Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- Why does "just hire a native speaker" stop working?
- What has to happen on a multilingual ticket? The Three Jobs
- How much does each approach cost per month?
- What do native-speaker hires buy?
- What does a translation vendor buy?
- What does an AI agent buy?
- What does this look like in real rollouts?
- Edel Optics: 75-79% resolution once the order data was connected
- YouGarden: 965 hours a month at 12,000 tickets
- TravelJoy: 80% resolution, up from 24% on Zendesk's own agent
- A six-market European retailer: 72% across all six
- What to do this week
- How do I get AI to name the failing job for me?
- When should you still hire a native speaker?
- The takeaway
- FAQs
Created time
Sep 7, 2026 02:59 PM
Title length (<60)
Author
Last optimised
Sep 15, 2026
Ecomm?
Tags
Image
Publish date
Aug 15, 2026
Video
Slug
multilingual-customer-support-without-hiring-native-speakers
Featured
Type
Article
Ready to Publish
Ready to Publish
If your plan for multilingual customer service is a hiring plan, you are staffing a problem that no longer needs staff.
You can run multilingual customer support without hiring native speakers. Put an AI agent on the helpdesk you already use, and let it detect each customer's language (ours handles 95 of them) and reply in it.
Connect it to your help center and your order data. Route the conversations it should not handle to a human, with translation on top.
That handles the everyday volume. Across our four rollouts, 66% to 80% of tickets are resolved without ever reaching a human, and hiring then shrinks to a decision about a handful of languages.
Two camps publish the standard advice. Translation platforms and outsourced desks want you to staff up first: Phrase's step one is to hire as many local-language agents as possible, and Hugo's whole answer is a trained multilingual offshore team.
Helpdesk vendors put translation tooling in place of the hires. The demand behind both is real: CSA Research's 2020 survey of 8,709 consumers across 29 countries found that 76% of online shoppers prefer to buy products with information in their native language, with 40% never buying from a site in another one.
My take is that a multilingual ticket needs three separate jobs done, and only one of the three has ever needed a native speaker.
Work out which job is failing and the hiring question usually answers itself.
I'm Mike, co-founder of My AskAI, and we run AI support agents for 200+ ecommerce and SaaS businesses, inside Zendesk, Intercom, Freshdesk, Freshchat, Gorgias and HubSpot.
Our agents have resolved more than 1,000,000 tickets, at a 72%+ resolution rate on a rolling 30-day basis. The numbers in this post come from four of those rollouts: a European eyewear retailer, a UK garden center, a travel-software team and a six-market European retailer.
Why does "just hire a native speaker" stop working?
TL;DR: Coverage is a grid of every language your customers write in by every hour they write in. Each hire fills one square of it.
The helpdesk camp already argues against hiring. Freshworks says on its own multilingual page that "manually hiring native speakers for every language is costly and hard to scale". Its FAQ says small businesses don't need to hire native speakers for every language right away.
Zendesk puts the in-country team at the heavy end of the spend (and it sells the helpdesk those teams sit in). Its own page says "if you have a significant presence within a particular market, you may need an in-country support team of native speakers—but this might also require outsourcing".
Helpware, which staffs multilingual desks for a living, puts the case as fluency plus culture: "your CSRs must possess a deep knowledge of all the cultural nuances, references, and slang of each region where you want to expand".
That is true, and it costs a salary per region. It is also the argument I make for putting a person on the small number of conversations where the wording decides the outcome.
Coverage runs on two axes: languages across the top, hours down the side. One hire fills one square of that grid.
A Spanish-speaking agent in Madrid does nothing for the 3am Spanish ticket (Edel Optics sells into 53 countries, so good luck shift-planning that). Zendesk describes the same axis: your customers "could be in another time zone or continent and want to reach out to your business when your support team is no longer on the clock".
Below your top two or three non-English languages sits a tail that will never justify a headcount (Dutch, Polish, Danish, whatever yours turn out to be), at any volume you are likely to reach.
Those customers still write in, and they are the ones who wait longest for a reply. A hiring plan has nothing to say about them. I've sat on that call, and the answer comes back as "not yet, maybe next year".
The second reason the trained-human answer has weakened is information. An AI agent reads the whole help center, every historic ticket and the customer's live account state through an API.
The rented equivalent starts each shift with a training deck, a limited login and whatever the last handover note said.
