The guest in the corner
Almost every business website now keeps a small guest in its bottom right corner. It waves. It bounces. Sometimes it opens itself before you have read a sentence, introduces itself with a first name and a stock smile, and asks for your email address. Nobody would put up with that from a person behind a shop counter, and a growing pile of evidence says customers don't put up with it from software either.
In a Gartner survey of 5,728 customers, 64% said they would prefer that companies didn't use AI for customer service at all, and the worry at the top of their list was that it would make reaching a person harder (Gartner, 9 July 2024). A Qualtrics study of more than 20,000 consumers found that nearly one in five who used AI for customer service got no benefit from it, about four times the failure rate of AI use in general, and half feared it would keep them from a human (Qualtrics, 7 October 2025).
Those numbers are not a verdict on the technology. They are a verdict on its manners.
This zine is about those manners: the small, specific habits that make an AI chat box feel like good help instead of a pop-up with opinions. Most of them are old shop-floor courtesy, written down for software. Say who you are. Don't hover. Answer the question that was asked. Show your working. Don't sell to someone who came to you with a problem. And when you can't help, say so, and hand the person to someone who can, without pretending the hand-off is something it isn't.
Here is the twist. Manners are not a tone-of-voice setting. You can't write "be polite" into a prompt and hope. A chat box that says it checked the warranty needs something that actually checked the warranty. A box that promises a person needs a real route to one, and a way to know whether the message arrived. The part of the system that makes those promises true is the Agentic Harness, and it is why this is a zine about harnesses rather than about adjectives.
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 ↗
Good manners are the outcome. The harness is how you get them, reliably, on the hundredth conversation as well as the first. The customers in those surveys want the same few things, over and over:
would prefer companies didn't use AI for customer service. Gartner survey of 5,728 customers · 9 Jul 2024
who used AI for service saw no benefit; half feared it blocks a human. Qualtrics, 20,000+ consumers · 7 Oct 2025
expect an explanation for decisions made by AI. Zendesk CX Trends 2026 · 18 Nov 2025 · vendor-reported
Our house example throughout is Eve, Feniex's assistant, who answers Feniex's customers in website chat, on the phone, by text and by email. Her written habits are the model for the rules that follow. Where we quote her numbers, we print the date, the source and the error rate right beside them.
Nine house rules
Say it's AI, up front
The first thing a chat box says should make one fact impossible to miss: this is software. Not a friendly first name with a stock photo, not a "team member" who types suspiciously fast. A short label in the box's header and a plain line in its greeting do the job: "I'm an AI assistant."
In Europe this is now law, not just courtesy. Article 50 of the EU AI Act requires that people be told they are interacting with an AI system, at the latest at their first interaction, unless that is obvious from the context. Those transparency duties apply from 2 August 2026, and fines can reach €15 million or 3% of worldwide annual turnover (EU AI Act, Article 50; Travers Smith explainer). That is our summary, not legal advice; your lawyer should read the article, not our zine.
The practical case is just as strong. In April 2025 the support bot for the code editor Cursor, answering under the name "Sam", explained a logout bug by inventing a one-device-per-subscription policy. Users took it as company policy and some cancelled before a co-founder corrected it in public (AI Incident Database #1039). A clearly labelled bot can still be wrong, but people read it as a bot, and they check.
Eve's version is plain. She never pretends to be a person, and on the phone she opens with "Hello, this is Eve, Feniex's intelligence. How can I help you?"
Further readingwhat the assistant can and cannot doGreet briefly
A greeting is a doorway, not a brochure. One line that says what the box is and what it can help with is enough: "Hi, I'm this site's AI assistant. I can help with products, manuals and orders." Then the box waits. It does not queue three more bubbles, a carousel of suggested questions and a satisfaction survey.
Long greetings fail for a simple reason: the visitor arrived with a question already half-formed. Every sentence you put in front of it is a sentence they have to read past. A brief greeting also sets the rhythm for everything after it. If the first message is two lines long, the visitor learns that this box is going to be quick, and they write back the same way.
