Welcome, humans.

This is a special edition of The Neuron where we grab one topic and pull on it from every angle.

Today’s topic: the invisible tax language puts on global work, and what changes when AI can finally follow a meeting across dozens of languages.

Most companies already call themselves global. Their people sit in different countries, customers speak different languages, and a single project might bounce from Tokyo to Paris to São Paulo before lunch. Then everyone joins the meeting and quietly agrees to squeeze the whole company through one shared language.

Which works great, right up until “technically fluent” turns into “I understood 80% of that and nodded through the other 20%.”

The interesting shift is that AI is getting good enough at speech, language detection, transcription, and summarization that the common-language meeting may stop being the only practical option.

That could change way more than note-taking. It could change who gets heard, where companies hire, how sales teams sell, and whether a decision made in one language becomes useful to the rest of the company.

The problem was never just translation

Picture a team with six people from six countries. Everyone speaks English well enough to work in it. On paper, problem solved.

In practice, language changes the meeting before anyone notices. The person speaking their first language can jump in quickly, interrupt naturally, make a joke, soften a disagreement, and improvise when the conversation takes a turn. Someone working in a second or third language is often doing an extra job at the same time: translating the thought before they can contribute it.

Researchers studying multinational teams have found two different kinds of language friction. The obvious kind is vocabulary and grammar. People miss a word, struggle to phrase something, or decide a point is too hard to explain. That reduces participation.

The sneakier kind is what researchers call pragmatic and prosodic transfer: the tone, emphasis, politeness rules, and conversational habits from one language carrying into another. That can change how people interpret the exact same sentence. A Journal of World Business study found both types can interfere with how multinational teams process knowledge.

That distinction matters because a perfect dictionary still would not solve the second problem. You can translate the words and miss what the person was trying to do with them.

Accents are only the beginning

Now add accents, local slang, product names, customer names, internal acronyms, and the phrase every company eventually invents that means something completely different inside the building than outside it.

Then add code-switching, where someone moves between languages inside the same conversation, sometimes inside the same sentence. That is normal human behavior in multilingual groups. It is also still a difficult technical problem. A 2026 systematic review of code-switching speech recognition research found the field is growing, but datasets, language coverage, and evaluation are still uneven.

Translation demos love clean sentences. Work loves “yeah, tell Carlos el deck needs updating before EOD.”

And then the meeting disappears

The biggest loss often happens after the call. A decision gets made in Spanish, a follow-up gets typed in English, the CRM note is two sentences long, and three weeks later nobody remembers why the team chose option B.

So the real multilingual-work problem has three layers:

Participation: who feels fast and confident enough to speak.
Understanding: whether everyone interprets the same meaning, tone, and intent.
Memory: whether the useful parts survive the meeting in a form the rest of the company can actually find.

AI is starting to attack all three at once.

FROM OUR PARTNERS

Your meeting notes just learned 32 languages

Back-to-back meetings are hard enough before half your team starts switching languages mid-sentence. Granola, the AI notepad built for people who live in meetings, now supports 32 languages. Turn on multi-language mode once and it follows along whether the conversation moves between English, Spanish, French, Japanese, or others. Granola transcribes from your computer while you jot down what matters, then turns the conversation plus your notes into clear summaries, action items, and next steps. Use code THENEURON and get 3 months on us.

What AI is actually doing under the hood

“AI understands 32 languages” sounds like one feature. Underneath, several hard problems have to work in sequence.

First, the system has to turn speech into text. That is automatic speech recognition, or ASR. It sounds simple until five people have different microphones, one person is in a coffee shop, two people talk over each other, and the customer’s company name is a word the model has never seen before.

Second, the system has to know which language it is hearing. In a normal multilingual conversation that language can change halfway through. The model has to identify the switch quickly enough that it does not keep forcing Spanish words into an English transcript, or vice versa.

Third comes the part people usually call “translation,” although modern meeting AI can do more than produce a second-language copy. A language model can take the transcript, the notes someone wrote during the call, the calendar context, and the surrounding conversation, then turn that mess into a structured summary: decisions, objections, action items, owners, dates, and next steps.

That last step is where this gets much more interesting for work.

A transcript is evidence. A summary is a shared object.

A raw transcript says what was said. Useful workplace memory says what happened.

Imagine a sales call conducted mostly in Japanese. The account executive can stay focused on the customer instead of translating every sentence into polished CRM notes while the conversation is still happening.

Afterward, the important bits can be organized for the broader team: what the buyer cares about, what they objected to, what was promised, and who owns the next step.

Or imagine a product meeting where two engineers bounce between English and Spanish because that is how they naturally think through the problem. The rest of the company does not need a literal line-by-line translation. It needs the decision and the reasoning behind it.

This is the boring superpower: turning “I wasn’t in that room, and I don’t speak that language” into “I can still catch up in five minutes.”

