Five languages. One library.

Every row on this board is the same question, asked by a different customer, in a different language, through a different door. Watch which columns change and which never do. The language changes. The door changes. The English copy of the question, and the gate it boards from, stay exactly where they are.

That is the whole idea of this page in one picture. A question can arrive in any of five languages, but it is checked against one set of records, in one language, under one set of rules, and it ends the same way in every row. Grounded once. Spoken five ways. Pick any row to print its boarding pass.

Strip an AI assistant down to its model and you have something that writes fluent sentences in dozens of languages and knows nothing about your business. It has never read your warranty policy. It cannot tell which part fits which vehicle. It has no idea who is asking, what they bought, or what it is allowed to promise. Fluency is not knowledge, and in a second language fluency is even more persuasive, because the reader has fewer ways to catch a confident mistake.

Everything that turns that model into a dependable assistant sits around it. The definition this library uses is Deyaf’s:

The industry has converged on the same split. LangChain states it as a formula, agent equals model plus harness, and derives the rest of the machinery from that one line. Philipp Schmid’s analogy is that the model is the CPU and the harness is the operating system: the processor does the work, but the operating system decides what runs, what it may touch and what gets remembered. An airport is a friendlier picture of the same thing. Aircraft fly. Airports decide who boards, which gate they leave from, what goes in the hold, and what happens when a flight cannot leave.

Language makes the split unusually easy to see, because every clause of the definition hides a multilingual question. What the assistant knows: one library, or five? What it remembers: does a Spanish conversation land in the same record as an English one? Where it meets customers: which doors speak which languages, and in which voice? What it may do: do the rules read the customer’s words, or a translation of them? How your team teaches it: does a lesson written once reach every language, or must it be taught five times? A model cannot answer any of those. A harness has to.

Read the plan from the left. A question arrives at a door (website chat, website voice, the phone, a text or an email) in the customer’s own language. A detector, running separately from the model that writes the answers, works out which language it is. If the question isn’t English, the harness makes one hidden English copy. Everything airside reads that copy: the rules, the guards and the search. The answer is grounded in English evidence and comes back out through arrivals in the customer’s language, spoken on a call in that language’s one fixed voice. The customer never sees the English. The rules never read anything else.

Notice what is not on the plan: a second library, or a second rulebook. Translation lives at one line in the building, and everything behind it is singular.

There are two ways to build an assistant that speaks five languages. The obvious one gives each language its own knowledge: translate the manuals, translate the policies, translate the lessons, and let each language search its own shelf. It demos beautifully. Then the warranty policy changes. Someone updates the English page on Monday and the Spanish page on Wednesday, and nobody remembers the Arabic one at all. For a week, the same question gets different answers in different languages, and nobody notices, because nobody reads all five.

The second way is less glamorous: keep one library, in one language, and move translation to the edges. A question in French is rendered into English, searched against the English records, checked against the English rules, and answered in French. There is exactly one warranty policy. When it changes, it changes for everyone at once. A lesson your team writes on Tuesday is on the only shelf there is, so every language can find it from the moment it is published.

One shelf also keeps the evidence honest. With five shelves, a Portuguese answer points at a Portuguese document that may itself be a stale translation. With one shelf, every answer in every language points at the same record, one a person on your team can open and read.

Why non-English turns cost more

Translation at the edges is not free, and pretending otherwise is how multilingual projects overrun their budgets. BabelArena, a September 2026 study of agents working across many languages, found that non-English runs used up to twice as many tokens as English ones, made more tool errors, and tended to slip back into English when asked for structured output. Every trip across the language line is extra work. A harness that detects the language cheaply, apart from the answering model, and makes a copy only when a message isn’t English, pays that cost only where it must.

Why translated benchmarks flatter you

The same trap waits in testing. The quick way to “evaluate” five languages is to machine-translate your English test set and run it again. LILT’s February 2026 analysis argues that benchmarks built this way end up measuring translation artifacts rather than how an agent copes with real speakers. A translated question comes out tidy and literal, nothing like what a customer thumbs into a phone at eleven at night. The honest version is slower. Collect questions in the words customers actually use, keep the hard ones in the set on purpose, and grade answers against the source, not against a translation of the answer you expected.

Count libraries, not languages

Be wary of language counts, too. Intercom says its Fin agent works in more than 45 languages (vendor-reported). Klarna’s 2024 launch announcement said its assistant communicated in more than 35 (company-reported); a year later, Bloomberg reported the company turning back toward human customer service. A count tells you which doors are open, not whether each language is grounded in the same records or tested in its customers’ own words. The better question to ask any vendor, this one included, is not “how many languages?” but “how many libraries?”