The greeting should be the same wherever the assistant lives. A customer who met it in chat and later calls should meet the same identity, the same knowledge and the same rules, not a second personality with its own ideas. That consistency is a harness job: one identity, one library and one set of rules behind every door.
Further readingone assistant, every way inDon't pop up on every page
The launcher, that little button in the corner, is an offer. Opening the panel on its own is a demand for attention, and it is the rudest thing chat boxes routinely do. It covers the content the visitor came to read, it can steal keyboard focus from someone using a screen reader, and on a phone it can take over the whole screen.
The polite default is simple: the box stays closed until someone asks. The exceptions are narrow. If a visitor shows a real sign of being stuck, such as searching twice with no result or coming back to the same troubleshooting article, one quiet line beside the launcher, once per visit and easy to dismiss, is fair. Checkout pages get nothing but a visible launcher that keeps clear of the pay button. Try your own pages in the chooser a few pages on.
A well-mannered box
You're chatting with an AI assistant. It answers from this site's manuals and policies.
- A label that says AI, and a close button that works with Escape as well as a click.
- A greeting of one line, saying what the box can help with. Then it waits.
- One missing detail, asked once, with a hint about where to find it.
- The answer first, then its source, as a tile the visitor can open. The box says what the record proves and no more.
- An honest way to a person, visible from the start and named for what it really does.
Say less, show more
Answer first, then stop
The most useful sentence in any reply is the answer, so it goes first. Not "Great question!", not a restatement of what was asked, not three sentences of background. The answer, one useful next step, and then silence.
This is exactly how Eve is written to talk. She answers first, gives one useful next step, then stops, and her replies usually run two or three sentences. She respects a customer's budget and recommends the simpler option when the simpler option is enough.
Stopping is the hard part. A language model left to its own devices will keep going, adding caveats and alternatives until the answer is buried in the middle of the fourth paragraph. A harness keeps replies short the way a good editor does: with written rules, and then by checking what actually went out against them, rather than hoping the model feels concise today.
Further readingfour steps in about three secondsShow where the answer came from
In Zendesk's CX Trends 2026 research, 95% of consumers said they expect an explanation for decisions made by AI (Zendesk, 18 November 2025, vendor-reported). The simplest explanation a chat box can give is its source: the manual page, the warranty policy or the product record, linked right under the answer.
Showing a source only works if there is a real one. In February 2024 a Canadian tribunal ordered Air Canada to pay C$812.02 after its website chatbot told a grieving passenger he could claim a bereavement fare after travelling, which contradicted the airline's own policy page. The airline argued the chatbot was responsible for its own words. The tribunal disagreed and held the company responsible for the information on its website (McCarthy Tétrault on Moffatt v. Air Canada). A box that had to point at the policy page could not have said what it said.
Eve checks the applicable record before any exact product, compatibility, warranty, price or software claim, because a convincing product name is not evidence. In website chat she can place a product or resource tile under the answer, so the visitor can open the very thing she used. And she says only what the result proves. If a lookup fails, the honest reply is "I couldn't check that", never "that doesn't exist".
Further readingthe model is one step of fourPop-up or not?
Pick a kind of page and what the visitor is doing. The chooser says whether the chat box should open itself, and why. It encodes this zine's house rule: the box opens when it is asked, offers one quiet line when there is a real sign of trouble, and otherwise waits.
Should the box open itself? No.
A first visit is the worst moment to interrupt: they haven't seen the page yet. The launcher in the corner is the offer.
| Page | Just arrived | Reading a while | Signs of being stuck | Pressed "Ask a question" |
|---|---|---|---|---|
| Home page | Stays closed | Stays closed | One quiet line, once | Opens: they asked |
| Product page | Stays closed | Stays closed | One quiet line, once | Opens: they asked |
| Help article or manual | Stays closed | Stays closed | One quiet line, once | Opens: they asked |
| Order status page | Stays closed | Stays closed | One quiet line, once | Opens: they asked |
| Checkout | Stays closed | Stays closed | Stays closed | Opens beside the form, never over it |
| Blog article | Stays closed | Stays closed | One quiet line, once | Opens: they asked |
One thing at a time
Ask for one detail at a time
When a chat box needs details, the rude way is a form in disguise: "Please provide your order number, email, full name, postcode, product and a description of the issue." The visitor has to scroll back, assemble a list and hope they got every format right.
