AI disclosure is the question nobody settles before launch. Somewhere in your company, software is already writing to customers. It drafts the follow-up, picks the send time, decides who gets which message, and increasingly it answers the phone. Whether you tell anyone is the part with real consequences.
This is not a legal guide. Rules on automated communication vary by state and by channel, and they change. This is about the practical question underneath the legal one, which is what you owe the person on the other end and what happens to trust when you get it wrong.
Summary
- Disclosure and consent are different problems. Consent is mostly settled law. Disclosure is mostly a trust decision you are making on your own.
- On email, the useful line is not whether AI wrote the draft but whether a human is accountable for the claim.
- On voice, disclosure is harder, because a caller who feels deceived reacts far more strongly than a reader who suspects a template.
- Most disclosure fails not because it is absent but because it is written defensively, which makes an ordinary practice sound like a confession.
- Data handling is the part teams forget. What the model retains matters more than what the message admits.
Table of contents
- What AI disclosure actually means
- Where this is a legal question and where it is a trust question
- Email, and the accountability test
- Voice, the harder case
- Writing disclosure that does not kill the conversation
- Data handling, the part people skip
- Key takeaways
- AI disclosure FAQ
- The bottom line
What AI disclosure actually means
Disclosure is telling the person you are communicating with that AI is involved. That sounds simple until you try to write the policy, because AI involvement covers a huge range.
A spell checker is AI by most definitions. So is the model that picked the send time. So is the system that drafted the whole message. So is the voice on the phone that sounds like a person and is not one.
Nobody thinks the spell checker needs a notice. Most people think the voice does. The interesting work is in the middle, and the way to get through it is to stop asking how much AI was used and start asking a different question. Would this person feel misled if they found out exactly how this message was produced?
That question is answerable. It also tracks fairly closely to where the law is heading, which is a convenient accident rather than a coincidence.
Where this is a legal question and where it is a trust question
Two things get tangled together, and they behave differently.
Consent is largely settled. If you are making automated or prerecorded calls or sending marketing texts in the United States, there are rules about permission, about when you may call, and about honoring an opt-out. Those are obligations, not preferences, and the relevant federal rules sit in 47 CFR 64.1200. We cover how that applies to AI calling specifically in our guide to TCPA compliance for AI outbound calling.
Disclosure is mostly not settled. Outside specific state rules and specific contexts, there is often no statute telling you to announce that a message was AI-assisted. Which means you are making a brand decision, and you should make it deliberately rather than by default.
The trap is treating an unsettled question as a settled one in your favor. Absence of a requirement is not the same as absence of a consequence. The consequence just arrives as a customer who feels tricked rather than as a fine.
Email, and the accountability test
Email is the easier channel, because readers have assumed a machine was involved for twenty years. Nobody believes the newsletter was hand written.
So disclosure on every AI-assisted email is not useful. It would be noise, and it would train people to ignore the notice in the cases that matter.
The line that holds up is accountability rather than authorship. Ask who is answerable for the claim in this message. If a person reviewed it and stands behind it, the fact that a model produced the first draft is a process detail. If nothing was reviewed and the message asserts something specific about price, availability, results or someone’s account, you have a problem, and the problem is not disclosure. It is that you are making unreviewed claims.
Two cases do deserve explicit handling. A message that appears to be personal correspondence from a named individual should actually involve that individual. And an automated reply that could be mistaken for a human response to a specific complaint should say it is automated, because the alternative is a customer waiting for a reply that already came.
Voice, the harder case
Voice is different, and the difference is not subtle.
Modern voice agents pass as humans on a short call. That is a product achievement and an ethical exposure at the same time. When someone discovers mid-conversation that they have been talking to software, the reaction is rarely mild. People feel embarrassed, which converts quickly to anger, and the anger attaches to your brand rather than to the technology.
There is also an asymmetry worth noticing. A reader who suspects a template shrugs. A caller who realizes they were fooled tells other people.
Practically, the teams that handle this best do one of three things.
They disclose up front, briefly, and move on. This costs less than people fear when the phrasing is confident.
Or they disclose on request and train the agent to answer the question honestly and immediately. If a caller asks whether they are speaking to a real person, the agent says no. Any design where the agent deflects that question is a decision to deceive, whatever the intention was.
Or they use a voice that does not attempt to pass, which sidesteps the problem and costs some warmth.
