade County, Florida; Florida
- Khoong et al. (2019), Assessing the Use of Google Translate for Spanish and Chinese Translations of Emergency Department Discharge Instructions, JAMA Internal Medicine
- Taira et al. (2021), A Pragmatic Assessment of Google Translate for Emergency Department Instructions, Journal of General Internal Medicine
- Freitag et al. (2021), Experts, Errors, and Context: A Large-Scale Study of Human Evaluation for Machine Translation, TACL
- Google Cloud Translation: Creating and using glossaries (Advanced)
- DeepL: Using the Languages API
- DeepL API reference: Request translation