Asking for one missing detail at a time is kinder and more accurate: the order number, and then, only if it is still needed, the email on the order. It lets the visitor fix one wrong answer instead of re-sending the whole list, and it means the box never asks for something it already has. In the same Zendesk research, 74% of consumers said they are frustrated by having to repeat information, and 81% expect a conversation to carry on without backtracking (vendor-reported).
Memory is manners here. The harness keeps this conversation in a bounded window, so "the one I mentioned earlier" still means something three messages later. Eve asks for one missing detail at a time, and she accepts corrections without a fuss.
Further readingfour kinds of memory, each with one jobNever sell into a support answer
Someone with a broken product is not a lead. Tacking "Have you seen our new model?" onto a troubleshooting answer tells them the box cares more about the next sale than about this problem. Eve's rule is flat: she never adds a sales pitch to support. She acknowledges frustration once, then fixes what she can.
There is a sharper reason to keep selling out of a chat box's reach: it must not make commitments the business hasn't made. In December 2023 a user told a Chevrolet dealership's chatbot to agree with anything the customer said. It then agreed to sell a 2024 Tahoe for one dollar and called the offer legally binding (AI Incident Database #622). The dealership did not honour it, and the bot came down.
The harness lesson is short: answering is not acting. Prices, discounts, orders and exceptions go through records and rules that the business controls, never through whatever a model can be talked into typing.
Further readinganswering is not actingRewrite it!
Five rude chat-box messages, every one invented for this zine, though you have probably met their cousins. Press Rewrite it on each (or tab to it and press Enter) to see the harness-mannered version and the house rule it follows. The scorecard keeps count.
No. 1 · Illustrative example · a store's home page, three seconds after arrival
Pops open, uninvited:Hey there! I'm Jessica, your personal shopping buddy! Before you go anywhere, pop your email in for an exclusive deal!!
Wrong with it: opens itself · poses as a person · asks for an email before helping.
Rewritten, waiting behind the launcher:Hi, I'm this site's AI assistant. I can help with products, manuals and orders. Ask whenever you like.
Rules 1 & 2: say it's AI; greet brieflyNo. 2 · Illustrative example · a yes-or-no question
Visitor asked: "Does the Model 40 take the standard mounting kit?"Great question! Choosing the right mounting kit is a really important decision, and there are lots of factors to consider. First, let's look at the history of our mounting systems… (four more paragraphs)
Wrong with it: buries the answer · never stops.
Rewritten:Yes, the Model 40 takes the standard mounting kit. Want the install page from the manual?
Rule 4: answer first, then stopNo. 3 · Illustrative example · a warranty question
Visitor asked: "What's the warranty on the Model 40?"Absolutely! All of our products come with a lifetime warranty, no questions asked!
Wrong with it: invents a policy nobody checked · nothing to point at.
Rewritten, with its source:The Model 40 is covered for two years from purchase, according to the warranty policy, section 2. Here's that page.
Rule 5: show where the answer came fromNo. 4 · Illustrative example · checking on an order
Visitor asked: "Where's my order?"To help you I need: full name, email, phone, order number, postcode, date of purchase, product serial and a detailed description of your issue.
Wrong with it: a form pretending to be a conversation.
Rewritten:I can check that. What's the order number? It's at the top of your confirmation email.
Rule 6: one detail at a timeNo. 5 · Illustrative example · a product that stopped working
Visitor wrote: "My charger stopped working."So sorry! While you're here, our new Pro charger is on sale this week! Connecting you to a live agent now… please hold.
Wrong with it: sells into a support answer · promises a connection that isn't coming.
Rewritten:Sorry about that. The manual's reset steps fix most charging faults: hold the button for ten seconds. If it still won't charge, I can take your details so the team can follow up.