What does not work is refusing to decide, which in practice means the agent evades and a caller eventually posts the recording. Our own platform leaves the choice with the customer and puts the wording in the script you approve, which is covered in our legal and compliance overview.
Writing disclosure that does not kill the conversation
Most disclosure reads badly because it is written by someone worried about liability. Hedged, passive and apologetic phrasing signals that something is wrong even when nothing is.
Compare these. “Please be advised that this call may be conducted using automated artificial intelligence technology” sounds like a warning. “Hi, this is Ava, I am an AI assistant with Bigly Sales” sounds like an introduction.
The second one is more honest and performs better, for the same reason. Confidence reads as normal. If you sound like you are admitting something, people conclude there was something to admit.
Three rules that hold up. Say it early rather than when challenged. Say it in one short sentence. Then continue as though it is unremarkable, because it is.
Data handling, the part people skip
Disclosure debates focus on the message. The more consequential question is what happens to the conversation afterward.
If customer conversations pass through a third party model, ask three things and get the answers in writing. Is this data used to train the provider’s models. How long is it retained. Who can access it internally.
For regulated conversations there is a fourth. Can recordings and transcripts be produced on demand for your compliance team, and for how long.
A team can be scrupulous about announcing the AI and still be quietly sending customer health or financial details into a system with unclear retention. The disclosure question is visible. This one is not, and it is the one that turns into a real incident.
Key takeaways
- Consent is a legal obligation. Disclosure is mostly a brand decision, so make it deliberately.
- The workable test is whether the person would feel misled if they knew exactly how the message was produced.
- On email, ask who is accountable for the claim rather than who wrote the draft.
- On voice, decide your position in advance and never let an agent dodge a direct question about being human.
- Write disclosure as an introduction, not a warning. Defensive phrasing creates the suspicion it was meant to avoid.
- Get retention, training use and access answers in writing before customer conversations reach a third party model.
AI disclosure FAQ
What is AI disclosure?
AI disclosure means telling the person you are communicating with that artificial intelligence is involved, whether that is a drafted email, an automated reply or a voice agent on a call. It is separate from consent, which governs whether you were permitted to contact them at all.
Do I legally have to disclose that a caller is talking to AI?
It depends on where you are calling and why. Consent, calling hours and opt out handling are governed by federal and state rules, while explicit AI disclosure requirements vary by state and by context and are still developing. Treat the absence of a clear rule as a decision you have to make rather than permission to say nothing, and get advice for the states you actually operate in.
Is AI disclosure the same as consent?
No, and conflating them causes most of the confusion. Consent is permission to contact someone and is largely settled law. Disclosure is telling them how the message was produced and is largely a trust decision.
Should every AI assisted email include a disclosure?
No. Readers have assumed automation in email for decades, and a notice on every message becomes noise that trains people to ignore it when it matters. Focus instead on whether a human is accountable for the claims the message makes.
What should an AI voice agent say if someone asks whether it is human?
It should say no, immediately and plainly. Any design where the agent deflects that question is a choice to deceive, and it is the single fastest way to turn a routine call into a complaint people share.
Does disclosing AI hurt conversion rates?
Less than teams expect when the wording is confident and brief. Most of the measured damage comes from disclosure written defensively, which makes an ordinary practice sound like an admission. An introduction performs differently from a warning.
What is the difference between disclosure and transparency?
Disclosure is the specific act of saying AI is involved in this interaction. Transparency is broader and covers what data you collect, how long you keep it, whether it trains a model and who can see it. You can disclose well and still be far from transparent.
What should we ask an AI vendor about data handling?
Whether your conversations are used to train their models, how long data is retained, who internally can access it, and whether recordings and transcripts can be produced on demand for your compliance team. Get the answers in the contract rather than from a sales call.
Who should own the AI disclosure policy internally?
It needs a named owner, usually shared between whoever owns compliance and whoever owns brand, because it is simultaneously a legal exposure and a trust decision. Policies with no owner default to whatever the implementing team decided that week.
The bottom line
The companies that handle this well are not the ones with the longest notice. They are the ones that decided in advance, wrote it in one plain sentence, told their agents to answer honestly when asked, and got their retention terms in writing.
The ones that struggle treated silence as the safe option, which it is right up until the moment somebody finds out.
If you are deciding how to handle disclosure on live calls, the wording sits in the script you approve rather than in our defaults. Book a demo and we will walk through how other teams in your industry have written it.
