In February 2025, a federal court in California refused to dismiss a wiretapping case against Google over an AI voice agent that answered calls inside other companies' contact centers. The plaintiffs were not Google customers. They had called their own bank and their own mobile carrier. The claim that survived was that the AI vendor itself was a third party listening in on a conversation it was never invited to.
That ruling is why "is an AI receptionist legal?" now has a longer answer than it did two years ago. The technology is lawful in every US state. The failure is almost never the technology — it is a greeting written by someone who assumed there was one rule to satisfy. There are three, and most deployments get at least one of them wrong before the first call is answered.
Almost every mistake I see comes from collapsing recording consent, AI disclosure, and outbound calling consent into a single "consent" checkbox. They are separate obligations, and satisfying one does nothing for the other two.
| Obligation | What it governs | Triggered by | Practical fix |
|---|---|---|---|
| Recording consent | Whether audio or a transcript may be captured | State wiretap statutes — about a dozen are all-party | Spoken notice on every call, before capture starts |
| AI disclosure | Whether the caller must be told they are speaking to a machine | A handful of state bot laws, plus the EU AI Act | Say it in the first sentence, in every state |
| Calling consent (TCPA) | Whether a business may place an automated or AI-voiced call at all | Outbound calls, callbacks, and AI text follow-ups | Written consent before marketing; answering inbound is outside it |
Read that table the other way and the most useful fact in this article falls out: the TCPA governs calls you place, not calls you answer. The FCC's rules target outbound automated and artificial-voice calling; answering an inbound call is a different activity. Most "AI receptionist compliance" pages skip that distinction and scare business owners into buying written-consent software they do not need.
The federal floor is permissive. Under the Wiretap Act, 18 U.S.C. § 2511(2)(d), a party to a conversation may record it. Your business is a party to its own calls. If that were the whole picture, no notice would ever be required.
It is not. Roughly a dozen states require all parties to consent before a call is recorded, and several more are mixed or unsettled. The working list: California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Oregon, Pennsylvania, and Washington — with Michigan and Oregon frequently argued either way.
Two are traps for anyone reading statutes instead of case law:
So think about the Ohio contractor with a 513 area code taking a call from a snowbird with a Pennsylvania cell number. The strictest rule on the line generally wins, and number portability killed area codes as a jurisdiction proxy years ago. The only defensible default is one notice on every call, written for the strictest state you serve.
It requires a spoken notice at the start of the call that sits inside the recording — you announce, the caller continues, and the audio you keep contains the announcement. Massachusetts draws the line at secrecy rather than paperwork: a recording made with the knowledge of all parties is not an interception, even without an affirmative yes.
Which produces the architectural mistake that voids more "compliant" deployments than any other: capturing audio before the disclosure finishes playing. If your stack streams from the first ring, transcribes immediately, and plays the notice two seconds later, then "by staying on the line you consent" is false for those two seconds — and those two seconds are captured evidence. The sequence has to be answer, notice, then capture.
The notice is not a formality appended to the call. It is the first thing that happens on it, and nothing is retained until it has finished.
Here is where the search results get loud and wrong. There is no final federal requirement that an AI voice agent announce itself. The FCC's 2024 declaratory ruling, FCC 24-17, settled that AI-generated voices count as "artificial" voices under the TCPA — but that ruling is about outbound calling consent, not a conversational disclosure. The follow-up proposal that would have required an in-call AI disclosure still had no final rule as of late 2026.
State law moved instead, unevenly. Maine requires disclosure when an AI could pass for a human. Utah requires it when a caller asks, and up front for licensed professions in high-risk interactions. Colorado attached a disclosure duty to high-risk AI systems beginning in mid-2026 and is actively rewriting that statute, so treat it as a moving target. California's AB 2905 covers automatic dialing-announcing devices and artificial voices, which reaches outbound dialers more than inbound answering. And the EU AI Act, Article 50, applies from 2 August 2026: people must be told they are interacting with an AI unless it is obvious. Take calls from Europe and that is your floor now.
None of that is a reason to stay silent. It is a reason to stop treating the disclosure as a legal checkbox and start treating it as a trust decision. A caller who works out mid-sentence that they are talking to a machine does not become a lead. They hang up and they remember.
The honest operational reading: one sentence, every call, every state. It costs four seconds and removes an entire category of argument — including the argument a plaintiff's lawyer makes when they claim the caller was deceived into disclosing account details to software.
