Call recordings and transcripts can be licensed to AI companies, but only where callers were told the call might be recorded and your notices and contracts cover the new use. The legal risk sits in consent and disclosure, so the first question about a call archive is what callers were told, and the audio comes second.
This guide is general information, not legal advice. Talk to a privacy lawyer before you license any recordings.
Why AI labs want real calls
Labs building voice agents and support assistants need examples of how real conversations go: a confused customer, an agent who asks the right question, an escalation, a refund approved or refused. Scripted recordings made for training sound scripted. A support team's archive shows the messy version, with the outcome attached.
The most useful call data links three things:
- The conversation, as audio or a transcript
- What the agent did in your systems during or after the call (a refund issued, a ticket escalated, an account changed)
- How it ended: resolved, called back, churned
That sequence is an episode, a situation followed by a human action and an outcome. Episodes built from support work are worth more than loose transcripts, because a model can learn which actions led to good outcomes. Our guide to episodes, records and environments explains the difference in detail, and the customer support page lists the records labs ask for most.
The consent problem
Two recent California lawsuits show where the trouble starts.
In July 2024, a class action against Patagonia alleged that its contact center vendor Talkdesk intercepted and analyzed customer calls without proper consent. According to Customer Experience Dive, callers heard that calls "may be recorded for quality training purposes" but were not told the calls went to Talkdesk or how its AI used them. A separate suit claimed Home Depot used Google's Cloud Contact Center AI to record and analyze calls without adequate consent, as reported by the same outlet. In that case, Ambriz v. Google, the court let the claims proceed at the pleading stage on the theory that Google could use call data to improve its AI models, according to Debevoise & Plimpton.
These are allegations, and neither case sets a price or a final rule. They do show that a standard "this call may be recorded" message may not cover sending recordings to a third party for model training. Before we offer any recordings, we check:
- What the recording notice said, and in which states and countries the callers were
- Whether your privacy notice mentions sharing or AI training
- Whether the calls were made on behalf of a client (a common setup for outsourced call centers)
- Whether the recordings contain payment card numbers, health details or other sensitive data
Our default rule is simple: call recordings are included only where callers were told the call may be recorded. Where the notice or the contracts don't support licensing, we leave those calls out.
Outsourced call centers and client data
Many call centers answer phones for other companies. In that setup, the recordings usually belong to the client, or your contract limits how you can use them. Under GDPR, a company that processes data on a client's instructions is a processor, and a processor that starts deciding its own purposes for the data becomes a controller for that processing under Article 28(10) of the GDPR. In practice, outsourced call centers need written permission from each client whose calls they want to license. Our guide on selling data you hold for customers covers how to ask for it.
What gets removed
Voices identify people, so raw audio is the hardest call data to de-identify. Most of what we prepare is transcripts with:
- Caller and agent names replaced by stable pseudonyms
- Phone numbers, emails, addresses, account and card numbers removed
- Dates generalized to a month
- Rare details that could single someone out (an unusual order, a named small business) removed or generalized
- A re-identification test run before every sale, with the report kept
Free text is the riskiest part of any dataset, and a transcript is all free text. That is why call data takes longer to review than an invoice export. Our guide to de-identification vs anonymization explains why removing names is only the start.
What call recordings sell for
No public price list exists for contact center recordings. Speech data vendors quote per project, and call archives are priced in private, negotiated licenses that depend on consent, language, volume, transcription quality and whether outcomes are linked. We won't quote a per-hour figure we can't source.
The factors that move value, in our view:
| Factor | Effect |
|---|---|
| Clear recording notice and rights | Without it, the data can't be sold |
| Outcomes linked to calls | Turns transcripts into episodes |
| Agent actions in your systems | Shows what good handling looks like |
| Volume and languages | More coverage for buyers |
| Industry (insurance, finance, telecom) | Specialized domains are harder to find |
The customer support data prices page collects what public signals exist. For a range on your own archive, the calculator gives an estimate that we confirm on a call.
Calls keep coming
A support team takes new calls every day. After the first license sells your history, the SDK can run monthly, pull only new tickets and transcripts, scrub them on your machine and upload a batch. Each license with an active refresh term pays for every accepted batch. A growing team creates more records each month, so batches get bigger without extra work. See how recurring data revenue works.
Check what your calls are worth
Get an estimate with the valuation calculator, or start with the seller hub. The first call starts with your recording notices and client contracts, because those decide what can be sold.
Frequently asked questions
Is a "this call may be recorded" message enough consent?
It may not be. The Patagonia and Home Depot lawsuits both argued that a recording notice did not tell callers their calls would be shared with or analyzed by a third party's AI. A lawyer should review your notices before any recordings are licensed.
Can we sell the audio, or only transcripts?
Transcripts are easier, because voices can identify people and are hard to scrub. Most of what we prepare is de-identified transcripts linked to actions and outcomes. Audio needs a separate review and stronger consent.
We run a call center for clients. Can we sell their calls?
Usually only with each client's written permission. Recordings made for a client are typically the client's data or restricted by your contract, and using them for your own purposes can change your legal role.
Do you remove payment card numbers?
Yes. Card numbers, account numbers, phone numbers and email addresses are removed before any data leaves your control, and every dataset gets a re-identification test before sale.