That shift also explains why meeting notes matter more than live translation alone. Live translation helps the current conversation. Structured notes can make the conversation reusable later, by people who were never there.

There is already evidence that lowering language friction changes behavior

An NBER study of multinational companies found that language barriers reduced interactions between foreign and domestic managers. When domestic managers were given English training, those interactions increased. Different intervention, same lesson: when the communication cost drops, people exchange more knowledge.

A smaller experiment on global teams found something similar from the software side. Researchers gave some teams machine-translated logs of subgroup conversations before the main meeting.

Those teams showed better meeting quality than groups that did not get the translated context. The study was small, so treat it as directional evidence, not the final word. But it points at the bigger opportunity: translation can unlock context that used to stay trapped inside language-based subgroups.

FROM OUR PARTNERS

One setting, 32 languages

Turn on Granola’s multi-language mode once, and it can transcribe meetings across 32 languages while you take the notes that matter. Afterward, Granola turns both into summaries, action items, and next steps. Code THENEURON gets you 3 months on us.

What a truly multilingual workplace changes

If this technology keeps improving, the most important change may be that companies can stop treating language fluency as a proxy for professional fluency.

Those are very different things.

Global sales gets less “telephone game”

International sales teams constantly translate between the customer’s language and the company’s operating language. The seller has the call, writes notes, summarizes the notes for a manager, then someone else turns that summary into a proposal or handoff.

Every step loses a little information. The wording gets cleaner while the original nuance gets thinner.

Multilingual meeting AI can shorten that chain. The original conversation becomes the source, and different people can pull the version they need from it later. Sales operations gets structured action items. Product gets feature requests. Leadership gets account risk. Nobody needs one exhausted account executive to manually retell the same meeting four different ways.

Hiring expands when “great English” stops being an unofficial job requirement

Plenty of jobs require strong communication. Fewer jobs actually require every candidate to do their best thinking in the same language as headquarters.

Better multilingual transcription and summarization could widen the pool for interviews, onboarding, internal mobility, and cross-border collaboration. A candidate can answer a complex question in the language where they can be most precise, while the hiring team still gets usable notes and follow-up context.

There is an obvious caution here: high-stakes hiring decisions should never be outsourced to a transcript or summary. Accents, phrasing, and translated language can introduce new bias if people mistake the AI output for a perfect representation of the speaker.

Collaboration gets less performative

A lot of multilingual work has a strange hidden ritual. People spend energy proving they can participate in the company language instead of spending that energy on the problem.

If the documentation layer gets good enough, teams can be more flexible about how the conversation itself happens. People can switch languages when it helps them explain something precisely, then let the system create the common record afterward.

That does not mean every meeting turns into 12 languages at once. It means language choice can become a workflow decision instead of a status signal.

Company knowledge stops belonging to whichever language the meeting happened in

This may be the biggest one.

Most company knowledge already lives in conversations: customer calls, project reviews, hiring interviews, design critiques, one-on-ones, and “quick” meetings that somehow become the place a major decision gets made.

When those conversations are searchable only by the people who attended, language is one more wall around the information. A multilingual AI layer can turn the wall into a doorway. The original conversation stays intact, while summaries and search can make the useful parts accessible across the organization.

In other words, multilingual AI is slowly turning language from a storage format into an interface.

FROM OUR PARTNERS

Bring 32 languages into the same notebook

Granola transcribes meetings across 32 languages and turns your notes into summaries and next steps. Code THENEURON gets you 3 months on us.

What AI still cannot translate for you

There is a tempting version of this story where language barriers simply disappear. The technology is useful precisely because reality is messier than that.

AI can get closer to the words. Humans still have to handle the relationship.

A colleague saying “that might be difficult” can mean “maybe,” “absolutely not,” or “I am trying to disagree without embarrassing you,” depending on the person, team, and culture. A transcript can capture the sentence perfectly and still miss the social information wrapped around it.

Names and internal jargon will still break. Low-resource languages will improve at different speeds. Code-switching will remain harder than clean single-language speech. Noisy rooms and overlapping speakers will create errors. And any company using transcription needs clear expectations around consent, transparency, access, and retention.

That means the winning workflow probably looks less like “AI translates everything, humans stop thinking” and more like this:

Let people communicate naturally when doing so improves precision or participation.
Use AI to create a common record that everyone can search and understand later.
Keep humans responsible for meaning when the stakes are high, especially hiring, legal decisions, customer commitments, and sensitive feedback.
Build a shared company vocabulary for names, acronyms, products, and recurring terms so the system has a fighting chance at getting them right.

The next test is not whether AI can translate enough words. It is whether companies can finally let people contribute in the language where they think best without turning that choice into a collaboration penalty.

If that works, the 32-language workplace will feel surprisingly ordinary. The meeting happens. People speak naturally. The useful parts survive. Everyone else can catch up.

And one of the oldest bottlenecks in global work quietly moves into the background.