What has to happen on a multilingual ticket? The Three Jobs
TL;DR: Every multilingual ticket needs three jobs done: understand the question, know the answer, say it back correctly. Hiring covers all three in one language. A translation vendor stops at two. An AI agent covers all three in 95 languages, once you connect it to your knowledge and your data.
The three approaches buy different jobs. I call the three jobs Understand, Answer and Express.
Understand is reading what the customer asked, including the idiom, the typo and the half-sentence. Answer is knowing your policy, your product and this customer's live account state. Express is replying in their language, in your voice, well enough that nobody screenshots it.
All three have to happen at the hour the ticket arrives, and across our customer base that is every hour of the week.
A hiring plan cannot buy that. Answer is the job no translator does. Most of what we do on a multilingual rollout is Answer work: connecting knowledge and live data.
How much does each approach cost per month?
The sum below is an illustration built on two assumptions: a team taking 3,000 tickets a month, and 600 of them arriving in a language nobody on the team reads. Both are round numbers to divide and nothing more (your own split comes out of step 1, further down). Run the same three columns on your numbers.
Native-speaker hires | Translation vendor | AI agent (My AskAI) | |
Monthly cost at 3,000 tickets | One salary per language, per shift, at your market's rate | $18,000 on the illustration above, at $0.15/word (600 tickets × an assumed 200 words) | $439 on Pro ($199 base + 2,000 extra tickets × $0.12) |
Languages covered | One per hire | The set you contract for | 95, auto-detected per message |
Coverage hours | The shifts you staff | The vendor's hours, plus turnaround | Every hour tickets arrive |
Understand | Yes | Yes | Yes |
Answer | Yes, once trained | No. It works from your glossary with no view of your order table | Yes, once connected to your knowledge and live data |
Express | Yes | Yes | Yes |
Who reads the ticket text | Your employee | A sub-processor outside your stack | Your AI agent, inside your helpdesk |
Time to live | A hiring cycle per language | Contract, then a glossary build | Days |
Our Pro plan is $199 a month with 1,000 tickets included, then roughly $0.12 for each extra ticket. It comes with 5 team seats, so more than 5 people needing a My AskAI seat crosses to Scale.
We charge per ticket, resolved or not, so the rate does not climb as the AI gets better. AI Tagging, AI Actions, Tasks, Live Translation and Image Processing are charged per use on top, on every plan including Pro.
What do native-speaker hires buy?
Hiring buys Understand and Express outright, and Answer over time (a trained person picks your product up over the first few months).
What it costs is a salary per language, per shift, so do the grid arithmetic before you write the job spec. Three languages on three eight-hour shifts is nine seats, before holiday and sickness cover (and before anyone quits). Follow-the-sun rotas and overlapping regional shifts move that number around, and you are still multiplying languages by shifts to get it.
It breaks on the tail, where the volume never justifies a seat, and on the clock, where nobody is on shift at 3am.
What does a translation vendor buy?
A translation vendor or an outsourced multilingual desk buys Understand and Express. Answer stays on your side of the wall.
The vendor has your glossary and your tone guide. Your order table stays behind your own auth (nobody gives a sub-processor that access). A "where is my parcel" ticket comes back to you or waits.
Smartling states that "in the United States, professional translators usually charge between $0.15 and $0.30 per word".
It also adds a hop: every message goes out and comes back, so first reply time grows by the vendor's turnaround.
Your customers' ticket text leaves your stack and lands with a sub-processor, and your security reviewer will ask about that. Put the same question to an AI vendor.
We hold SOC 2 Type II and GDPR, encrypt at rest with AES-256 and in transit with TLS. Each customer runs in isolated containers, and we never use customer data for model training.
What does an AI agent buy?
An AI agent buys all three jobs. Ours handles 95 languages, auto-detected per message, and replies in the customer's language by default.
You can also lock it to one language where brand consistency wins. It works at 3am because there is no shift to staff.
An AI agent only wins the Answer job when it is connected to your knowledge and your live data.
So connecting the data is the first thing we set up on a multilingual rollout.
Where a conversation has to reach a human who cannot read it, Live Translation turns the customer's message into the agent's language as an inline internal note, then translates the agent's reply forward on send. That runs on Intercom, at $0.05 per ticket. A Shopify subscription platform on Intercom uses it so its agents can pick up non-English merchant threads on handover.
What does this look like in real rollouts?
