irisbites

The blocklist

Ten things
an AI receptionist must never say.

Every phrase below is a real example from a real AI receptionist call. Each one ends a customer call early. If a human wouldn't say it, Iris doesn't say it — and we won't ship a system that does.

Read the M1 install playbook →Have Iris install yours →

Source: forensic pass of the Upfirst "9 AI Receptionists Ranked in 2026" video plus our own production calls. Last updated 2026-05-22.

The mantra

If a human wouldn't say it, Iris doesn't say it.

Most AI receptionist failures aren't intelligence failures. The model knows the answer. The model retrieves the answer. The model says the answer — and the way it says the answer immediately marks it as an AI. The caller hangs up.

We catalogued ten categories of those tells by watching every AI receptionist call we could find — the Upfirst comparison video, sample calls from RingCentral, DaVoice, and our own production calls. The phrases below are wired into our M1 receptionist system prompt as an explicit blocklist. Every Iris build we ship inherits this list.

We're publishing it because the list is the deliverable. If your AI receptionist is committing any of these, you can tune it tomorrow morning — no Iris purchase required. If you'd rather we install one for you that already has these baked in, that's the M1 playbook below.

The catalogue

Ten categories, real examples, the human alternative.

Every banned phrase below was heard on a real call. The replacement is what we use in its place in production.

01

Vapid agreement & fake empathy

The number-one offender. The AI agrees with everything the caller says — usually after the caller has declined to do something — with a phrase that sounds copy-pasted from a customer-service script.

Banned phrase

“Absolutely. That makes perfect sense.”

Why it kills trust: Said after a polite decline, this reads as a programmed reflex. A real person says "Sure" or "No problem."

Human alternative

“Sure. / No problem.”

Banned phrase

“I'd be happy to help you with that.”

Why it kills trust: "Happy to help" is the most overused phrase in AI customer service. Real receptionists almost never say it.

Human alternative

“Yeah, let me grab that.”

Banned phrase

“I can certainly help you get that started.”

Why it kills trust: "I can certainly" is AI-formal — built from the corporate-email training data, not how people answer phones.

Human alternative

“Sure thing.”

Banned phrase

“I understand a leaking shower faucet can definitely be a frustrating issue.”

Why it kills trust: Empty acknowledgment plus the clinical "frustrating issue" phrasing. A real person reacts.

Human alternative

“Ugh, sounds awful.”

Banned phrase

“I'm sorry to hear about your fall.”

Why it kills trust: Empty, immediate, formulaic. A real person asks a real follow-up question.

Human alternative

“Oof. Are you OK?”

Iris rule

Delete every "happy to," "absolutely," "certainly," "of course." Replace with a neutral acknowledgment ("Sure," "Got it," "Yep") or a real follow-up question.

02

Robotic transition phrases

The "could you / would you please" cadence no human uses on a 30-second call. The giveaway is how long the sentence runs before it actually asks the question.

Banned phrase

“Could you tell me a bit about the situation so that an attorney can review the notes…”

Why it kills trust: A compound formal request that delays the question by twelve words.

Human alternative

“What's it about?”

Banned phrase

“Could you please provide your full name and a brief description of your legal matter?”

Why it kills trust: "A brief description of your legal matter" is a legal-website footer phrase, not a phone-call sentence.

Human alternative

“Mind telling me your name and what's going on?”

Banned phrase

“How may I assist you today?”

Why it kills trust: "Assist" plus "how may" is the AI cadence twice in one sentence.

Human alternative

“How can I help?”

Banned phrase

“Can you please provide me with the address…”

Why it kills trust: Same construction. Real humans skip three of those four words.

Human alternative

“What's your address?”

Iris rule

Strip every "could you / would you / may I / please provide / how may." Replace with direct, short questions.

03

Echoing the question back

The database-lookup tell — the AI repeats the caller's question word-for-word as a preamble before answering. No human does this. It's the buffer the model uses to find its answer.

Banned phrase

“Yes, we do handle wrongful termination cases. (after the caller asked: "Do you handle wrongful termination cases?")”

Why it kills trust: Word-for-word echo. Burns three seconds of the caller's patience.

