AI & automation

Call recordings, transcripts, and AI: where the data goes

A modern business phone system holds more than audio: call records, recordings, voicemails, transcripts, message history, and AI-generated summaries. This guide maps what exists, where it lives, and the questions that determine whether an AI feature is safe to turn on.

Updated August 30, 2026

Every phone system produces data as a side effect of working. Some of it is unavoidable (a call cannot be billed without a record that it happened), some of it is optional (recording is a switch), and some of it is new with AI features (transcripts and summaries did not exist in the analog world). Privacy questions get easier when the categories are kept separate, because each has its own sensitivity, retention logic, and legal footing.

What a phone system actually stores

Metadata (the CDR layer) is often more revealing over time than any single recording: it shows who talks to whom, how often, and when.
DataContainsExists when
Call detail records (CDRs)Numbers, direction, timestamps, duration, routing outcomeAlways; this is how calls are billed and audited
RecordingsFull call audioOnly when recording is enabled, by rule or per call
VoicemailsCaller audio plus metadataWhen a call reaches a mailbox
TranscriptsText of recordings, voicemails, or live callsWhen transcription or an AI agent is enabled
MessagesSMS and MMS content and mediaWhenever texting is used; carriers and platforms both handle it in transit
AI artifactsSummaries, extracted fields, sentiment, agent conversation logsWhen AI features run over calls or messages
Contact dataNames, numbers, notes, CRM-synced recordsWhen imported or synced

The processing path behind AI features

Transcription and AI features add processing steps, and each step is a place where data crosses a boundary. A recording is transcribed by a speech-to-text engine; a transcript may be summarized or answered by a language model; an AI receptionist streams live audio through the same kind of pipeline in real time. The operative questions are the same regardless of vendor: which subprocessors touch the audio and text, in which regions the processing runs, whether the data is stored by those subprocessors after processing, and whether any of it is used to train models beyond your account.

  • Speech-to-text: live audio or stored recordings leave the telephony platform for an ASR engine, in-house or third-party.
  • Language models: transcripts, message threads, or caller utterances are sent to a model for summarization, answering, or extraction.
  • Agent access: tools reached over MCP or APIs let external AI agents read call logs and messages; the permission scope of those tokens is part of the privacy surface.
  • Storage: recordings and transcripts persist in the platform (and its backups) until a retention rule or a deletion request removes them.

Retention: how long everything lives

Retention is a policy decision disguised as a storage default. Keeping everything forever maximizes hindsight and maximizes exposure; a breach or a subpoena reaches whatever exists. Reasonable practice sets retention per data category rather than globally: CDRs are commonly kept for years for billing and dispute purposes, while recordings and transcripts are kept only as long as their business use (quality review, dispute windows, regulatory minimums in some industries) actually requires.

  • Look for per-category retention controls: recordings, voicemails, transcripts, and messages should be separately configurable.
  • Deletion should propagate: removing a recording should remove its transcript and derived summaries, and reach backups on a stated schedule.
  • Legal holds override policy: litigation or investigation can require preserving data past its normal window.
  • Some sectors have minimum retention rules (financial services call-recording mandates are the classic example); minimums and maximums must be reconciled deliberately.

Who can see what

Inside a business, call data has an audience problem: recordings and transcripts of one team's calls are rarely appropriate reading for everyone. Access control that matches the org chart (per-number, per-team, or per-role visibility), plus an audit log of who listened to or exported what, is the difference between a system of record and a liability. On shared numbers the calculus is explicit: the whole point is shared visibility, so the boundary moves to who is assigned to the number.

Outside the business, the platform operator itself is an audience: support staff may be able to access account data under defined procedures, and lawful-process requests (subpoenas, warrants) reach whatever the platform retains. A provider's law-enforcement guidelines and transparency practices are part of its privacy posture.

Recording consent
One-party and two-party (all-party) consent state laws govern recording and, by extension, live transcription of call content. Covered in depth in call recording; AI listening does not relax any of it.
GDPR / UK GDPR
Call recordings and transcripts containing an identifiable person's voice or data are personal data: they need a lawful basis, are subject to access and erasure requests, and constrain transfers outside the EEA/UK.
CCPA/CPRA and state privacy laws
California and a growing list of US states give consumers access and deletion rights that reach call data held about them, with business-to-business nuances.
HIPAA
If calls carry protected health information, the phone platform is a business associate: a signed BAA, and features configured to honor it, are prerequisites, not nice-to-haves.
TCPA and wiretap-law AI claims
A wave of litigation has tested whether AI transcription and analytics of calls constitute wiretapping under state statutes such as CIPA in California. The theories are contested, but they are a live reason to disclose recording and AI processing clearly.

What to ask a provider before enabling AI features

  1. Which subprocessors receive audio, transcripts, or messages, and where do they process it?
  2. Is customer content used to train models shared beyond our account, and where is that committed in writing?
  3. What are the retention defaults per data category, and can we shorten them?
  4. Does deletion cascade to transcripts, summaries, and backups, and on what timeline?
  5. What access controls and audit logs exist for recordings and transcripts internally?
  6. Can AI features be enabled selectively (per number, per team) rather than account-wide?
  7. Is a DPA available, and a BAA if health information is in scope?