Rules 7 & 8: no pitch; an honest hand-offRules restored
- 1 · Say it's AI, up front
- 2 · Greet briefly
- 4 · Answer first, then stop
- 5 · Show where the answer came from
- 6 · One detail at a time
- 7 · Never sell into support
- 8 · Honest route to a person
Show them the door
Make "talk to a person" easy, and honest
The fear behind that Gartner figure is specific: that the AI stands between the customer and a human. So the route to a person should be visible from the first screen, work on the first request, and describe itself truthfully. Gartner's own poll of 163 customer service leaders found 95% planned to keep human agents (Gartner, 10 June 2025); the people are there, and the box should lead to them.
Truthfully is where most boxes fail. "Connecting you to an agent now…" followed by a spinning wheel and nothing is worse than no offer at all. In January 2024 the parcel company DPD switched off the AI part of its chat after, following a system update, it swore at a customer and wrote a poem about how useless the company was. The customer had been trying to reach a person (ITV News, 19 January 2024).
For Eve, what really happens is a hand-off, not a live connection. She takes the person's name and contact details and hands the matter to the Feniex team. On the phone the rule is the same: no transfers, by design. Every unanswered question becomes a durable item, and when contact details are available a support ticket is filed. She says "with the support team" only after the ticket system confirms it accepted the ticket. Step through it on the insert overleaf.
Further readingwhat happens when she doesn't knowMind the box itself
Manners include the widget's own behaviour, and for many visitors that behaviour is the whole experience. The launcher is a real button with a name a screen reader can announce, such as "Ask a question", not just "button". The panel can be opened, used and closed with the keyboard alone: focus moves into it when it opens, Escape closes it, and focus returns to the launcher afterwards.
New messages are announced through a polite live region, so a screen reader doesn't talk over someone who is still typing. The panel never covers the page's own controls, text can be zoomed without clipping, and the colours pass contrast. Typing dots, bouncing launchers and sliding panels calm down for anyone who has asked their device for reduced motion.
None of this is exotic. It is ordinary web accessibility applied to a small window that happens to talk. But a chat box that can't be closed with the keyboard is a box that traps people, and no amount of polite wording fixes that.
Cut out and keep
- Say it's AI, up front.
- Greet in one line, then wait.
- Stay closed until asked.
- Answer first, then stop.
- Show the source.
- Ask one thing at a time.
- Never sell into support.
- Make the way to a person easy and true.
- Keyboard, screen reader, focus, reduced motion.
The honest hand-off
A visitor asks for a person. Step through what a well-mannered box says, one message at a time, and choose what the ticket system does at the end. Watch the status card: the visitor should never have to guess what state the assistant is in.
- Can I talk to someone about returning this?
- I can't approve returns myself, so I'll pass this to the team. What name should they use?Says what will really happen: a hand-off, not a live connection.
- Dana.
- Thanks, Dana. What's the best phone number or email for their reply?One missing detail at a time.
- [gives an email address]
- Filing a ticket for the team…Saving is not delivery. Nothing is promised yet.
- If the ticket system confirms it:Done. It's with the support team, and they'll reply to the email you gave.
- If the ticket system doesn't answer:I couldn't confirm the team has it yet, so I won't tell you it's with them. Your question is saved on our side, not lost.
House style: how Eve minds hers
Every rule in this zine has a working example. Eve is Feniex's assistant, a warm, composed public face for a business built on a catalog, manuals, software, warranties and a dealer network. She helps people choose suitable equipment, make a buying decision, or get support for equipment they already own. Eve runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop.
She meets customers in website chat, where answers stream in with product and resource tiles; in website voice; on the phone; and by text and email. It is one Eve at every door, with the same identity, the same library and the same rules. That matters for manners: a customer treated well in chat should not meet a different personality on the phone.
Her manners are written down, not left to mood. She answers first, gives one useful next step, then stops. She asks for one missing detail at a time and accepts corrections. She checks the exact record before any product, compatibility, warranty, price or software claim, and says only what the result proves. She acknowledges frustration once and never adds a sales pitch to support. She never pretends to be a person, and on a call she never reads web addresses aloud, because nobody can use a link they only heard.