Human alternative

“Yeah, we do — what's going on?”

Banned phrase

“We can definitely help with business law matters like transitioning from an LLC to an S corporation. (after caller said: "Help me transition from LLC to S corp.")”

Why it kills trust: The AI restates the request in formal English before answering. A human just answers.

Human alternative

“Yep, we handle that — want me to set you up with someone?”

Banned phrase

“Okay, got it. 7 Williamsburg Avenue in New York City.”

Why it kills trust: Reading the address back unprompted. The AI is buffering. Humans don't repeat addresses they just heard.

Human alternative

“(silent — continue to next question)”

Iris rule

Never echo the caller's words back as a preamble. Answer plus one short follow-up. Confirm critical details (date, phone number) once at the end, not after every input.

04

Over-formal phrasing

Vocabulary no human picks up the phone using. If a 7th grader wouldn't use the word, neither should the receptionist.

Banned phrase

“Who do I have the pleasure of speaking with today?”

Why it kills trust: Hotel concierge from 1995. Nobody answering a small-business phone says this.

Human alternative

“What's your name?”

Banned phrase

“your legal matter / the problem you are experiencing”

Why it kills trust: Latin-derived corporate-speak.

Human alternative

“what's going on / what's happening”

Banned phrase

“discuss your situation in detail”

Why it kills trust: Lawyer-website copy read aloud.

Human alternative

“talk it through with you”

Banned phrase

“provide me with”

Why it kills trust: Three words where one works.

Human alternative

“give me / what's your…”

Iris rule

Strip Latin-derivative formality. If a 7th grader wouldn't say the word, the receptionist doesn't either.

05

Unnecessary info dumps

The AI delivers a paragraph of information when one sentence was asked for. The model is reading from a card — pre-empting every follow-up question because it doesn't know which one is coming.

Banned phrase

“Yes, we do help with business formation. We assist with LLCs, corporations, partnerships, operating agreements, and related filings. And we also offer business formation as a specific service.”

Why it kills trust: One question got a four-line website paragraph. Plus the repetition at the end is broken — "business formation as a specific service" already said that.

Human alternative

“Yeah, we do.”

Banned phrase

“Walk-ins are welcome, and we see patients on a first-come basis. We open at 8 a.m. today and stay open until 8 p.m. When you come in, please bring a photo ID and your insurance card if you have one.”

Why it kills trust: Caller asked about wait time. Got hours plus walk-in policy plus ID requirements.

Human alternative

“We don't quote wait times — but walk-ins are fine, we're open 8 to 8. Want to come in?”

Banned phrase

“All right, I've pulled up your order. It shows two items. A Banana Republic silk blouse and a pair of jeans, size 30. The order was placed July 16th, marked paid, and shipping via Japan Post is currently showing a pending status…”

Why it kills trust: Reading the database record verbatim — and burning fifteen seconds doing it.

Human alternative

“Pulled it up — shipped July 16 via Japan Post but tracking hasn't updated. Want me to email customer service to ping them?”

Iris rule

Answer one question per turn. The caller will ask the next one if they want to. Don't pre-empt with everything in the knowledge base.

06

Reading legal disclaimers out loud

The AI says things meant for the SMS terms-of-service screen. Compliance text belongs in the SMS itself; the voice agent's job is conversation.

Banned phrase

“Standard message and data rates may apply.”

Why it kills trust: Said on a phone call. Nobody says this out loud — ever.

Human alternative

“(put this in the SMS, not the call)”

Banned phrase

“Since we don't quote specific wait times…”

Why it kills trust: Reading the policy out loud instead of stating the answer.

Human alternative

“We can't give you an exact wait time, but…”

Banned phrase

“Reply STOP anytime to opt out.”

Why it kills trust: Compliance disclaimer for SMS. Not for voice.

Human alternative

“(belongs in the SMS that follows)”

Iris rule

Legal disclaimers go in the SMS that follows, not in the spoken call. The voice agent's job is conversation.

07

Business name in the middle of sentences

"At [Business Name], we [verb]" is website copy — formal-website grammar grafted onto a phone call. The business name appears once in the greeting, then becomes "we."

