Multilingual customer service and the phone problem

Every other channel leaves evidence. A bounced website visit shows up in your reports. An abandoned form shows up. A call that ends in confusion shows up nowhere.

That invisibility is why the problem persists at businesses that have otherwise done the language work well.

How bad is the missed call problem generally?

Bad enough to matter, and it varies enormously by industry.

CallRail's benchmark report, announced 14 January 2025 and based on analysis of 1.1 million de-identified conversations, found missed call rates of 32% in health care, 28% in legal, 14% in home services, and 9% in real estate. Its definition is worth carrying with the number: a missed call is an unanswered inbound call that went to voicemail or was detected as abandoned. Patient Prism, scoring 11,552,668 patient calls across 8,280 dental locations in 2025, found 31 of every 100 calls hung up before reaching an agent at all.

The spread between 9% and 32% is the useful part. A single cross-industry number tells you nothing. Your own rate is the only one worth acting on. Both of those samples also skew toward businesses sophisticated enough to have bought call tracking software, so the true market numbers are plausibly worse.

Here is the number you will be quoted instead, and it is worth being able to dismantle in one sentence. The claim is that 62% of business calls go unanswered. It comes from a blog post by 411 Locals dated 18 January 2016, which describes its method in its own words: "We monitored the phone calls of 85 businesses, operating in 58 industries, for a period of 30 days." That is roughly one and a half businesses per industry. No country is stated, "monitored" is never defined, and no raw data was published. The 62 is 37.8% forwarded to voicemail plus 24.3% getting no response. The same post then pivots into recommending you hire a virtual secretary, which is what 411 Locals sells.

It gets attributed to Forbes, to Ruby Receptionists, and to a ServiceTitan analysis of contractor phone lines. None of those published it. A statistic with three incompatible parents is a statistic with no parents.

Now add a language barrier to whatever your real baseline is. A customer who calls, hears a language they cannot use, and hangs up is not counted as a missed call by any of these systems. The phone was answered. The call still failed.

What actually fails on these calls?

Four things, in roughly this order.

Nobody on shift speaks the language. The most common and most fixable, because it is usually a scheduling problem rather than a hiring one.

The right person cannot be reached fast. The call goes on hold while somebody goes looking. Two minutes of hold is a hangup.

The menu routes to a dead end. A press two for Spanish option that rings a phone nobody carries is worse than no option, because it made a promise first.

The greeting is in one language. A customer hearing only English on your voicemail assumes that is the answer for the whole business, and does not leave a message.

None of these need technology. They need knowing who is working and one handoff.

How do I find the gap on my own schedule?

What is the actual fix, in order?

Start with the cheapest thing that works.

1. Write down who speaks what. You will find capability the schedule does not reflect. 2. Cover your busy hours. Use the weighting above. This alone solves most of it. 3. Build a thirty-second handoff. Whoever answers should know who to get and how. A name and a location, not a search. 4. Record a bilingual voicemail greeting with a callback promise you keep, such as within the hour during business hours. 5. Offer text as an alternative. Underrated. Plenty of customers read and write a second language more comfortably than they speak it over background noise. 6. Consider a professional interpretation line in health care, legal, or anything where accuracy has consequences. These services are routine in clinical settings.

Notice that none of the first five cost anything.

One caution on the callback, because it is weaker than it used to be. A TNS survey from July 2022 found 75% of Americans never answer calls to their wireless phone from a number they do not recognize. Capturing the number and ringing back is not a substitute for resolving the call while the person is on the line, and it is even weaker when your callback will come from a number the customer has never seen, in a language they were not expecting.

The scale of what a missed connection costs is documented on the dental side, and the shape transfers. Patient Prism's funnel shows that of every 100 calls, 13 people walk away without booking, 4.8 receive a followup call, and 0.7 come back and book. Fewer than one in a hundred people who reached out ever walks back through the door after walking away.

Should I use an AI voice agent for this?

Cautiously, and not as your first move.

The honest numbers are not encouraging. Coval, a company running roughly 3.1 million evaluation metrics weekly, reports that around 95% of voice agents work in a demo and about 62% survive their first week live. That is a vendor figure from a company selling evaluation tooling, so read it accordingly.

Multilingual makes it harder, and the reason is measurable rather than impressionistic. Word error rate is the industry's headline metric and it is a weak proxy. Deepgram's own explainer works through a 36-word phone call with nine errors, a 25% word error rate, in which the transcript is perfectly usable for classifying the call except that it renders "declined" as "designed," which is the one word that mattered. Deepgram's conclusion, verbatim: "When a speech recognition API fails to recognize words important to your analysis, it is not good enough, no matter what the WER is."

Now apply that to your business. The words most likely to be mistranscribed on a bilingual call are exactly the ones with no English equivalent: your dish names, your service names, the customer's name. Deepgram also notes that "a 25% word error rate is about average for off the shelf speech recognition APIs."

These tools are genuinely useful in a narrow band: taking a message after hours, confirming a reservation, answering hours questions. They fail on the complicated call, which is exactly the call a second-language customer is most likely making.

Fix the human routing first. If unanswered volume remains, then look at tooling.

How do I know whether my phone is losing customers?

Count. Nobody counts, and that is why the problem survives.

Get call tracking on your main number so you have a record of every call, its length, and whether it was answered. Then listen to twenty calls. That is enough to see the pattern and it takes an hour.

Listen for how long before someone competent picks up, what happens when the language does not match, and how many calls end in under fifteen seconds. Short calls are usually failures. Maple's dataset puts the average restaurant call at one minute 45 seconds, which is a reasonable reference point for what a completed call looks like.

Log the language of every call for two weeks. That gives you a demand picture nothing else will, and it pairs with what the language of your reviews signals.

Confirm the number ringing is the right one, too. Owners are often surprised, which is why checking your Google listing regularly belongs on a schedule.

Then track it over time as part of a small standing set of numbers, because measuring whether multilingual marketing is working depends on call data more than on web traffic.

How do I keep this from breaking when staff change?

Write it down, and attach it to a shift rather than a person.

The usual failure is that one employee has quietly handled every Korean-language call for three years. She leaves, the capability leaves with her, and nobody notices for a month because the calls simply stop converting.

Put language coverage on the schedule the way you would a certification. Make it visible, and make it somebody's job to notice when a shift has none.

The same discipline applies to anything that depends on a person remembering, which is why a review system has to survive staff changes to be worth building.

If Korean-language calls are a real part of your volume, the search side matters too, because Korean-language search brings a specific kind of caller who has already decided and just needs to confirm something.

What to do this week

Ask every employee what languages they speak and write it on the schedule.

Run the coverage map. Capability list, shift count, uncovered shifts, then weighted by when your phone actually rings.

Call your own business three times, once in each language you claim to serve, at different times of day. Note exactly what happens.

Record a bilingual voicemail greeting with a specific callback promise.

Add a text option to your Google Business Profile and your website, and make sure someone watches it.

Then set up call tracking if you do not have it, and listen to twenty calls next month.

Questions about what happens when someone calls you?

Email me at eric@seod.com with your business name and phone number and tell me which languages you serve. I will call you as a customer, in English and in Korean, and send you back exactly what happened, including how long it took and where it broke. I do this myself.

I answer these personally. If your phone handles it well, I will tell you so and stop there.

Or keep reading more on multilingual and community marketing.