Where customer conversation data goes when you deploy an AI agent

An AI agent call typically copies your customer’s words into up to seven stores: telephony recording, call log, agent platform, transcription, model provider, your CRM and their phone carrier. Each runs its own clock, from zero to indefinite. OpenAI and Anthropic keep API logs up to 30 days by default. Twilio keeps recordings until you delete them.

Sources last checked: 23 September 2026. Every external figure on this page links to the publisher that produced it, and was re-read at that source before publication.

• Stores per call: up to seven, mapped in the data-flow table below.
• Longest clock you do not control: Australian carriers keep call metadata (not content) for at least 2 years under the Telecommunications (Interception and Access) Act 1979.
• The rule: your customer’s conversation stays in existence until the slowest store deletes it. We call this the longest-clock rule.
• Who can compel it: the customer (APP 12 and GDPR Art. 15), a litigant against you, a litigant against your vendor, and government agencies. US law reaches US providers wherever the server sits (the CLOUD Act).

Where does my customers’ conversation data actually go?

An AI voice agent is several companies chained together, and each keeps its own copy. Vendor names are examples of each layer, using published defaults as read on 23 September 2026. Your contracts and settings may differ.

Store 5 is the one you are least likely to see: in the common speech-to-text, model, text-to-speech design, every turn of the conversation is sent to the model as text, so the model provider holds a transcript even if you never switch recording on. If you want a named inventory of the stack behind a particular agent, how an AI sales agent is assembled shows which layers are involved.

What counts as conversation data, and what does not

Conversation data is everything the call produces that can be linked to the person on it. That includes the audio and the transcript. It also includes the derived records: the summary, the sentiment tag, the lead score and the objection your CRM logged. The OAIC is explicit: “Privacy obligations will apply to any personal information input into an AI system, as well as the output data generated by AI (where it contains personal information)” (OAIC guidance, 21 October 2024).

Three things sit outside that definition. Aggregate, de-identified findings (such as a finding about the best follow-up interval) are not the customer’s data. Carrier metadata is not content; the TIA Act’s retention scheme expressly excludes “the content or substance of a communication”. Training is a separate question from storage: a provider’s no-training default says nothing about how long its logs sit. An AI summary of a call is personal information whenever the call was, so deleting the audio while keeping the summary does not delete the conversation.

The longest-clock rule: how long does my customer’s call actually exist?

A customer’s conversation stays in existence until the slowest store in the chain deletes its copy. Legal holds extend that further. Worked with published defaults: a call happens on day 0 on a stack of Twilio (US1, recording on), Vapi on its Usage-only package, the OpenAI API with default settings, and your CRM. The customer asks for deletion on day 20.

1. Vapi raw data: already gone. The 14-day window closed on day 14.
2. OpenAI abuse-monitoring logs: may still exist. They age out on the provider’s schedule, by day 30.
3. Twilio recording: exists until you delete it. Delete it on day 20 and the metadata stays visible until day 60.
4. Twilio call log: a separate object from the recording. It is viewable in Console and API for 13 months from creation, roughly day 395, and after that still retrievable through Bulk Export. Twilio publishes no deletion date for it.
5. CRM transcript and summary: exist until you delete them. Nothing ages them out.
6. Customer’s carrier: keeps the call’s metadata to at least day 730. You cannot touch it, but it holds no words.

Result: act on the request the day it arrives and the words can be gone by day 30. The fact of the call can survive to day 730 or later. Leave the CRM row or the Twilio recording unmanaged and the words have no end date at all. Your retention policy is only as short as the longest clock in your agent’s supply chain.

Who can subpoena or compel an AI call recording?

Four kinds of requester reach different parts of the table.

• The customer. Under APP 12 an Australian organisation must give access within a reasonable period, which the OAIC says “should not exceed 30 calendar days” (APP guidelines, 12.67). Under GDPR the controller “shall provide a copy of the personal data” (Art. 15(3)) within one month, extendable by two (Art. 12(3)).
• A litigant suing you. Your CRM copy is your own record, so store 6 is reachable through ordinary discovery in a dispute you are party to.
• A litigant suing your vendor. In US courts a subpoena can command a non-party to produce “electronically stored information” in its “possession, custody, or control” (FRCP 45(a)(1)(A)(iii)). You need not be party to the case.
• Government. In the US, the Stored Communications Act requires a warrant for content held 180 days or less. Subscriber details and “telephone connection records” are reachable by administrative subpoena, and a provider must preserve records for 90 days, renewable once for another 90, on request. In Australia, stored carrier metadata is available to law-enforcement and national-security agencies under the TIA Act.

