If a customer says we recorded them without consent: preserve the audio, find what the greeting actually said, check one timestamp, then reply. The Office of the Australian Information Commissioner asks complainants to come to you first and allow 30 days, so there is room to get this right. General information, not legal advice.
This is the incident-response path only. The law itself is set out jurisdiction by jurisdiction in Australian call recording laws for AI agents.
What probably happened on that call
Two documented behaviours sit next to each other on an agentic voice stack, and this complaint lives in the gap.
Behaviour one: recording starts on a call event, not on a disclosure event. Twilio documents the record attribute on <Dial> with a default of do-not-record, and describes record-from-answer as the option that “will start the recording as soon as the call is answered” (Twilio, TwiML Dial, read 18 September 2026). LiveKit documents the same shape at the room layer: “Auto egress is a separate configuration path. It starts recording automatically when a room is created” (LiveKit, Egress overview). Neither event is “the disclosure finished playing”.
Behaviour two: the greeting is interruptible, on purpose. Twilio documents welcomeGreetingInterruptible, which “controls what interruptions from the caller are permitted during the welcome greeting”, with a default of any; interruptSensitivity defaults to high, “the most responsive and easiest to trigger” (Twilio, ConversationRelay). LiveKit Agents states that “the user can interrupt the agent at any time” (LiveKit, turn detection and interruptions).
Put the two together and the consequence is structural, not a vendor defect: where a deployment starts recording at answer and leaves the greeting interruptible, caller audio exists before the disclosure sentence has finished. We are not asserting that any named product recorded your caller before your disclosure, because we have not seen your call; we are saying those two documented behaviours, combined in the ordinary way, produce that result by construction.
The first-second rule: if the recording start timestamp is earlier than the moment the disclosure finished playing, the complaint is factually right about the first second of audio, whatever the legal answer turns out to be. Establish that before arguing anything else.
Is this as urgent as it feels?
Usually less urgent than it feels at 11pm, and we are not going to invent a penalty clock: no published national countdown starts merely because a customer says this. What is real and published is a small set of windows you can read at source.
- 30 days, and generally 12 months. The OAIC tells complainants to complain to the organisation first and that “we think 30 days is a reasonable time” to respond (OAIC, complain to an organisation or agency); on a separate page it states that “generally, a complaint must be about a matter that occurred less than 12 months ago” (OAIC, what you can complain about). That is your working window, not a fine.
- 13 months, then 40 days. Twilio: “you are able to retrieve resources via GET to the /Calls endpoint for 13 months after the resource is created” (Twilio, Call resource), and on a separate page that recording metadata is “kept after deletion for a retention period of 40 days” (Twilio, Recording resource).
- Your own retention policy. Vapi states that artifacts “are automatically cleaned up based on your retention policy” (Vapi, call recording), and that
artifactPlan.recordingEnableddefaults totrue. The clock most likely to destroy your evidence is one you set yourself.
It is genuinely urgent in three cases: the audio sits inside a retention window shorter than the complaint cycle; the recording has left your control; or the call involved health or other sensitive information. Otherwise this is a day of careful work, not an emergency.
The next 24 hours: the four-artefact hour
Before you reply and before you form a view, put four things in one folder, read-only, and suspend any retention job that would delete them. Call it the four-artefact hour. It is free and does not involve us.
| Artefact | What it proves | Where it lives |
|---|---|---|
| 1. The audio and its start timestamp | When capture actually began | Twilio: start_time on the Recordings resource, or RecordingStartTime in the recording status callback, with Trim defaulting to do-not-trim so leading silence survives. LiveKit: the egress output file. Vapi: call.artifact.recording. |
| 2. The greeting rendered on that call | What was said, not what the template says today | Twilio: the welcomeGreeting value in the TwiML actually served. LiveKit and Vapi: the session transcript. |
| 3. Whether the caller cut in, and where | Whether the disclosure completed | LiveKit truncates history “to include only the portion of the speech that the user heard before interruption”. Twilio: reportInputDuringAgentSpeech, which has defaulted to none since May 2025, so those events exist only if you enabled it. Vapi: transcript turn boundaries. |
| 4. The consent artefact | Whether consent exists independently of the audio | Vapi Enterprise returns compliance.recordingConsent with a grantedAt timestamp in the end-of-call report. On most stacks there is no such field and the answer is whatever your CRM stored. |
Then, in order: acknowledge the complainant in writing today, saying what you are checking and by when; do not delete the recording, even if that is what they asked for, because it destroys the evidence of what the greeting said; check the timestamp; do not decide the law tonight.
The arithmetic that tells you how big the gap is
Time the rendered greeting rather than guessing: count the words in the disclosure, divide by your text-to-speech read rate in words per minute, multiply by 60. With example inputs to replace with your own, 34 words at 165 words per minute is 34 ÷ 165 × 60 = 12.4 seconds, all of it in the file before the sentence ends.
Who actually owns this complaint, and it is usually not us
Australian recording law is state and territory law and differs materially between jurisdictions. Which Act applies depends on where the parties were when they spoke, so there is no single national answer to give this customer; the jurisdiction table is on our Australian call recording laws page. We re-read the Queensland provision at source on 18 September 2026: section 43(1) of the Invasion of Privacy Act 1971 (Qld) makes a person “guilty of an offence against this Act if the person uses a listening device to overhear, record, monitor or listen to a private conversation”, liable “on conviction on indictment to a maximum penalty of 40 penalty units or imprisonment for 2 years”. Section 43(2) lists the cases where subsection (1) does not apply, the first being “(a) where the person using the listening device is a party to the private conversation”, with further carve-outs at (b) to (e) for unintentional hearing by telephone, certain Commonwealth officers, authorised police use, and government network radio in listed duress situations (reprint current as at 1 July 2024). It creates an offence, so that route runs through police and a lawyer, not a regulator with an online form.