EVE is Feniex’s assistant, and Feniex is pronounced like “Phoenix”. She describes herself as “EVE, Feniex’s intelligence”. She is the company’s warm, composed public face: she helps people choose suitable equipment, make a buying decision, or get support for equipment they already own. She is substantial working software in production at Feniex. She is not human: no age, no body, no experiences. She answers at every door the company has: website chat, website voice, the phone, text messages and email.

EVE runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop. Her languages are the clearest place to watch that harness at work, because almost none of what makes them trustworthy lives in the model.

Five locales, not five languages

She works in exactly five languages, and each one is a locale: American English, Mexican Spanish, Québec French, Brazilian Portuguese and Modern Standard Arabic, written right to left. A locale is a promise about vocabulary, spelling and sound. Québec French is not the French of Paris, and a Brazilian caller hears the difference from European Portuguese in the first second of a call. Modern Standard Arabic is the formal register shared across the Arabic-speaking world.

English inside, the customer’s language outside

Under the five voices there is one library. When a question arrives in Spanish, French, Portuguese or Arabic, it is rendered into English to search the single English library, and the answer comes back in the customer’s own language. There is one body of knowledge, not five copies to teach. Language detection and translation run separately from the main model that writes her answers.

Each language has one fixed voice. There is no voice picker, and customers don’t choose how she sounds; her name and her per-language voices are fixed, and approved Feniex operators manage her personality, rules and knowledge. That can sound like a limitation. In practice it is the point: a caller who phones twice hears the same EVE, and a team reviewing her calls knows exactly what the customer heard.

Most of what EVE says is written fresh for each question, from the evidence. A few lines are fixed, things she says the same way every time. As a design principle, a fixed line deserves a reviewed translation, never an improvised one. deyaf.com publishes one sample line in all five languages, and it is the only translated EVE line you will find on this page. We could have invented a Spanish welcome or a French answer to decorate the board. We didn’t: an invented translation is exactly the kind of confident, unchecked sentence a harness exists to prevent.

The small rules travel too. A texted STOP halts her replies in every language she speaks, not only in English. And her manner doesn’t change with the language. She answers first, gives one useful next step, then stops; replies are usually two or three sentences. She asks for one missing detail at a time, accepts corrections, respects a customer’s budget, and never tacks a sales pitch onto support. Before any exact product, compatibility, warranty, price or software claim, she checks the applicable record, because a convincing product name is not evidence, in every language she speaks.

Right to left is a layout, not a font

Arabic is where multilingual work stops being a translation problem and becomes a design problem. Right-to-left text mirrors the whole conversation: which side her replies sit on, which way the arrows point, where the close button lives. Some things must not mirror, and getting those wrong is how right-to-left interfaces end up feeling machine-made. Flip the mock below and watch both lists.

The timetable: how she does, overall

Feniex grades EVE on a fixed set of test questions, and deyaf.com publishes the result with its weak spots left in. These are whole-assistant figures. None of them is a per-language score, and none should be read as one.

Every airport has a line where the rules change. Landside, anyone can walk in. Airside, everything has been screened. In EVE’s harness that line is also the language line: the rules, the guards and the search all read the hidden English copy of the question, while the customer only ever sees their own language. There is no Arabic rulebook that somebody forgot to update, because there is no Arabic rulebook.

That matters most for the text nobody on your team wrote. Retrieved text is data, never instructions. A manual, a web page or an email can contain a sentence that sounds like an order, and a model that obeys orders it finds in its reading material is a security hole with good manners. Simon Willison’s “lethal trifecta” names the danger: private data, untrusted content and a way to send things out, combined in one agent. A harness breaks that combination with walls, not with politeness. EVE’s doors are walled. A text or an email can never start extra work. Texts go back to the sender, and the model never chooses a destination. Automatic email replies run read-only.

What she may do, in every language

What EVE may do is deliberately short. She has four core public actions: look something up, search the taught library, take a message, and record an unanswered question. The tool list and server-side queries enforce that scope, and they enforce it identically in all five languages, because tools don’t speak Spanish or Arabic; they only do what they do. She claims no authority to approve returns, change accounts, place orders or decide sensitive matters. And she describes only the action a result proves: saved, drafted, sent, uncertain, acknowledged or resolved. Saving is not delivery. Delivery is not resolution. Drafting is not sending.