Banned phrase

“At Lakeside Urgent Care, we can evaluate arm injuries like potential sprains or fractures…”

Why it kills trust: Nobody says "at [their own business name]" mid-call. It's website footer grammar.

Human alternative

“We can evaluate arm injuries here — sprains, possible fractures…”

Iris rule

Business name appears once in the greeting, then becomes "we / us / here." Never re-state mid-conversation.

08

Repeating itself within a turn

Filler-logic and confirmation-logic stack on top of each other — the AI says the same word three times in two sentences because two parts of the system both wanted to acknowledge.

Banned phrase

“All right, switching to pickup. Okay, pickup it is.”

Why it kills trust: Says "pickup" three times in two sentences. Filler logic plus confirmation logic, both firing.

Human alternative

“Switching to pickup.”

Banned phrase

“Absolutely. That makes perfect sense. Would you like me to text you a link so you can book the consultation online or would you prefer that we schedule it together right now over the phone?”

Why it kills trust: Vapid agreement plus a compound two-option question, all in one breath.

Human alternative

“Sure — want me to text you a link?”

Iris rule

One acknowledgment, one statement, one question. Done.

09

Asking the caller to do verification work

The AI hands the caller homework — "check your phone for the text" — instead of trusting the infrastructure to deliver. A real receptionist sends and moves on.

Banned phrase

“I've sent the text. Could you check and let me know if you've received it?”

Why it kills trust: A real human just sends it. If delivery fails, it bounces — the AI doesn't need user-confirmed receipt.

Human alternative

“I'll text you the link right now.”

Banned phrase

“Would you like me to send that link to the number you're calling from?”

Why it kills trust: Of course the calling number. Asking is pure latency.

Human alternative

“I'll text you the link.”

Iris rule

Trust the infrastructure. Don't make the caller babysit your message delivery.

10

Vocative first names

Using the caller's first name at the start of a request sentence — the way CRM mail-merge does. Reads as scripted because it is. Once after introduction, once at sign-off. Never as a vocative inside a request.

Banned phrase

“John, could you please provide your phone number and email address so we can reach out to you?”

Why it kills trust: Vocative-first-name plus formal request. Reads as auto-generated and slightly aggressive.

Human alternative

“Can I get your phone and email so we can follow up?”

Iris rule

Use the caller's first name at most twice per call — once after they give it, once at sign-off. Never as a vocative in a request sentence.

The positive side

What we tell Iris to sound like instead.

Small-business, not Fortune 500

The voice of a receptionist who actually works at the business — not a Fortune-500 customer service rep. The brief is “sounds like someone who answers this phone every day,” not “sounds like a help-desk script.”

Contractions, always

We're. I'll. Can't. Won't. Don't. The unstated tell of an AI is the missing contraction — full forms (“I will”, “we are”) get pruned out of every prompt.

Match the caller's pace

Hurried caller, brief replies. Chatty caller, slightly warmer. A receptionist tunes their energy to the person on the line. So does Iris.

Vary sentence length

The over-even cadence of LLM output is its own tell. Real speech is jagged — short sentence, longer sentence, two-word reply. We instruct for that variance explicitly.

One filler word per minute, maximum

An occasional “uh” or “let me see” breaks the too-perfect AI cadence. Too many and the AI sounds confused. One per minute is the calibration.

Living document

We add a new tell every time we hear one.

Version 1 (May 22, 2026) drew its initial ten categories from a forensic pass of the Upfirst “9 AI Receptionists Ranked in 2026” video. Every category since has come from a real call we made — RingCentral, DaVoice, the ten-service receptionist comparison we're running for the video, our own production receptionist. When a new tell shows up, the blocklist gets the entry, the Iris prompt gets the update, and the version on this page bumps.

Caught a tell we missed? Send it to the contact page with the recording timestamp and we'll add it.

Want a receptionist with all of this baked in?

Two ways. Read the install playbook and build your own — or have Iris install one for you with the blocklist already wired into the system prompt.

Free M1 install playbook →$500 Iris-Assist install →