TL;DR: Four live rollouts, resolving 66% to 80% of tickets without escalating to a human. In one of them, connecting order data lifted resolution from 20-30% to 75-79%.
We count a conversation as resolved when the AI handled it without handing off to a human. Hold anyone else's numbers to the same definition.
Edel Optics: 75-79% resolution once the order data was connected
Edel Optics is a European eyewear retailer selling across 53 countries, running Zendesk. The agent auto-detects language, defaults to German and switches for everyone else. Its knowledge is the Zendesk help center plus a User Data API connection.
Resolution sat in the 20-30% range until we connected the User Data API, which gave the agent order, delivery, return and tracking information. It now runs at 75-79%.
The agent read and wrote German, English and everything else from the first day. Connecting the order data is the Answer job.
The same rollout runs 92% AI CSAT across 4,067 tickets and saves around 150 hours of agent time a month.
YouGarden: 965 hours a month at 12,000 tickets
YouGarden is a UK online garden center on Freshdesk, taking about 12,000 tickets a month through our agent. It resolves 66% of them, peaking around 82%, at 78% AI CSAT across 11,785 tickets. That is 965 hours saved a month, or roughly six full-time agents' worth of work.
Mamunur Rahman, their Head of Customer Service, described the effect to us:
"My AskAI has fundamentally changed how we support our customers. The quality and consistency of responses are extremely high, and it's allowed us to scale support without compromising the experience we're known for at YouGarden."
TravelJoy: 80% resolution, up from 24% on Zendesk's own agent
TravelJoy is a travel-advisor SaaS on Zendesk. Its AI resolution rate is 80%, up from 24% on the helpdesk's own AI agent, and the team saves 193 hours a month.
A six-market European retailer: 72% across all six
A baby-and-family-textiles retailer we work with sells across six European markets on Gorgias and Shopify, and runs a single multilingual configuration for all six.
It resolves 72% of about 1,000 tickets a month at 94% AI CSAT, with Self-Learning's auto-drafted articles covering roughly 200 of those tickets.
Our own AI resolution rate benchmark study puts the field median at 70% across 195 rated deployments from 38 vendors. It aggregates self-reported rollouts, so take it with a grain of salt.
What to do this week
TL;DR: Before you write a job spec, count last month's tickets by language and by hour. The count usually settles the question.
- Pull last month's tickets by language and by hour (~1 hour in any helpdesk's reporting). You will see which languages are real volume, which are the tail, and how much of it lands outside your team's shifts. Hugo recommends the same first move: "use data analytics to identify all of the languages used by your customer base. Analyze customer location data, website language preferences, and support ticket language requests".
- Name which of the three jobs is failing for each of your top three non-English languages (~30 minutes). If the failing job is Answer, no amount of language hiring will fix it.
- Run one language in internal-notes mode for two weeks (~15 minutes to set up). The AI drafts every reply as an internal note for a human to read before anything reaches a customer. Both Edel Optics and YouGarden started this way before switching to direct replies.
- Connect the live data source behind your top non-English ticket type (a sprint, or less). At Edel Optics this one change moved resolution about 50 points. Order status, subscription state, delivery tracking: pick the one that generates the most tickets.
- Decide your handover language rule before you go live (~30 minutes). When the AI hands a Portuguese conversation to an agent who reads no Portuguese, what happens? Live Translation is one answer, and routing to a named agent is another.
How do I get AI to name the failing job for me?
Steps 1 and 2 both come out of one ticket export. Paste yours into any AI chat tool with the prompt below and it will do the sorting for you (it cannot judge answer quality, so read the output as a shortlist rather than a verdict).
You are helping me decide how to cover multilingual customer support.
Here is my ticket data for last month, exported from my helpdesk:
[paste ticket counts by customer language and by hour of arrival]
Here is what my support team can already read and answer today:
[list the languages your team covers, and the hours each person is on shift]
Work through this in four steps.
1. Split my languages into three groups: real volume, thin tail, and one-off.
2. For each language with real volume, name which of these three jobs is
failing, and say why:
- Understand: reading what the customer actually asked, idiom and typos
included.
- Answer: knowing my policy, my product, and that customer's live account
state (order status, subscription, delivery).
- Express: replying in their language, in my brand voice.
3. Say how much of each language arrives outside my team's shift hours.
4. For each language, recommend one of: hire a native speaker, use a
translation vendor, or put an AI agent on it. One sentence of reasoning
each, referring back to the failing job you named in step 2.