| What the complainant is actually claiming | Who owns it | What you do |
|---|---|---|
| You mishandled my personal information | OAIC, under the Privacy Act 1988 and the Australian Privacy Principles | Respond yourself within 30 days. The OAIC expects them to come to you first. |
| You broke the recording law in my state | The state or territory surveillance devices or listening devices Act: an offence provision, not a complaints scheme | Get a lawyer. Establish where each party was before anyone opines. |
| My business turns over under $3 million | Possibly nobody federally: the OAIC states the Privacy Act “covers organisations with an annual turnover of more than $3 million, and some other organisations” | Do not relax. The state recording Act has no turnover threshold. |
| It was health information and I am a patient | OAIC plus state law: the Privacy Act “applies to all private sector health service providers anywhere in Australia”, and in NSW, Victoria and the ACT those providers “must comply with both Australian and state or territory privacy laws when handling health information” (OAIC, 1 December 2025) | Treat it as sensitive information; escalate. |
| You are a contractor to a state agency | The state body named on that same OAIC page: the NSW Information and Privacy Commission, OVIC, the Queensland Office of the Information Commissioner, the Tasmanian Ombudsman, and the NT and ACT equivalents | Check the contract. It usually binds you to the state regime. |
If the recording also left your control, run an OAIC Notifiable Data Breaches assessment in parallel. If the complaint is really about the call being made at all, that is a different instrument, covered in inferred consent under the Spam Act.
The next 7 days
Reproduce it, scope it, then decide. Place a test call, speak over the greeting, and compare the recording start timestamp against the transcript point where the disclosure was cut. If those do not line up, the configuration produced the complaint, not that one call.
Then count it, and be precise about the scope: over a stated window, on a stated set of numbers or assistants, how many inbound calls had first caller speech before the disclosure token finished. That is a per-configuration count, not a company-wide rate, and reporting it as the latter is how a correct count becomes a wrong conclusion. We publish no figure of our own, having measured none on a disclosed window.
Finish the week by writing the response, deciding on deletion once the complaint closes, and checking where the audio landed. If the file crossed a border you did not intend, that is a second question, covered hop by hop in AI voice agent data residency.
The fix that stops it recurring
One sentence: make the disclosure uninterruptible, and start recording after it finishes. Every parameter below is on the vendor page already linked.
| Stack | Parameter | Documented default | Set it to |
|---|---|---|---|
| Twilio ConversationRelay | welcomeGreetingInterruptible |
any |
none, or dtmf to keep a key press as the escape hatch |
| Twilio ConversationRelay | interruptible, interruptSensitivity |
any, high |
Leave both. Barge-in is a feature everywhere except across the disclosure. |
| Twilio Voice | <Dial record> |
do-not-record; record-from-answer starts at answer |
Start recording by API after the disclosure, so RecordingSource reads StartCallRecordingAPI |
| LiveKit Agents | turn_handling.interruption.enabled |
Interruptions enabled; adaptive mode on LiveKit Cloud | false on the disclosure say() only, not on the session |
| LiveKit | Auto egress in CreateRoom |
Starts at room creation | An explicit StartEgress after the disclosure |
| Vapi | compliancePlan.recordingConsentPlan |
artifactPlan.recordingEnabled defaults to true |
The consent plan, where “call recording only begins after consent is granted” — Vapi documents it as an Enterprise feature; on other plans, a squad whose consent assistant sets recordingEnabled false |
The trade-off is real: a 12-second disclosure that cannot be skipped will cost you abandons. The lever is length, not interruptibility. Opening lines that survive that pressure are in our AI agent disclosure script post.
Doing this yourself costs roughly a day of engineering time per stack, plus an ongoing job: every greeting change, new assistant and migration re-opens the same gap, and none of the three vendor configuration pages linked above documents a build-time or deploy-time check that fails when the disclosure becomes interruptible again (read 18 September 2026). That maintenance, not the fix, is what makes the complaint recur. Zian AI supports private model deployment on customer infrastructure, which keeps the recording inside the customer environment; the same audit questions apply to any vendor (auditing a voice platform for leaked keys and PII).
Frequently asked questions
Do I have to tell people my AI agent is recording the call in Australia?
It depends on where the parties were. Recording is governed by each state and territory surveillance devices or listening devices Act, and those Acts differ materially, so there is no single national rule. A spoken notice at the start of the call is how most operators establish consent, rather than a duty every Act imposes.
Does a caller cutting in over the greeting make the recording unlawful?
Not by itself. It makes the complaint factually accurate about the first second of audio, which is a different thing from unlawful. That turns on the Act applying where the parties were and the exceptions inside it. Treat the timestamp as a fact and the legality as a question for a lawyer.
Should I delete the recording a customer is complaining about?
Not before you have preserved a copy and answered them. If the person asked for deletion, acknowledge the request, hold the file while the complaint is open, then delete once the matter is resolved. Deleting first removes the only evidence of what the greeting said and when capture began.
Do I have to report this to the OAIC?
A complaint about consent is not automatically a notifiable data breach. The OAIC asks people to complain to the organisation first and to allow 30 days for a response, and states that generally a complaint must be about a matter that occurred less than 12 months ago. See the OAIC on what you can complain about.
The Privacy Act does not cover my business. Does that end the matter?
No. The OAIC states that the Privacy Act covers organisations with an annual turnover of more than $3 million, and some other organisations, so a smaller business may sit outside that route. The state or territory recording Act still applies, and it is the one that creates an offence rather than a complaint.
Next step
If the recurrence fix is what you want handled rather than the complaint, Zian AI builds inbound and outbound voice agents with private model deployment as an option, so disclosure and recording start can be sequenced inside your own environment. Apply For Partnership.