A retention clock only protects you if it ran before anyone asked.

When a court overrides deletion: the OpenAI preservation order

On 13 May 2025, in the New York Times copyright litigation, US Magistrate Judge Ona T. Wang directed OpenAI to “preserve and segregate all output log data that would otherwise be deleted on a going forward basis”. This applied whether the data would have been deleted “at a user’s request or because of ‘numerous privacy laws and regulations’” (order, ECF 551).

A stipulated order filed 9 October 2025 ended the ongoing duty “as of September 26, 2025” (stipulation and order). Data already segregated stayed preserved, except requests from the EEA, Switzerland or the UK. Accounts on a list of named domains stayed under preservation going forward. So for 136 days, from 13 May to 26 September 2025, a published 30-day deletion default did not decide how long output logs lived. A court did. A vendor’s retention promise is a default, and a court order in a case you are not party to can suspend it.

Does keeping the data in Australia keep it out of US reach?

Not on its own. The CLOUD Act requires a US provider to disclose data in its “possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States” (18 U.S.C. § 2713). An Australian region with a US-incorporated vendor changes where the data sits. It does not change who can compel it.

Residency is also narrower than it sounds. OpenAI offers Australian storage, but regional processing only in the US, Europe and UAE. Twilio’s AU1 comes with the rollout caveat quoted in the table. APP 8 governs your disclosure overseas in the first place, and that side is covered in the enterprise security review questions for AI sales agents. For EU data, GDPR Article 48 pushes back: a third-country court order is only recognised “if based on an international agreement, such as a mutual legal assistance treaty”. A US processor holding EU data can therefore receive an order that GDPR does not recognise.

How to shrink the map: settings that change the answer

Most of the table can be changed, but each lever leaves something untouched.

Doing this yourself is documentation work. Seven stores times four questions (what, where, how long, who can compel) gives 28 answers you need in writing, each checked against a contract clause rather than a sales call. The asset-by-asset exit view is in what you own when you leave a lead generation vendor. For recording retention in finance and health, see call recording compliance. For how a managed programme handles these stores at corporate volume, see AI appointment setting for corporate sales teams. In a default AI agent stack, two copies have no end date: the telephony recording and the CRM transcript. Setting a retention rule on both closes most of the gap.

This page is general information about published vendor terms and laws, not legal advice. Confirm your obligations with your own counsel.

Frequently asked questions

Can my customer ask for a copy of their AI call?

Yes. In Australia, APP 12 gives individuals a right of access, and the OAIC says a reasonable response period “should not exceed 30 calendar days” (APP guidelines, paragraph 12.67). Under GDPR the controller must provide a copy within one month, extendable by two. The request covers transcripts and AI summaries as well as audio.

Can a court force an AI vendor to keep data I deleted?

Yes. On 13 May 2025 a US magistrate judge directed OpenAI to preserve output log data “that would otherwise be deleted”, including data a user asked to delete (court order). The duty ended as of 26 September 2025; data already segregated stayed preserved.

Does storing AI call data in Australia protect it from US subpoenas?

Not if the vendor is a US provider. The CLOUD Act requires disclosure of data in a provider’s possession, custody or control “regardless of whether” it is located inside or outside the United States (18 U.S.C. § 2713). Residency changes location, not compellability.

How long do phone companies keep records of AI calls in Australia?

At least two years for metadata such as number, time and duration, under section 187C of the TIA Act. The content of the call is not part of the retention requirement (Home Affairs data retention guideline).

Is an AI-generated call summary personal information?

Yes, if it relates to an identifiable person. The OAIC states that privacy obligations apply to personal information input into an AI system and to “the output data generated by AI (where it contains personal information)” (OAIC guidance).