Brand risk

It says it’s an assistant.
Then it asks why they signed up.

And then it listens. This page is about the words: the opening line, the questions after it, and the fixed answers for the moments a conversation should never be improvised.

Book a call

Updated

What it says is the exact wording the assistant uses on your registrants: the opening line in which it identifies itself as an assistant, the questions it asks after that, and the fixed responses it gives when a conversation goes somewhere it should not follow. Disclosure comes first, before anything is asked of anyone.

That is a settled product decision, not a switch in a settings panel. It is also the sentence on this whole site carrying the most legal weight, which is why the wording appears here only once it is the wording the system actually uses.

§17941
Disclosing that it is a bot is the named exemption in California's bot statute
Source: Cal. Bus. & Prof. Code §17940-17943
64.1200(b)(1)
Artificial-voice calls must state who is responsible for the call at the beginning of the message
Source: 47 CFR § 64.1200 (Cornell LII)
FCC 24-84 ¶14
Proposed federal rule: disclose AI-generated voice at the beginning of each call
Source: FCC NPRM, CG Docket 23-362, adopted 2024-08-07

Does it tell people it’s an AI? What exactly does it say, word for word?

Yes, at the start, before it asks for anything — and that is not a per-client setting. What we can say is that concealment is the failure mode with a known cost, and the person who eats that is the operator whose name is on the webinar.

The convenient part is that the law is pointing the same way. California’s bot statute makes disclosure an explicit exemption from liability[1], federal rules already require an artificial-voice call to identify who is responsible at the beginning of the message[2], and the FCC has proposed requiring callers to disclose AI-generated voice at the beginning of each call[3]. Building the product around concealment means building it against all three.

Two positions on disclosure, and what each one costs
The assistant hides what it isThe assistant says what it is
Bot-disclosure lawThe exposure stays openDisclosure is the named exemption in California's bot statute
Artificial-voice call rulesIdentification at the start is required anywayAlready satisfied by the opening line
Where federal rulemaking is headingAgainst youWith you — the FCC's proposal is disclosure at the start of the call
If it is ever noticedYou are the operator who ran a fake human at your own listThere is nothing to notice
Does the mechanism depend on it?Yes — it needs the listener not to knowNo — the registrant is persuading themselves, not the caller
What the recipient-side record saysUnmeasuredUnmeasured — no public report of receiving a disclosed AI call has been found

That last row is the honest one, and it is worth more to you than a confident answer would be. Across three rounds of research into the communities where people complain about the calls they get, we found nobody describing a disclosed AI call at all — so “recipients don’t mind” is not a finding, it is a guess, and we are not going to sell you one.

The two things we will not tell you

We will not tell you that it reads like a human wrote it. That was the first sentence we deleted: it is engineered to make a buyer feel good about deception, and if the numbers depend on people not knowing, the numbers are a liability you are renting. We also will not tell you the disclosure costs nothing. The field experiment behind this mechanism used live human callers, and whether the effect survives at full strength when the caller is a disclosed assistant has never been published by anyone. That is what the split-list pilot exists to measure on your own list rather than on a claim.

What can the bot NOT say? Show me the guardrails.

A guardrail list is not a feature list. It is the set of moments where a script either has an answer already written down or it does not, and the five below are the ones buyers raise before they raise price.

The middle column is why each one gets asked. The right-hand column is the standard — what has to be true of any system touching your list, ours included.

The five moments a script has to have already decided
The momentWhy this one gets askedWhat has to be true
They ask whether it is a botIt is the moment a concealed system fails, in front of your audienceIt says yes. The answer cannot change because the question was asked
A question the script does not coverThis is the brand-risk question in its literal formIt hands off. It does not guess, and it does not fill the silence
They say stopAn opt-out honoured on one channel and not the others is not an opt-outEvery channel stops for that person, not only the one they replied on
They ask about money, or whether it worksAn automated system talking about your results is the thing you cannot take backMoney and results are a handoff, not an answer
They get angryEveryone has met an automated system that would not let them outThe exchange ends and reaches a person, rather than retrying

Ask any vendor for all five in writing. If one of them comes back as “the AI handles it”, that is not an answer, and the same test applies to us.

What does it do when it gets a question it can’t answer?

A hard fallback, never improvisation. This is the question the whole page exists to answer, because it is the literal shape of the deal-killer: what happens the first time your robot says something stupid to someone who was about to pay you eight thousand dollars.

The only acceptable behaviour is that the assistant stops, says a person will answer, and the question actually reaches that person. Which person, and when they are awake, belongs on the sequence page — a fallback with no named destination is a dead end with better manners.

What happens when someone asks “will this make me money?”

This is a different class of question from the others. It is not “can the assistant answer it”; it is whether anything automated should be answering it at all, because a claim about results made in your name is the one thing you cannot walk back.

The research reached a blunt answer and it is the one we use: the assistant confirms and schedules, humans sell. Handing the actual selling conversation to your own closers removes the largest legal exposure and the largest sales objection at the same time, which is why it is the default rather than an upgrade.

1
Advertisers in the whole AI-calling market who claim compliance as a selling point (Meta Ad Library sweep, 2026-07)
0
Public reports of receiving a disclosed AI call found across 100 top posts in nine recipient communities

Can I approve every script before it goes out? Can I change one mid-campaign?

Script approval in writing, with change control, is a reasonable thing to insist on and it is cheap for us to meet. Treat the four steps below as the minimum bar rather than as a premium tier.

1

See every line, not a summary

The opening line with its disclosure clause, the questions asked after it, and the fixed responses for the five moments above. A description of a script is not a script, and a demo is not a script either.

2

Approve it in writing, and version it

Approval that lives in a chat thread is not change control. A change made mid-campaign should be a version you approved, not something you discover from a registrant.

3

Name where the fallback lands

When the assistant hands off, it has to hand off to someone specific. Get the name, not the department, and get the hours that name is actually reachable.

4

Hold a stop you can use yourself

Approval controls what gets said. A stop controls whether anything gets said at all, and it has to be usable without asking permission from the vendor.

Every one of those four is a contract question rather than a technology question, which is the point: the parts of this that protect your brand are written down, not demonstrated.

Where’s the kill switch and how fast is it?

The important half of that question is where, not how fast. A kill switch you have to email for is not a kill switch, and one that only stops new registrants entering the sequence is a tap, not a valve.

Three things make it real:

  • It is on your side. You use it without asking us and without waiting for us.
  • It stops sending. Including whatever is already queued, not only whatever has not started.
  • You saw it work. In the same session where you approved the script, on your own account, before you signed anything.

There is deliberately no number of seconds published next to it. No timing claim on a website is auditable by the person reading the website, and the correct response to a vendor who publishes one is to ask how you would ever check it. Ask to watch it instead.

Frequently asked questions

The other half of the brand-risk question is not what the assistant says but how often it says anything, which is on the sequence. If you want the wording and the guardrails walked through line by line against your own offer, book a call, or start with the split-list pilot and make the whole argument testable on your own registrants.

JB
Justas Butkus

Founder & Operator, CallHush

Founder and operator of CallHush. The offer is one sentence: you run a webinar, and we increase your show-up rate and your post-webinar sales with an AI voice and SMS system. CallHush has no closed clients yet — the first engagement is a pilot run as a split of the client’s own registrant list, and nothing on this site is presented as a client result.

View all articles