Sell call recordings and call center data for AI training

What AI buyers want from call data

Voice agents and customer service agents need examples of real conversations: interruptions, confusion, escalations and resolutions.

• Audio with speaker separation
• Transcripts with timestamps
• Call outcomes, dispositions and QA scores
• Scripts and the SOPs agents followed

Transcripts are easier to de-identify than audio. A voice itself may identify a person. Calls paired with the QA rubric and outcome teach a model what a good call looks like, which is closer to what enterprise buyers describe wanting: QA processes and human feedback.

What we know about prices

We found no public 2025 or 2026 price for a company selling its call recordings to an AI buyer. We will not guess one. Related signals:

• Protege runs a marketplace that pays data partners a revenue share per use and is strongest in healthcare and media (SiliconANGLE).
• Scale AI's Dataset Collection Initiative invites companies to submit non-public datasets. Pay depends on quality and use case, with no amounts published (CX Today).
• Enterprise programs publish company-level tiers from $100K+ for a broad data package. Recordings are not on its published list.

All-party consent states

Some states require every party on a call to consent to recording. Per the Reporters Committee for Freedom of the Press (RCFP):

A caller who heard "this call may be recorded for quality and training" agreed to something. Whether that covers licensing the call to an AI company is a separate legal question. See call recording consent and AI.

Ambriz v. Google

In February 2025, a federal court in Northern California let a California Invasion of Privacy Act (CIPA) claim proceed against Google. The claim targeted its contact-center AI, based on that AI's capability to use call data to improve its models (Goodwin). The case was allowed to proceed. It was not a final ruling. It shows that using call data for model improvement can draw wiretap-style claims in California.

Other legal risks

• CCPA. Callers and staff are consumers. A transfer for money may be a "sale" unless the data is de-identified (Cal. Civ. Code 1798.140).
• HIPAA. Calls with patients may contain PHI, which can't be sold without authorization. De-identification uses Safe Harbor or Expert Determination, and HHS says risk "is not zero" (HHS).
• GLBA. Financial calls may contain nonpublic personal information that requires notice and opt-out (FTC).
• Card data. PCI rules say sensitive authentication data must never be stored after authorization. Recordings that captured it are a problem on their own.
• Changed terms. The FTC warns that quietly changing terms to allow AI training could be unfair or deceptive (FTC).
• Employees. Agents' voices and performance data are employee data. See employee data and AI training.

General information only, not legal advice.

How to prepare call data

1. Map each recording to the state of every party and the consent language used.
2. Exclude calls from all-party-consent states unless counsel approves.
3. Exclude health, financial-account and payment calls.
4. Prefer transcripts. Redact names, numbers, addresses and account details.
5. If audio is required, discuss voice de-identification with the buyer.
6. Pair calls with scripts, QA rubrics and outcomes.

Operating vs. shutting-down companies

Consent rules apply whether you are operating or closing. An operating call center can update its disclosures for future calls, with counsel's help. Historical recordings stay bound by the consent given at the time. For most companies, SOPs, QA rubrics and de-identified transcripts are a safer place to start. See selling SOPs.

Want to know what else your company could license? Take the 60-second estimate.

Frequently asked questions