A voice agent processes some of the most sensitive data your organisation handles: recorded speech, transcripts, and whatever callers disclose while trying to get something done. Treat a voice line as a system that collects personal data by design.

What gets processed

  • Call audio — the caller’s voice, which in most jurisdictions is personal data in its own right and biometric data in some.
  • Transcripts — the literal text of what was said, including anything volunteered.
  • Action data — whatever is passed to your systems, which can include account numbers and addresses.
  • Summaries — derived from the above.
Callers disclose things nobody planned for — medical details, financial hardship, information about third parties. Any assessment based only on the fields you designed will understate what the line actually holds.

Settle these before launch

Most jurisdictions require callers to be informed that a call is recorded or processed automatically, and many require it before the conversation begins. Requirements differ by country and sector — confirm what applies to yours.
Decide retention per data type. Audio, transcripts, and summaries do not need identical periods, and “keep everything indefinitely” is rarely defensible under data protection law.
Access to transcripts should be scoped to people with a reason to read them. Reviewing calls for quality is a legitimate reason; general availability across a company is not.
If your organisation has obligations about where data is processed, settle this during scoping rather than after building. It is the question most likely to force an architecture change late.See Deployment options.
Individuals can request erasure in many jurisdictions. Know how you would locate and remove one caller’s records across audio, transcripts, summaries, and anything your actions wrote downstream.

Minimise at the source

The cheapest data to protect is data you never collected.
  • Do not collect what the call does not need. If the agent never needs a national ID, do not ask for one.
  • Avoid putting sensitive values into action parameters that get logged widely.
  • Where verification is required, prefer confirming a detail the caller supplies over asking them to read out a full number.

Regional obligations

Saudi Arabia’s Personal Data Protection Law, and the equivalents across the Gulf, place specific requirements on processing personal data — including rules on transfers outside the country. If you operate under PDPL or a comparable regime, these constraints shape which deployment configuration is appropriate.
Nothing on this page is legal advice, and nothing here should be read as a compliance claim about Voho. Whether a given configuration satisfies a given obligation depends on your contract, your sector, and your jurisdiction. Involve your legal and data protection people early, and get commitments in writing.

Questions worth asking any vendor

Including us:
  1. What exactly is retained, and for how long, per data type?
  2. Where is data processed, and does that change under load?
  3. Is call data used to train models? If so, can that be switched off?
  4. What certifications exist, and what is their scope?
  5. What are the notification commitments if there is a breach?
  6. What happens to data when the contract ends?
A vendor that answers these precisely and in writing is a different proposition from one that answers them warmly in a sales call. Ask for the written version.