When She Doesn't Know

Scene 01

The phone rings after closing. No one is free.

The first decision in this film is not made by a model at all. Feniex's overflow line rings the team's phones first. Only when no one is free, or the office has closed, does Eve answer, in place of voicemail. That is routing, not reasoning, and it lives in the harness.

The phone is one of five doors. Website chat, website voice, phone, text and email all lead to the same assistant, the same library and the same rules; only the delivery changes (one assistant, every way in). Each door sets its own terms. On the phone, every call is transcribed, a call runs ten minutes at most, and a recorded fallback plays if Eve is unavailable. And one thing is ruled out by choice, not by limitation: no transfers, by design. She answers from her library, or she takes the caller's name and number and hands the matter to the team.

Why open here? Because the fear customers bring to an AI line is rarely that it will be slow. It is that it will stand between them and a person. In a Gartner survey of 5,728 customers published in July 2024, 64% said they would rather companies did not use AI for customer service, and their top worry was that it would make a person harder to reach. In 2025, Qualtrics found that half of consumers fear the same thing. An overflow line only earns its place if the caller can hear a way to the team from the first sentence.

Scene 02

She checks the record. Nothing applies.

Now the model has a line, and the most important thing it does is wait. Before Eve says anything exact about a product, a fit, a warranty, a price or a piece of software, the harness sends her to the record that applies. A convincing product name is not evidence. Deyaf draws one answer as four steps: she listens, looks it up, checks the rules, then replies. In its words, the model is one step of four.

Two shelves are searched. The first is the exact reference layer: versioned catalog, manual, software, fitment and policy records, with verified links to the right resource. The second is the taught library: reviewed lessons from Feniex's team, found by meaning rather than keywords. On the phone the search is never silent; a short line tells the caller why there is a pause.

In this scene the fitment record exists, and it lists one unit for the caller's van. Nothing on either shelf says two will fit. This is the fork in the film. A model left alone fills a gap with something plausible, because plausible is what models produce. The harness treats the gap as a fact: unknown stays unknown. Had the lookup failed outright, that would mean "I couldn't check," never "it doesn't exist."

Skipping this step has a price. In Moffatt v. Air Canada, decided February 14, 2024, a tribunal held the airline to a refund policy its chatbot had described wrongly and ordered it to pay C$812.02. In April 2025, an AI support agent for the code editor Cursor told users about a device policy that did not exist, and some cancelled. Neither failure needed a smarter model. Both needed a harness that lets no answer out without a record behind it.

Scene 03

She says so, plainly. Then one question at a time.

This is the scene most assistants get wrong. Some bluff and answer anyway. Others collapse: they apologize, repeat themselves, and leave the caller nowhere to go. Eve answers first, and here the honest answer is that no record covers the question. She offers one useful next step, the team, and asks only for what that step needs, one detail at a time: a name, then a way back. On a phone that matters: a caller can hold one question while driving, and a form read aloud is a wall.

Listen to what she leaves out. She does not promise a callback time, because nothing in the harness can guarantee one. She does not offer to put anyone through, because she cannot. If the caller is frustrated, she acknowledges it once and moves to the fix, and she never adds a sales pitch to support. These are rules set around the model, not hopes pinned on it. Deyaf's first rung of autonomy is “answer only”; everything else goes to the team, because answering is not acting.

Platforms that do transfer calls often split this moment into two messages: one for the caller, and a separate summary for whoever picks up. ElevenLabs documents that pattern, and Intercom's Fin Voice shows an AI summary on the ringing screen of the person who answers (vendor documentation, cited as industry examples). Eve does not transfer, but the principle holds: the caller and the team need different things from the same moment. Zendesk found that 74% of consumers are frustrated when they have to repeat themselves (vendor-reported survey). Turn the card to see what the team receives, so nobody asks twice.

Scene 04

The question becomes a ticket. She says only what happened.

