The state of multilingual customer support

The language gap in support

Customers are unusually explicit about this one. CSA Research's Can't Read, Won't Buy study — 8,709 consumers across 29 countries — found 76% prefer to buy with information in their native language and 40% will never buy from a business that offers only another language. The support-specific figure is the one worth pinning to the wall: 75% say they are more likely to repurchase when post-sale customer care is offered in their language.

The most useful finding in that study is the one that kills the standard objection. 60% of consumers who are confident in their English still prefer support in their own language. So "our customers speak English fine" is not a defense. It describes what customers will tolerate, not what they would choose.

Unbabel's 2021 Global Multilingual CX Report (n=2,754 across six countries — vendor research with published methodology, and now five years old) found 68% would switch to a brand offering native-language support and 64% would pay a higher price for it. The same survey found 92% say poor service in their own language damages trust — a useful warning against shipping bad translation and calling it access.

The supply side of the same gap

Support organizations mostly know. In Intercom's multilingual support research — which dates to December 2019, and surveyed 135 support leads plus 170 SaaS customers, so treat it as directional and old — 88% of support teams said they offered multilingual support while only 28% of end users saw it. In the same research, 85% of support managers said bilingual reps were hard to find, and 29% of businesses said they had lost customers for the lack of multilingual support.

Nothing since has loosened the constraint. The Bureau of Labor Statistics still counts only about 78,000 employed interpreters and translators nationwide, with roughly 2% projected growth over the decade, while the interpreting market grows from $11.7 billion in 2024 toward $17 billion by 2029 — an 8% CAGR against a 2%-per-decade workforce, as we broke down previously.

The two organizations that actually count

There is no national survey of how many contact-center calls arrive in a language other than English. There are, however, two large public operations that publish their own numbers, and both are more instructive than any vendor survey.

NYC 311, in its July 2024 Language Access Plan, reports that in FY2023 it handled 17,857,232 calls, of which 609,255 required interpretation — 3.41% — across 86 languages. Two details matter more than the headline: Spanish accounted for 87.7% of interpreted calls, and the top ten languages accounted for over 99% of interpretation minutes, with more than 96% of LEP callers requesting Spanish, Mandarin, Russian or Cantonese. A small share of total volume, heavily concentrated, with a very long and very thin tail.

NYC Health + Hospitals shows the trajectory. A New York State Comptroller audit published in November 2025 found interpreted minutes grew from 13.9 million in FY2019 to 35.6 million in FY2024 — a 156% increase in five years — across 255 languages, at a FY2024 cost of $24.1 million. That works out to a blended $0.68 per interpreted minute actually paid, a number worth remembering the next time a vendor quotes retail.

How to add translation to customer support calls

There are four honest ways to get a support call handled across a language barrier, and they are not alternatives so much as tiers. Here is what each actually costs.

Option 1: hire bilingual agents

The most intuitive answer, and the one whose economics are least understood. Start with what employers actually pay for the skill. California's state bilingual differential is $200 per pay period — about $1.15 an hour — and it requires the employee to use the language averaging 10 percent of the time, with explicit instructions to concentrate bilingual duties in as few positions as possible. Los Angeles County pays $0.57 an hour. The City of Los Angeles pays 2.75% of salary. Portland pays $1.00 an hour. Against the BLS median customer service representative wage of $19.08 an hour, that is a 2.6–6% premium — roughly $1,200 to $2,400 per agent per year.

The folk wisdom that bilingual employees earn 5–20% more does not survive contact with the peer-reviewed literature. Subtirelu (Language in Society, 2017) found job ads requiring Spanish advertise lower wages. Churkina et al. (PLOS ONE, 2023), using ACS data from 2005 to 2019, found a 0.8 percentage point wage penalty for bilingual workers, largest among Spanish speakers. Fry and Lowell (ILR Review, 2003) found no significant wage contribution from second-language ability. A 2023 analysis found only 2.9% of 764 U.S. occupations demand meaningful foreign-language proficiency. The premium is real in the specific public-sector schedules above and roughly zero in the labor market at large.

The real problem with hiring is arithmetic. Spanish covers 60.6% of non-English speakers, so one bilingual hire buys a great deal. Language number three is about 3% of that population. Languages four through fifty are fractions of a percent each — and every one of them still requires a full-time person to be available at 4pm on a Thursday. Then apply contact-center reality: 52% average annual agent attrition (Deloitte Digital's 2024 survey of 600 leaders), about $20,800 to replace an agent plus six months to average performance (SQM Group), and the 85% of managers who called bilingual reps hard to find. Hiring covers your top one or two languages. It does not cover a tail.

Option 2: conference in a human interpreter

The incumbent method. Put the caller on hold, dial the over-the-phone-interpreting line, enter a client ID and language code, wait, then run a three-party consecutive conversation. It works, it is defensible, and for high-stakes and regulated conversations it remains the right answer.

The prices are far more negotiable than most buyers realize, because public procurement documents publish them. On Washington State's DES contract, 911 Interpreters bids $0.49/min for Spanish; Lionbridge bids $0.59–$0.79 flat; Voiance holds $0.69/min flat on NASPO ValuePoint; LanguageLine's executed California CMAS price list is $0.97/min plus $5.00 per third-party dial-out. LanguageLine's own retail, no-contract Personal Interpreter service is $3.95/min for audio. Same service, four to eight times apart, depending almost entirely on whether you have a contract. Minimum billing increments matter too: a 15-minute minimum turns a 3-minute call into an effective $4.17 per minute. Our full breakdown lives in what live phone interpretation really costs.