Her first week on the phone
On the phone, Eve 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. 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.
As of September 24, 2026 · Feniex's internal Eve 3.0 report · machine-judged and provisional. Eve's report card, weak spots included ↗
We print the 10% beside the 86% on purpose. A zine about manners can't hide its own mistakes, and neither should an assistant's scoreboard. These are Feniex's own measurements, not an independent benchmark.
When she doesn't know
The polite answer to a question she can't settle is not a guess. The question becomes a durable unanswered item; with contact details a ticket is filed, and "with the support team" waits for confirmation. Answering the customer and teaching her are separate jobs: a person approves each new lesson before it joins her library, in a supervised loop rather than training on every raw conversation.
Further readingyour team answers once; she knows it for goodDear Zine
Should our box pop open on every page?
Dear Zine, our marketing lead wants the chat box to open itself on every page because it "drives engagement". I think it's annoying. Who's right? Popped Out
You are. An auto-opened panel counts every dismissal as a conversation, so the engagement it drives is mostly people closing it. Keep it closed until asked, allow one quiet line when someone is plainly stuck, and judge the box on questions answered, not panels opened. Start with one good job it does well.
Further readingstart with one good jobDo we really have to say it's AI?
Dear Zine, our bot is good enough to pass for a person. Do we really have to label it? Passing in Plain Sight
Yes. For visitors in the EU it is the law from 2 August 2026, and everywhere it is the thing that keeps a wrong answer from reading like company policy. Being good enough to pass is a reason to be proud of the label, not to hide it. On the phone, Eve's very first line says who she is, and she never pretends to be a person.
Further readingDeyaf's trust and control principlesHow do we stop the five-paragraph answers?
Dear Zine, every reply from our bot is an essay. Scrolling in Spokane
Write the rule down (answer first, one next step, stop), then measure against it on a fixed set of real questions, and re-measure after every change. When a person fixes a bad answer, make the fix a reviewed lesson so the same essay doesn't come back next week.
Further readingthe loop that makes her betterCan the box just connect people live?
Dear Zine, customers keep asking the bot for a human. Can't it just put them through? On Hold Forever
It can if someone is staffed to pick up, and then the box should say exactly that. If nobody is, don't imply otherwise. Eve doesn't connect anyone live; she takes the details, hands the matter to the team, and tells the person what she did. A truthful hand-off beats a pretend connection every time.
Further readingEve on the phoneWhat should the box remember about me?
Dear Zine, I don't want to repeat myself, but I don't want to be tracked either. Forget Me Not
Good manners remember this conversation, within limits, and treat older notes as history, never as current facts. Anything that changes, like an order's status, should be looked up fresh from the live record rather than recalled from memory. Contact details belong apart from anything the assistant publishes.
Further readingwhat an assistant should and shouldn't rememberHow do I know it's polite outside the demo?
Dear Zine, it behaves in the demo. How do I know it behaves on a Tuesday? Demo Day Nerves
Test it on fixed questions that include the hard ones, and publish the misses with the hits. Feniex's own report card for Eve shows 86% handled well beside a 10% material-error rate, as of September 24, 2026, from Feniex's internal Eve 3.0 report, machine-judged and provisional.
Further readingEve's numbers, measured not promisedWhere do I start with one of my own?
Dear Zine, I want a well-mannered box for my own business. Ready to Print
Start with one door and one job, and switch on each door deliberately. Deyaf's builder lets you describe your business, add what it knows and try a knowledge preview in your browser; the preview is not live AI, and live channels are switched on one at a time.
Further readinga front desk that knows Further readingswitch on each door deliberatelyAnswers by the Zine. Letters are invented for this issue.
A chat box with good manners is not a nicer prompt. It is a system that knows where its answers come from, knows what it may do, and knows when to hand over.
That system is the Agentic Harness. Eve is Feniex's working proof; Deyaf is how another business builds its own.
Good for one well-mannered assistant
Describe your business, add what it knows, and try it in your browser. About ten minutes to a saved preview, no account needed.
Build your assistantThe builder's preview is a knowledge preview, not live AI. Live doors switch on one at a time.