It's 8:12 on a Wednesday and the compliance lead has a headset on, not a statute book. Last night's after-hours file from the collections line opens on an account number. The recording line is missing. The transcript starts the same way. The agent said it after the consumer started talking, so it is not in the first seconds of the file anyone would ask to hear.
Call recording consent is not the FCC's AI-voice disclosure rule, and it is not solved by telling the caller they are speaking to an AI. If you capture the audio, or turn that capture into a transcript, the people on the line have to be told before the capture is the record. Two-party consent is the nickname for a shorter, stricter map, and the lists do not all match. The chat widget has a paired question, not the same statute: the transcript is the record, the visitor should know it is an AI, and you should be able to say what is stored.
The TCPA file is already written in TCPA + AI Voice in 2026. When and how to say the agent is an AI is the 2026 trust playbook. This piece is the recording and transcript map, plus a one-month audit of the first seconds.
Two-party means all-party, and only a short list is verified
As it stands, federal law is the floor. Under 18 U.S.C. § 2511, interception is lawful when you are a party, or one party has given prior consent, unless the purpose is criminal or tortious. The business on the call can be that party. Justia's 50-state survey, last reviewed in September 2024, notes that a violation of the federal prohibition generally carries up to five years, with a civil action under section 2520.
We will not print a fifty-state survey or a clean count of eleven or twelve. A few statutes are genuinely contested, or one-party on the phone, and those states stay off the verified list.
The states Justia labels all-party, without that hedge, are California, Florida, Illinois, Maryland, Massachusetts, New Hampshire, Pennsylvania, and Washington. The Reporters Committee's primary all-party list includes that same core.
- California. Penal Code section 632 makes it a crime to intentionally record a confidential communication, including by telephone, without the consent of all parties. A violation is a fine not exceeding $2,500, or county jail not exceeding one year, or state prison, or both. Section 632.7 is a separate cellphone and cordless rule that does not turn on confidentiality. In Kearney v. Salomon Smith Barney (July 13, 2006), the California Supreme Court applied section 632 to calls a Georgia office recorded with California clients.
- Florida, Illinois, Maryland, Massachusetts, New Hampshire, and Pennsylvania. Justia treats each as all-party. Florida Statutes section 934.03 requires prior consent of all parties. Illinois, 720 ILCS 5/14-2, generally requires all other parties' consent before a participant surreptitiously records a private conversation. Pennsylvania's exception, 18 Pa.C.S. § 5704, is prior consent of all parties.
- Washington. RCW 9.73.030 requires all participants' consent before a private telephone communication is recorded. Consent is considered obtained when one party announces, in any reasonably effective manner, that the communication is about to be recorded. If the conversation is recorded, that announcement has to be on the recording.
Connecticut and Nevada are phone-specific, not members of a fake total. Connecticut's criminal offense is one-party, but General Statutes section 52-570d generally creates civil liability unless all parties consent at the start, including a verbal notice on the recording. Nevada requires all-party consent for wire communications, which Justia reads as phone calls after Lane v. Allstate. Montana section 45-8-213 bars a hidden recording without the knowledge of all parties, and does not use the word consent.
Justia notes that courts disagree about which state's law controls on an interstate call. Assume the stricter law can be the one a claimant picks. If you do not know where the caller is sitting, play the announcement on every call.
What the voice agent has to say
A working line is "This call is being recorded and transcribed." It has to be early, audible, and, on a Washington reading, part of the file. Say the agent is an AI in the same opening if your script already does. The trust playbook owns that sentence. Do not let it replace the recording sentence. TCPA is about whether you may place an artificial or prerecorded voice call. Recording consent is about whether you may keep the audio and the words. Inbound and outbound are different TCPA postures. They are the same recording posture if you hit record.
Justia says that, depending on the state, staying on the line after a clear warning can be implied consent. If the recording begins at answer, the announcement has to begin at answer. "First seconds" is the operational test, not a statutory stopwatch. A transcript is not a workaround. If the agent transcribes the live call, treat that as recording it.
On a Verlingo voice agent, every call can be logged with a recording, a transcript, and an outcome. That is what this audit opens. HIPAA mode stores no recordings, no transcripts, and no call logs on the platform. Do not script a recording you are not keeping. Voice coverage is 25+ languages. The phone does not speak the chat widget's language count.
The chat transcript is the other record
There is no sourced two-party recording statute for a website chat widget. Do not import California or Washington onto the bubble because "record" and "transcript" rhyme. Some California courts have treated internet chats as presumptively not confidential unless the facts say otherwise. That is not a phone all-party rule, and it is not a reason to skip the disclosure.
On chat, the transcript is the record. A Verlingo chat agent hands off with the full transcript, which your retention policy has to own. Chat covers 100+ languages. The disclosure belongs before a name, an account number, or a date of birth. Check three facts: the widget said so before the transcript filled, you know what is kept after the session, and the handoff packet is the full transcript. Phone notice in the first seconds of the audio. Chat notice before the first stored message.
A one-month audit of the first seconds
Pull every voice call the agent handled in the prior 30 days that has a recording. Skip the calls QA already loved.
- Open the audio, not the summary. Pass only if the recording-and-transcript announcement is intelligible in the first seconds, before an account number, a date of birth, or a balance dispute.
- Confirm the announcement is on the recording, not only in a prompt the dialer claims it played. If the transcript starts before the notice, the capture started too early.
- Split inbound from outbound so one missing greeting cannot hide in a blended score.
- Count the share with the announcement in the first seconds, the share where the consumer spoke first, and the share with no recording. HIPAA mode should be zero stored recordings, transcripts, and call logs on the platform.
- Sample chat from the same month. The AI disclosure, and a line that the conversation is saved, should appear before the first user message. The handoff should carry the full transcript.
Write the misses down as script defects. Fix the greeting, replay ten calls the next day, and keep the sample until the first-seconds rate is boring. On a collections line, run it on the inbound number on the letter before you touch outbound.
Start small, measure it
Start with the inbound voice line and the chat widget already in production. Put the recording sentence in the first seconds of every voice recording, keep it on the file, and put the matching line on the widget before anything is stored. Measure for one month. If that rate is not where you would want it when someone asks for last Tuesday's call, the greeting is still wrong.
When the sample is clean, expand. This is field notes, not legal advice. Take the statute list to counsel before you treat it as your map. Then start the experiment, and let the first seconds of the recording make the argument.