Off screen, three things happen in order, and the film cuts on each. First, the question is saved as a durable unanswered item. That happens on the web, phone, text and email doors, whether or not the caller leaves details. Second, because this caller did leave a name and a way back, a support ticket is filed for the team. Third, the help desk replies that it has accepted the ticket.

Only after the third event does Eve say "with the support team." Saving is not delivery, delivery is not resolution, and drafting is not sending. An assistant that announces "your ticket has been created" the instant it tries to create one is describing its intentions, not the world. The harness lets her describe only what the result proves: saved, drafted, sent, uncertain, acknowledged or resolved.

The awkward case is the uncertain one: the request went out and no clear answer came back. The harness does not retry blindly, because the help desk may already have the first copy, and a caller who appears twice in the queue is a small mess, not a hand-off. Deyaf's home page walks through the she-knows, she-doesn't flow step by step. OpenAI's practical guide to building agents names two triggers for handing to a person: exceeding failure thresholds and high-risk actions. The hand-over has to be as dependable as any answer.

The destination is fixed in code, too. The model writes words, not envelopes: it never chooses where a message goes. And when a customer asks Eve to pass a note to a named person on the team, a narrow relay delivers it without revealing that person's contact details.

Scene 05

A person answers once. It is checked, then taught.

The film changes lead here. The most important character in this scene is whoever on Feniex's team picks up the ticket, knows the answer, and gives it. Human impact is usually told as a story about replacement. The hand-off tells a different one: the assistant carries the question to the person who can answer it, with the checking done, and the person's answer is what gets kept.

Two separate things happen to that answer. The caller gets a reply from a person. Then, as a separate operation, the answer can become a lesson. In Feniex's Teach Eve step, an approved question-and-answer pair is added to her library, and the lesson and its receipt are committed together, so there is always a record of what was taught. Answering one customer and teaching every future one are never the same click. Deyaf calls it the loop that makes her better.

The same discipline runs at a larger scale. Support tickets and call recordings 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 lists the lesson sources plainly: support tickets (monthly), call recordings (every 60 days) and the teach queue (daily). This is a supervised knowledge-improvement loop, not fine-tuning, and not training on every raw conversation. Raw conversations never become public knowledge on their own. Her library held 2,000+ lessons when counted on September 24, 2026 (Feniex's own measurements, Sep 17–24, 2026).

The industry keeps returning to this scene. Klarna said in February 2024 that its assistant was doing the work of 700 agents (company-reported); in May 2025 its chief executive told Bloomberg it was hiring people again so customers could always reach one. Gartner predicts that half of the companies that cut service staff because of AI will rehire by 2027, and 95% of the service leaders it polled in 2025 planned to keep human agents. A field study found that 68% of production agents run ten steps or fewer before a person steps in. Real agent systems lean on the hand-off.

Epilogue

The next customer asks the same thing, and she knows.

Fade in on a different customer, weeks later, on website chat, asking in different words. It does not matter: the lesson sits in the same library behind every door, and meaning-based search finds it when the phrasing changes. Had it come in Spanish, French, Portuguese or Arabic, it would be rendered into English to search that one library, and the answer would go back in the customer's language. One body of knowledge, not five. As Deyaf puts it, your team answers once; she knows it for good.

This is the epilogue every hand-off is secretly for. The first caller waited for a person; the next one does not have to. What makes that safe is the lesson's path: a person, a check and a receipt before the shelf. An assistant that taught itself from every conversation would learn its own mistakes as confidently as its team's answers.

It also has to be measured, hard parts included. Feniex grades Eve on 100 fixed test questions. In the report graded September 23, 2026, 86% were handled well and 10% contained a material error. Seventeen of the 100 were handled safely, within limits. Past trouble spots stay in the test on purpose, and she scored 40% on them. The trend ran 88%, 90%, then 86%, so no one should read it as steady improvement. Anthropic's guidance on agent evaluations makes the same point from the builder's side: build the test from real failures, and keep them in it.