People
Multilingual AI employees
Tobiloba Odejinmi · 16 Jun 2026 · 6 min · 925 words
Direct answer
Language is not a model setting you flip. A multilingual AI employee needs a source of truth in each language you will send, a reviewer who can defend that language, and a rule for what happens when retrieval is empty. If you only have reviewers in English, you can draft in other languages. You cannot ship in them unattended.
- Drafting in a language is cheaper than standing behind it.
- Legal and money language need an approved set, not a fluent guess.
- Name a reviewer per language you actually send.
- Empty retrieval in one language should escalate, not code-switch into confidence.
Language is not a dropdown
Vendors will show you a list of languages and a fluent paragraph in each. That is a demo of a model. It is not a job. A job has a customer, a tone, a legal line you will not cross, and a person who can say the reply is wrong.
I work with teams who already serve more than one language because the market is like that, not because a setting exists. The constraint is almost never whether the model can produce Yoruba, French, or Arabic. The constraint is whether anyone on your side will stand behind the sentence.
Tone is local. Templates are too.
Support in Lagos does not sound like support in Berlin. A hiring note that feels warm in one language reads like a brush-off in another. If you generate from an English template and “just translate,” you will ship the English company’s manners into a room that did not ask for them.
Write the tone rules in the language you send, or accept that a reviewer will rewrite them every time. I would rather maintain three short templates than one global voice document that nobody can apply. Voice documents are where multilingual projects go to become slides.
What the model can do versus what a reviewer must do
The model can draft. It can classify intent across languages better than most tired inboxes. It can pull a field from a form. It cannot decide that a refund phrase is safe in a language nobody on the team reads. That decision is a person, or it is a gamble.
Put the reviewer on the job description. If you do not have one for a language, the AI employee’s job in that language is “draft and escalate,” not “send.” This disappoints people who wanted a flag-shaped feature. It also prevents the only kind of miss you cannot explain: we do not know what we said.
Legal text is not a style
Terms, privacy lines, hiring notices, and anything that sounds like a promise should come from an approved set. NYC LL144 notice language is a good example. You do not invent candidate notice in a second language because the model is confident. You translate once, you have counsel look if the risk is real, and the system is only allowed to use that text.
The EU AI Act’s transparency duties do not get easier because the user wrote in another language. If you owe an explanation, you owe it in a language the person can use. That may mean you do not operate in a market yet. Say that. Do not let fluency hide a gap.
Handover across languages
Handover has to name who covers each language when the owner is out. Escalations need to land with the payload in the original language, plus a draft if you have one. Do not escalate a lossy English summary and throw the source away. The source is the work.
I also want evaluation sets per language you send, even if they are small. Ten real tickets. The expected shape. If you only evaluate English and then turn on four other languages on Friday, you did not go multilingual. You widened the blast radius.
How I would start
Pick the language that already has volume and a reviewer. Ship that job. Add the second language when the sources exist, not when a stakeholder asks for a flag on a slide. Count the pile by language. You will find that one of them is the real product and the others are politeness.
If the business is already multilingual in the human team, the AI employee should follow that map, not invent a wider one. Meet people in the inbox they use. Keep the owner human. Language is a people problem that happens to show up as text.
Questions people ask
Can one AI employee work in several languages?
It can draft in several. It can only finish in the languages where you have sources and a reviewer. Those are different claims. Write them separately in the job description.
Do we need a separate workflow per language?
Not always. You often need separate sources, templates, and escalation paths. One workflow with three empty knowledge stores is still three jobs you have not staffed.
What about machine translation?
Fine for internal understanding. Risky as the last step before a customer or a candidate sees it. If you translate a policy and then generate from the translation, you stacked two places to be wrong.
How do we handle mixed-language tickets?
Detect, route, and keep the original text with the case. Do not silently translate away the words the customer used. Reviewers need those words. So do you, later, when someone asks what was promised.
Is this a model problem?
Rarely first. It is usually missing sources, missing reviewers, or a tone that is acceptable in one market and rude in another. Fix the job before you swap models.
Written by
Tobiloba Odejinmi
Head of Engineering at 10mg Health. I have run engineering at Zeeh Africa and sold Insurpass and Shopl. I still write the code. If you have one process that still runs on people copying things, we can look at it in thirty minutes.