The cost that never appears on the invoice is time. Fagan et al. (Journal of General Internal Medicine, 2003) timed clinical visits at 28.0 minutes with no interpreter, 36.3 minutes with a telephone interpreter — 29.6% longer — and 26.8 minutes with an in-person interpreter. Grover et al. (2012) found emergency department stays of 116 minutes with an in-person interpreter versus 141 with telephonic. A 2026 study in the American Journal of Emergency Medicine covering 47,038 encounters found the delay begins before the interpreted conversation does: median time to physician self-assignment was 7.4 minutes for patients needing an interpreter versus 5.8 minutes without.

Option 3: real-time AI translation on the call itself

The newest tier, and the one that changes the arithmetic. Published per-minute rates for AI speech translation engines sit at $0.034/min (OpenAI's gpt-realtime-translate, 70-plus input languages), about $0.042/min (Azure Live Interpreter) and $0.04/min (Palabra AI's speech-to-speech API). Against competitively bid human interpreting at $0.49–$0.97 and vendor retail at $1.25–$3.95, that is roughly 15x to 100x cheaper at the engine level — before telephony, and before anyone's margin. Finished products cost more than raw engine time: TalkTool's published rate is 30¢ a minute, still well under the cheapest public-contract human rate.

Be honest about latency, because vendors are not. Sub-100-millisecond claims in this market are single-component numbers with the network excluded. Published research measurements of full speech-to-speech systems put ending offsets in the 2.66 to 4.73 second range (Meta's Seamless evaluations), and IWSLT 2024 computation-aware results run 1.92 to 4.07 seconds. For context, a human simultaneous interpreter's ear-voice span is 3.5 to 5.3 seconds, and a natural conversational turn gap is about 200 milliseconds. The fair statement: today's AI speech translation already matches or beats a human interpreter's lag, while remaining many times slower than unmediated conversation. On our own infrastructure we measure sub-2-second latency — that is our measurement of our system, not a claim about the category.

What it buys operationally is the removal of ritual. No dial-out, no hold, no client ID, no language code, no three-party consecutive turn-taking. The friction that Karliner's study identified as the binding constraint is exactly the thing this tier removes. That is the case for an AI phone interpreter on routine volume — and the case for keeping humans on the rest.

Option 4: translated text, as the fallback rather than the plan

Translated SMS and email cost roughly a penny per message — Twilio at $0.0083 per segment plus Google's neural machine translation at $20 per million characters — which makes async follow-up nearly free. Two cautions. In healthcare, 45 CFR 92.201(c)(3) requires a qualified human to review machine-translated text whenever it is critical to rights, benefits or meaningful access; HHS made this explicit in the 2024 rule's preamble. And machine translation quality degrades in exactly the languages LEP callers speak: Khoong et al. (JAMA Internal Medicine, 2019) measured 92% accuracy for Spanish and 81% for Chinese on discharge instructions, with clinically significant harm potential in 2% of Spanish and 8% of Chinese translations; Taira et al. (2021) measured 82.5% overall, ranging from 94% for Spanish down to 67.5% for Farsi and 55% for Armenian. Translated text is a fine fallback channel. It is not a substitute for handling the call.

The sequence to run

Put together, here is the order a support leader can actually execute — cheapest and most reversible first.

1. 1
   Audit your language mix before buying anything

   NYC 311 knows that 3.41% of its calls need interpretation, that Spanish is 87.7% of them, and that ten languages cover over 99% of its interpretation minutes. Pull the same three numbers from your own logs: how many calls, which languages, which queues and outcomes. Nearly every expensive mistake in this market starts with guessing the shape of the tail.

2. 2
   Cover the text channels with what you already own

   Your helpdesk almost certainly includes machine translation for tickets, email and chat — Zendesk, Intercom, Freshdesk and the live-chat platforms all ship some version of it. Turn it on before you buy a translation layer. Just do not assume it touches audio; the vendor documentation says plainly that it does not.

3. 3
   Put real-time translation on the phone line

   This is where the volume is (voice is 53%+ of interactions) and where the built-in tools stop. Real-time AI translation on the call removes the hold-and-dial ritual entirely, which is the difference between language access existing on paper and being used on a busy Tuesday.

4. 4
   Keep human interpreters for the high-stakes tier

   Consent, diagnosis, legal exposure, formal complaints — anything a regulator might read a transcript of. Negotiate a contract rate rather than paying retail; the public procurement rates above ($0.49–$0.97/min) are your benchmark. Define the escalation trigger in advance so agents are not making that judgment call under pressure.

5. 5
   Measure CSAT and resolution per language, then iterate

   No published benchmark exists for native-language versus machine-translated support CSAT — the industry does not collect it. So collect your own. Split first-contact resolution, handle time and CSAT by caller language. It takes one field, and it is the only way to know whether your coverage is working.

The bottom line

The multilingual customer support tools market is mature on text and nearly empty on voice — which is backwards, because voice is more than half of all interactions and the channel customers choose when the stakes are highest. The helpdesk vendors document their own voice gaps. The interpreting incumbents are excellent and expensive, and the research says friction, not availability, is what stops them being used. And the industry's biggest research franchises do not measure language at all, so nobody is being told how large the hole is.