Nobody Has to Hold

Every business telephone has two kinds of trouble hour. One is the busy hour, when every person who could pick up is already talking to someone else. The other is the empty hour, when the office has closed and the lights are off. For most of the telephone's history, both hours ended in the same place: a recorded voice asking the caller to leave a message after the tone, and a row of blinking lights waiting for the morning.

This almanac is about the alternative. It follows one phone line through a single turn of the clock and asks, hour by hour, who answers, what the line says while it thinks, what becomes of the caller's question, and what happens when the line's own helper is unavailable. The helper is an AI assistant, but the assistant is not the interesting part. The interesting part is the Agentic Harness around it: the rules, records, voices, limits and hand-offs that decide what the assistant may do on a call and what it must hand back to people.

Our worked example is EVE, Feniex's assistant, who has answered Feniex's overflow and after-hours line since September 17, 2026. Everything we say about her comes from what Feniex and Deyaf publish, and her figures carry their dates and their caveats. The business day on our dial, however, is invented. It is an illustrative day, not Feniex's opening hours, because the point of an almanac is the shape of the seasons, not the time on one particular wall clock.

The oldest rule on this line is also the simplest: the team's phones ring first. When a customer calls Feniex, the call goes to people before it goes anywhere else. EVE is not a gate that every caller must get past. She is the hand that catches what would otherwise fall: the call that arrives while everyone who could answer is already on another call, and the call that arrives after the office has closed.

That ordering is a design decision with consequences. The assistant never competes with the team for the easy calls, and her whole working life is spent in the two hardest conditions a phone line has: the rush and the dark. A line built the other way round, where the machine answers first and a person is a reward for persistence, is exactly what callers fear. In a Gartner survey of 5,728 customers, 64% said they would prefer companies did not use AI for customer service at all, and their top worry was that it would make a person harder to reach. Qualtrics found that about half of the consumers it surveyed share the same fear. An overflow line starts from the opposite promise: a person first, and when no person can come, something better than a tone.

Better than a tone in what way? Voicemail keeps the caller's voice, but it cannot answer a question, ask for the one missing detail that would make a callback useful, or say what happens next. On EVE's line, when no one is free or the office is closed, she answers instead of voicemail, and she opens the same way every time:

“Hello, this is Eve, Feniex's intelligence. How can I help you?”

The greeting on the overflow and after-hours line

Two things in that sentence are doing work. It gives her name, and it says what she is: Feniex's intelligence, not a person pretending to be one. Europe now treats that kind of candour as a duty rather than a courtesy. Under Article 50 of the EU AI Act, which applies from August 2, 2026, systems that talk with people must be designed so that people know they are dealing with an AI, unless it is already obvious. On a phone line nothing is obvious. The greeting settles it in the first breath, before the caller has spent a word.

The dial below is the whole line in one figure. Turn it through an illustrative day, noon at the top and midnight at the bottom, and it tells you, hour by hour, whose phone rings first, who answers when nobody can, and what becomes of the call. A switch lets you rehearse the one bad night every line should plan for: the night the assistant herself is unavailable.

Notice how few kinds of outcome there are. Busy or closed, the caller reaches EVE. If EVE is down, a recorded fallback plays. There is no hour on the dial where the phone rings into nothing, and no hour where the rules quietly loosen because it is late. That regularity is the first gift of a harness: the rules are written once, outside the model, and applied the same way at three in the morning as at eleven.

The Table of Hours is what you would write down first if you were planning a line of your own, long before anyone chose a voice. The rush is where the overflow rule earns its keep, the night is where the after-hours rule does, and the outage switch is the line's insurance, written before it is needed.

On a web page, a spinning icon tells you the machine is working. On a telephone there is no icon. There is only sound, or its absence, and silence on a call means one of two things to a caller: the line has dropped, or nobody is there. Either reading ends the call.

Voice engineers talk about this as a latency budget. A recent industry write-up puts the point at which a spoken reply starts to feel slow at around 800 milliseconds, with under half a second preferred, and notes that listening and thinking consume most of that budget. Those are targets for conversation. A caller asking whether a part is covered by a warranty is not making conversation; the honest answer needs a record checked, and checking a record takes longer than a breath.

EVE's measured speed on the phone over her first week was 1.6 seconds to her first word and 3.0 seconds to a finished reply (Feniex's own measurements, Sep 17–24, 2026; the report card, weak spots included). That is quick for an answer that has been looked up, and slower than a conversational gap. So her line is built never to go silent. After about 1.5 seconds she says "One moment." When a lookup starts she says "Let me check on that." While it runs, she says "Still checking" about every 4 seconds.

