AI or a Human Interpreter: Deciding Language by Language

The question is usually put as a single choice for the whole event: are we using AI, or are we hiring interpreters? Framed that way, it has no good answer, because the languages in most rooms do not carry the same weight. One of them is the reason a family is there. Another is a courtesy for two visitors who would have coped anyway.

The better question is asked once per language.

Why per language is a real option

On a single broadcast, an interpreter can take one output language and speak it themselves while every other language stays on AI. Everyone on that language hears the person by default and can switch to the AI if they prefer it. While they stay with the person no model runs on that language, so it costs you nothing in translation — a listener who switches opens an AI session and that language then bills normally for as long as somebody is on it.

One thing to settle before the day: the languages an interpreter may take are an organization setting, and it cannot be changed once a broadcast has started. Tick them in advance, even the ones you are unsure about.

That is what makes the decision granular. You are not buying a philosophy for the event; you are spending an interpreter's time where it changes the outcome and letting software cover the rest. See human interpreters for how the mechanics work.

Where a person is not optional

Start here, because these are the cases where the question is already answered and the only mistake available is to talk yourself out of it.

• Somebody is legally entitled to an interpreter. A hearing, a formal proceeding, anything with a legal consequence for the person in front of you. The entitlement is to a qualified human being, and no AI stream becomes one by being convenient or cheap.
• Consent is being given. Medical, legal, financial. If somebody is agreeing to something on the basis of what they just heard, a person should have said it.
• The words themselves are the product. A poem, a liturgy, a contract read aloud, testimony. Where phrasing carries meaning that a paraphrase loses, a machine that is nearly always right is not nearly good enough.
• It is a two-way conversation. Live translation of this kind carries one speaker to an audience. If both sides need to be understood, that is an interpreter's job and not this tool's.

Where AI is genuinely the better answer

• Languages you would otherwise not offer at all. The honest comparison for a fourth or fifth language is rarely "AI or a professional" — it is "AI or nothing", because nobody was going to book four interpreters.
• Recurring, unglamorous and long. A weekly service, a monthly staff meeting. Booking a person every week for a language with six listeners is not sustainable, and those six are still better served than they would be by silence.
• When you do not know who is coming. A person has to be booked in advance; a language can be enabled on the day and costs nothing if nobody picks it.
• Captions for people who are deaf or hard of hearing. A different need with a different answer — see accessibility and live captions. Note that this is not a substitute for a sign language interpreter where one is required.

The questions that actually decide it

What happens if a sentence comes out wrong?

This is the whole decision compressed into one question. If the answer is "somebody is briefly confused", AI is fine. If the answer involves a consequence — a wrong dose, a signed document, a safety instruction misheard — you are in the first list above.

How many people, and how much does it matter to them?

Audience size is a weak signal on its own. Four people who cannot follow anything without help matter more than forty who are getting a convenience. The per-language figures tell you which languages are being used, and after a couple of events that is worth reading — but which of them people depend on is something you still have to ask. See how to choose which languages to offer for how to read the numbers you do get.

Do you already have the person?

Most organizations that ask this question already have somebody who has been quietly interpreting for two rows, for years. Giving them a channel is usually the highest-value change available, and it costs nothing in translation. They are also, in our experience, relieved: whispering for an hour is harder than speaking into a phone.

Can you sustain it?

An interpreter who can come once is a good week. An interpreter who can come every week is a plan. If the answer is "once", put them on the language that matters most that day and leave the rest on AI, rather than building a dependency you cannot keep.

What happens when the person does not arrive

Worth knowing before you rely on it. If an interpreter does not connect, or their connection drops mid-service, listeners on that language move to the AI translation on their next reconnect, which happens by itself: their app asks for the language again, finds nobody covering it and takes the AI instead. Expect a short silence while that happens — a retry, a fresh session opening for a language that until now had none, and the few seconds of buffer every translated voice is held for. Usually under half a minute. Nobody is stranded on a dead channel.

This is the practical argument for having AI underneath a person even when you have booked one: the fallback is automatic, and the alternative to it is a language that simply goes dead.

A workable default

For an organization starting out, this ordering causes the fewest regrets:

• Run AI on every language for the first two or three events.
• Read the per-language figures and find out which language people actually depend on.
• Put your person on that one — usually somebody you already have.
• Leave the rest on AI, and let people switch if they prefer it.
• Revisit when the room changes, which it will.

And if the event is in the first list — a proceeding, a consent conversation, a two-way meeting — book the professional and do not let a price comparison talk you out of it. We would rather say that here than have you find out on the day.

For how the two run together on one broadcast, see human interpreters. For what the AI side costs, see the pricing page — what a person costs is your own arrangement with them.