That is the shape of Ambriz v. Google. The plaintiffs did not allege the AI misheard them. They alleged the vendor behind the AI was an uninvited third party on a confidential call, and the claim proceeded because the system had the capability to use the data for the vendor's benefit. Notice that the business using the contact center was not the defendant. If you are an agency deploying voice agents on client numbers, read that twice. You are the vendor in that story.
For calls a business places, 47 C.F.R. § 64.1200 is the operative text. Informational calls using an artificial or prerecorded voice need prior express consent. Marketing calls to a wireless or residential number need prior express written consent that names the seller, covers artificial voice, and states that consent is not a condition of purchase. Statutory damages run $500 per call, up to $1,500 for wilful violations, with no cap on class size.
Three details sink well-meaning operators:
None of this touches the calls your receptionist answers at 9 p.m. on a Sunday — which is why reception is the safest place to start with voice AI. The caller initiated contact, so the consent question that dominates outbound never arises.
You do not need three disclosures. You need one sentence, delivered before anything is captured, that does three jobs:
"Thanks for calling Ridgeline Plumbing. This call is recorded and you're speaking with an AI assistant. How can I help?"
Business name, recording notice, AI identity, then the question. If you serve more than one language, the notice plays in the caller's selected language before capture begins, not after. If a caller objects to being recorded, the agent confirms, stops recording and transcription, keeps the call going, and logs the decline with a timestamp. An all-party consent the caller cannot actually refuse is not consent.
What to log per call, as separate events rather than one boolean: the exact disclosure text and its version number, the timestamps the notice started and finished, whether the caller continued or objected, the resulting recording state, and — for outbound — the consent record you relied on plus the DNC scrub date.
When a complaint lands eighteen months later, "recording-notice event at 00:06, wording version 4.2, caller continued, recording state on" is a defence. "consent: true" is not.
Every white-label voice deployment has the same structural problem: one prompt builder, one greeting template, many client brands in many states. Change the shared greeting and you have changed it for every tenant at once — efficient right up until the day it is catastrophic.
It also sells. "Compliant by default, with an audit trail per call" closes deals that "we have AI answering" does not, especially with insurance, legal, and healthcare-adjacent clients whose own counsel has already frightened them. For the operating model behind that pitch, see our breakdown of the white-label AI receptionist platform and the regulated-vertical version of this problem.
Answer all six and you are ahead of most funded companies in this category. Fail the first and nothing else on the list matters yet.
A short list, because fear is doing more damage in this market than the law is:
This is operational guidance from people who deploy phone agents, not legal advice. Statutes change and courts reinterpret them — the all-party list itself is genuinely debated. Have a lawyer read your greeting once. Then stop rewriting it every time a competitor's blog scares you.
Yes, in every state. No statute prohibits an AI from answering a business phone. What varies is the notice you must give — recording consent, and in a few states an AI disclosure — plus the separate rules that apply if the same system ever places outbound calls.
There is no final federal requirement as of late 2026. Maine, Utah, and Colorado impose state duties in defined situations, California's dialer rule covers artificial voices, and the EU AI Act's Article 50 has applied since August 2026. Because the caller's location is not reliably knowable, the practical answer is to say it on every call.
The commonly cited list is California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Oregon, Pennsylvania, and Washington. Michigan and Oregon are argued either way, and Nevada's statute reads one-party while its courts require all-party for phone calls. Verify against current statute before relying on it.
Not to the answering itself. The TCPA's artificial-voice restrictions govern calls a business places, and inbound calls are outside that prohibition because the consumer initiated contact. The moment your agent returns a call, sends a marketing text, or runs a campaign, you are back inside the framework.
Everything above compresses into about fourteen seconds of audio. That is the whole compliance surface for most service businesses: a notice that plays before capture, an AI disclosure a caller can hear without straining, a way to decline, and a log proving all three happened.
Nida ships voice agents with the disclosure and recording notice as a locked first step rather than a prompt instruction, configured per workspace, with the call log stored beside the lead so the record and the transcript never drift apart. You can see how the pieces fit on the pricing page, or start with the cost and ROI math for contractors if you are still deciding whether to answer the phone at all. If you are weighing underlying voice platforms, our Vapi, Retell, and Bland comparison covers what each gives you on disclosure configuration.
Start with one number and one greeting. Call it from a cell phone in an all-party state and listen to the first five seconds. If the notice is not there before you start talking, you have found your first fix — and it is a fifteen-minute fix, not a legal project.
← Back to the blog