None of those phrases is clever, and that is the point. A caller who hears "Let me check on that" knows three things at once: they were understood, something is being done, and the wait has a reason.

The cues are also a rule about honesty. "Let me check on that" is said when a check is actually starting; it is not a verbal tic to buy time for a guess. A line that says "checking" while the model improvises is worse than a silent one. Coupling the phrase to the lookup is something only the harness can do, because only the harness knows whether a lookup has begun. And every read the line makes has a time limit of its own, so "Still checking" is never a loop without an end.

Every call on the line ends in one of two ways. Either EVE answers from her library — the catalog, manuals, software, warranty and policy records, and the lessons the team has reviewed and taught her — or she takes the caller's name and number and hands the matter to the team. There is no third ending in which she improvises.

The first ending depends on evidence. Before any exact claim about a product, its compatibility, warranty, price or software, she checks the applicable record, and she says only what the result proves. "A convincing product name is not evidence" is how the rule is written. The cost of skipping it has a precedent. In Moffatt v. Air Canada, decided on February 14, 2024, a tribunal held the airline responsible for what its website chatbot had told a grieving customer about bereavement fares, advice that contradicted the airline's own policy page, and ordered it to pay C$812.02. The sum was small. The principle, that a company owns what its assistant says, was not.

The second ending is the hand-off, and here the line makes its most deliberate choice: no transfers, by design. EVE does not patch a caller through to a desk phone, and she never says she is connecting anyone. She takes the details, and the matter goes to the team as a durable item that cannot be lost in a shift change. When contact details are available, a support ticket is filed, and she says the matter is "with the support team" only once the ticket is confirmed as accepted. Until that confirmation comes back, the honest sentence is a smaller one, and she says the smaller one.

Why refuse transfers, when many phone systems offer them? Our reading, not a measurement, is simple arithmetic. On an overflow line, a transfer lands on the same busy team the call just overflowed from: in the middle of a rush, every person who could take a transferred call is the reason the call came to EVE in the first place. After closing there is no one to transfer to at all. A transfer on this line would be a promise of a person that the hour cannot keep. A taken message with a filed ticket is a promise it can.

The hand-off also spares the caller the worst part of calling back. Zendesk's CX Trends 2026 report (vendor-reported) found 74% of consumers frustrated at having to repeat information. Every call on EVE's line is transcribed, and the conversation record is one archive that operators can read across phone, chat, text and email, so the person who follows up can begin where the caller stopped rather than at the beginning.

Two more things frame every call. The first is a limit: ten minutes at most, a fence set in advance so that no single conversation can hold the line indefinitely. The second is the transcript, which is how the line learns. Calls, like tickets, 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; then she is measured again. It is a supervised loop that a team runs on purpose, the loop that makes her better, and not a machine quietly absorbing every conversation it hears.

A phone line does not choose who calls it. EVE speaks five languages: English, Spanish, French, Portuguese and Arabic. Callers in Spanish, French or Portuguese are answered in that language's own voice, one fixed voice per language, rather than an accent painted over the English one.

What does not multiply is the knowledge. A question in French is rendered into English to search the one English library, and the answer returns in French. There is one body of knowledge, not five copies drifting apart season by season. Detection and translation run separately from the main answering model, so the language is settled before the answer is written. The rule in short: English inside, the customer's language outside.

We have shown only the one line deyaf.com prints in every language, and we have left it at that on purpose. An almanac that made up its foreign weather would not deserve the trust of its local readers.

Almanacs print records so that next year's readers can tell a good season from a bad one. These are EVE's, from her first week on the phone. They are Feniex's own measurements, machine-judged and provisional, and the one success rate in the table is printed beside the error rate that belongs with it.

As of September 24, 2026 · Feniex's internal Eve 3.0 report · machine-judged and provisional. Read it in full at EVE's report card, weak spots included.

Read the table the way a farmer reads a rainfall record: as the conditions of one week, not a promise about the next. One hundred and two calls is a first week's harvest, not a yearly average. The 93% of callers who stayed and talked is the figure that speaks to the old fear of an assistant on the line. The 60 handed to the team speaks to the other fear: 60 of the 102 calls ended with the caller's details taken and the matter in the team's hands, never a transfer.

The report-card figure needs the most care. It comes from a fixed set of 100 test questions rather than from the calls themselves, and its 86% "handled well" only means something beside the 10% that contained a material error. One answer in ten on that test was wrong in a way that mattered. That is why, on this line, the hand-off matters as much as the answer, and why a figure is never printed here without its date, its source and its weak spot.