Where my data does not support a conclusion, write "unverified, check this
yourself" instead of guessing.
Finish with one table: language, monthly tickets, share arriving outside
shift hours, failing job, recommendation.When should you still hire a native speaker?
TL;DR: Three cases where the hire is still the right call, and one where a translation vendor beats the other two approaches.
Thin-coverage languages that are also high-stakes need presence in the market, which only a hire buys. Say you sell into one market where a mis-worded reply loses the account (a distributor, a regulator, your biggest customer's home country).
A person who lives in that language is a fair line on the budget, whatever the volume says. Licensed, regulated or legally binding conversations in-language belong to a person too.
So does a team with nothing written down. With no help center and no ticket history, an AI agent has nothing to work from. A trained human spends the ramp-up period building that knowledge instead.
Where there is a ticket history but no articles, our agent drafts starter knowledge from the last 5,000 historic tickets.
Marketing, brand and long-form localization are where a translation vendor beats the other two approaches outright. Phrase sells a translation management system with glossaries, QA tooling and a second linguist reviewing the output. I send people there for campaign copy.
Freshworks makes the narrower version of the same point about support: "adding a layer of human review for critical communications helps maintain accuracy and brand voice".
Culture is the other case for a person. Freshworks again: "translating words alone isn't enough to create meaningful customer experiences".
The takeaway
TL;DR: Multilingual support is three separate jobs. Work out which one is failing before you write a job spec.
The hiring plan never quite closes because a headcount number cannot tell you which of the three jobs is failing.
What you are short of is three jobs getting done in a language nobody on your team reads, at the hour the ticket arrives. Understand, Answer and Express: that is how I break a multilingual ticket down.
Count last month's tickets by language and by hour, then name the failing job for your top three non-English languages.
Want to see what an agent does with them before you commit? Our trial runs 30 days, every feature unlocked, no card.
FAQs
What does multilingual support mean?
Multilingual support is support delivered in more than one language. An international customer asks a question and gets an answer in the language they think in.
It covers every channel a customer might reach you on: chat, email, help center articles and social. In practice it also means detecting which language a customer is writing in, so nobody has to pick one from a dropdown (ours does that per message).
What is bilingual customer support?
Bilingual support is two languages, normally your local language plus one high-volume second one, which in the US is usually English plus Spanish.
It stays a hiring answer for as long as the mix really is two languages, because you can staff it with a handful of people. The moment a third language arrives you are back on the languages-by-hours grid, where hiring stops scaling with the problem.
Multilingual vs multi-language: what's the difference?
The terms are used interchangeably in support tooling, and nobody will misunderstand you either way. Where a distinction is drawn, it runs like this:
Term | What it describes | Typical example |
Multi-language | A system holding content in several languages | A help center with per-locale articles |
Multilingual | An interaction handled in the customer's language | A ticket answered in Polish |
Can you outsource multilingual customer support?
Yes, and it is a real market. Providers like Hugo and Helpware sell trained multilingual desks. They are clear about what you are buying: fluency plus cultural knowledge, in a team you don't have to recruit (which, if recruiting is your bottleneck, is worth a lot).
The Answer job stays with you, because an outsourced desk works from a glossary and a training deck with no view of your live order data. If the wider outsourcing decision is the one you are weighing, we have a longer piece on its pros and cons.
Do I need a multilingual help center as well as a multilingual agent?
Yes. They do different jobs. The help center serves the customer who searches before writing in, and Zendesk is right that "offering self-service content in various languages not only improves CX but also reduces the number of support tickets".
The agent picks up from there. At professional translation rates, translating a whole help center into six languages is a budget decision (do that arithmetic before you commission it). An AI agent that answers in 95 languages does not need the articles translated to answer from them.
What are multilingual customer support best practices?
Start with your language-by-hour mix, then name which of the three jobs is failing in each language.
Pilot in internal-notes mode so a human reads the first non-English replies (both Edel Optics and YouGarden started that way). Then connect the live data behind your biggest non-English ticket type and decide the handover language rule before you go live.
How should multilingual customer support tickets be routed and measured?
Route by detected language at the point a human takes over, so an escalated conversation lands with someone who can read it or with translation attached. Measure resolution rate per language, because a healthy overall number will hide a language that is failing.
Resolution means the AI handled the conversation without escalating to a human. Our benchmark study puts the field median at 70%. Hold each language against that.