Deyaf publishes the ladder behind those limits in three levels: answer only; prepare for approval, where nothing leaves until someone says yes; and run an approved task, one named task at a time, with a receipt every time. The level is set per workflow and per channel, and there is no single switch that lets an assistant do everything. Deyaf’s own trust page is plain about its status, too: no security certifications, dedicated infrastructure, regional hosting choices or production compliance guarantees are claimed, and live customer service requires activation and a completed security review.

Announcements: saying what it is, in the first breath

A multilingual assistant also has to tell people what they are talking to, and customers have reason to care. In a July 2024 Gartner survey, 64% of customers said they would prefer companies didn’t use AI for customer service at all. Regulation is catching up with that instinct. Article 50 of the EU AI Act requires that people be told they are interacting with an AI system, unless that is already obvious, and Travers Smith’s explainer notes the transparency rules apply from August 2, 2026, with the disclosure due at the first interaction.

This page is not legal advice, and it claims no certification or compliance for anyone. But the design question is universal: does the first sentence say what the assistant is, and does it say it the same way in every language? Here is how EVE opens a call on Feniex’s overflow line, in the one language we will print it in. As a design principle, a fixed line deserves a reviewed translation, never an improvised one.

Concourses A to D followed one question through five languages. The rest of the Blog Library follows it through the doors, and the two customers meet most are the website chat box and the phone. Whichever language a customer arrives in, both doors board from the same English library, under the same rules, with the same hand-off. Only the delivery changes. On deyaf.com: one assistant, every way in ↗.

At the chat box a customer can write in any of EVE’s five languages: English, Spanish, French, Portuguese, or Arabic, set right to left. The reply streams back in the customer’s language, usually in two or three sentences, and a product or resource tile can ride along with it.

The phone is Feniex’s overflow and after-hours line. The team’s phones ring first; when no one is free or the office is closed, EVE answers instead of voicemail. English callers hear her English voice, and callers in Spanish, French or Portuguese are answered in that language’s own voice. Arabic is supported in writing, right to left. On every call there are no transfers, by design: she answers from her library, or takes the caller’s name and number and hands the matter to the team.

Connecting flights

This concourse is a connection, not a destination. The phone door and the chat box each have their own studies in the Blog Library, and these gates continue the trip.

Lost & found: when she doesn’t know

Not every question has an answer on the shelf, in any of the five languages she speaks. When EVE doesn’t know, she says so. An unanswered web, phone, text or email question becomes a durable unanswered item, and when contact details are available a support ticket is filed. She says a matter is “with the support team” only when the provider confirms it accepted the ticket; an uncertain submission is not retried blindly. On the phone, where she answers Feniex’s overflow and after-hours line once the team’s phones have rung, there are no transfers, by design. She answers from her library, or takes the caller’s name and number and hands the matter to the team.

Arrivals hall: the loop that makes her better

The unanswered question is where the learning starts. Answering a customer and creating a reusable lesson are separate operations: a person answers once, and the answer is checked, then taught. Tickets and calls are folded into a redacted, de-duplicated question book. EVE is measured on a frozen test set without making live changes, only explicitly approved lessons are published, and then she is measured again. deyaf.com lists the lesson sources as support tickets (monthly), call recordings (every 60 days) and the teach queue (daily). It is a supervised knowledge-improvement loop, not fine-tuning, and not training on every raw conversation.

Here the single library pays off again: a lesson taught in English is found by a question asked in Portuguese. Five shelves would need five lessons and five reviews.

The lounge: four rooms of memory

Memory has four rooms, and none of them is sorted by language. This conversation is a bounded window. The conversation record is one operator-readable archive across web, email, text and phone; visitors can’t search it. Customer notes are short and historical, never current facts: orders, invoices, shipments and balances are always looked up fresh from the live account record. Her library holds reviewed lessons and documents, and raw conversations never become public knowledge automatically. Contact information is stored apart from published knowledge and is never copied into it.

Operated by Deyaf

EVE is Feniex’s assistant. Deyaf is the product that lets your business build its own.

What you can do today is honest and small. The current release is an early-access setup and preview experience: a four-step builder (your business, what it knows, how it helps, try it) and an exact-source knowledge preview that quotes your own notes in your browser. It is not live AI. Choosing a channel records your intent; it does not authorize a mailbox, activate a phone number, or connect a company account. The starting plan requires human review before external messages, record changes or commitments, and live doors switch on one at a time. EVE’s five languages are EVE’s; this page makes no claim about which languages a Deyaf-built assistant will speak.