Who is on the line? EVE is Feniex's assistant, warm, composed and not human. She helps people choose suitable equipment, make a buying decision, or get support for equipment they own. The same EVE answers website chat, website voice, text and email; the phone is one door among several, with the same identity, library and rules behind each. On a call she answers first, gives one useful next step, then stops. She asks for one missing detail at a time, and she never reads a web address aloud, because a string of slashes and dots is no use to a listener.

None of that behaviour lives in the model alone. EVE runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop. It is worth being exact about the term, so here is the definition this almanac uses. An Agentic Harness is everything around the AI model: what the assistant knows, what it remembers, where it meets customers, what it may do, and how your team teaches it. (How an Agentic Harness works, end to end.) LangChain's shorthand for the same idea is "Agent = Model + Harness". On a phone line, the harness is by far the larger term.

Lay the almanac over that definition and every chapter turns out to be a harness part. The rule of the first ring and the after-hours rule are where it meets customers. The library and the evidence rule are what it knows. The transcript and the conversation record are what it remembers. No transfers, the ten-minute limit, the hand-off and the fallback are what it may do. The question book is how the team teaches it. Even the hold beat belongs to the harness rather than the model, because it is tied to whether a lookup has actually begun. The model writes the sentences. The harness decides which sentences are allowed to be said, and when.

Anthropic's engineering guide to building agents names customer support as one of the most natural places for them, because a support conversation pairs talk with actions such as looking up records, and because success can be measured. The telephone adds the hardest condition of all: it happens in real time, out loud, with no screen to show where an answer came from. So the rules have to live in the harness, where they can be checked, rather than in anyone's hopes for the model.

Almanacs end with instructions: what to sow, and when. If you are planning an answering line of your own, here is the order in which we would plant it. It is this almanac's editorial guide, drawn from the line described above, not a Deyaf procedure.

1. Decide who rings first. Write down that people answer first, and name the exact conditions under which the assistant takes the call: every line busy, or the office closed. Never let the assistant become a gate.
2. Write the greeting once. Name the assistant and say what it is in the first sentence. Keep it the same at every hour.
3. Stock the shelf before the season. An assistant can only answer from what it has: records it can check, and lessons your team has reviewed.
4. Plan the silences. Decide what the line says while it looks things up, and never let it say "checking" unless it is.
5. Choose honest endings. Decide which matters are answered and which become a message and a ticket, and when the words "with the team" may be said.
6. Set the fences. A maximum call length, and a list of things the assistant may never approve.
7. Plan for the dark. A recorded fallback for the night the assistant is down, written in daylight.
8. Keep the record. Transcribe calls, fold them into reviewed lessons, and measure before and after anything changes.

None of this retires a team. Gartner reported in 2025 that 95% of the customer-service leaders it polled plan to keep human agents, and predicted that half of the organizations planning big AI-driven headcount cuts in service would abandon those plans by 2027. The overflow model assumes the people stay: they are, after all, the first ring. What changes is the hour that used to end in a tone.

Deyaf is built from EVE: the harness that runs EVE at Feniex, packaged so another business can have an assistant of its own. Deyaf's early-access builder lets your business set up its own; live doors switch on one at a time. Choosing a channel there records your intent. It does not activate a phone number or connect a company account.

Does EVE answer every call to Feniex?A reader who likes to talk to people

No. The team's phones ring first. EVE answers only when no one is free or the office is closed, and then she answers instead of voicemail.

Can she put me through to a person?A reader in a hurry

No transfers, by design. She answers from her library, or takes your name and number and hands the matter to the team. She says it is "with the support team" only once the ticket has been confirmed as accepted.

What will I hear while she looks something up?A reader who dislikes silence

About 1.5 seconds in, "One moment." When a lookup starts, "Let me check on that." While it runs, "Still checking," about every 4 seconds.

What happens if she is not working?A cautious reader

A recorded fallback plays. The rule is written in advance, so an outage is planned for rather than improvised.

Is the call kept, and how long can it last?A reader who reads the small print

Every call is transcribed, and a call lasts at most 10 minutes. Transcripts help the team follow up without asking you to start again, and, reviewed and redacted, they help the team teach her.

Are the hours on the dial Feniex's hours?A reader with a watch

No. The dial runs an illustrative business day, invented to show the rules. We do not publish Feniex's opening hours here.

Are these figures independent?A sceptical reader

No. They are Feniex's own, machine-judged and provisional, dated September 24, 2026, and the success rate is always shown beside the 10% material-error rate. See the full report card, with its trouble spots.

Can I have a line like this today?A reader with a busy office

Deyaf's current release is an early-access setup and preview experience; live doors switch on one at a time. Deyaf sets out what you get today, and what switches on later.