PIPEDA, Recordings, and US Subprocessors: a Plain Note for Owners

Recording, in Ontario

Ontario is one-party consent for most business calls. We still put a short recording disclosure on the line, and the agent will say it is automated if asked. That is the caller-facing piece. The shop-facing piece is that recordings and transcripts exist so we can dispatch, book, and debug. They are not there to build a marketing dataset.

Who owns the file

The recordings and transcripts belong to you. We do not sell them. We do not use them to train foundation models. If you want them deleted, that is a request we honour. Minimum collection is the default: name, number, address, postal code, symptoms, unit notes, and whatever else the script needs to book or dispatch.

Where the call actually goes

Voice AI does not run entirely in Canada on this stack. Telephony and SMS go through Twilio. The live voice pipeline is Vapi. Conversation and summaries go through Anthropic. Call logs sit in Supabase. The app is hosted on Render. Several of those process data in the United States. The security page lists each subprocessor and its region.

If a vendor tells you the opposite without naming those layers, they are selling a feeling. We would rather lose a buyer who needs a Canadian-only residency story than invent one. A HIPAA or SOC 2 scoped build is a separately priced engagement. The default plan is PIPEDA-aligned handling, not a covered-entity stack.

• Disclosure on the line, even though Ontario is one-party consent for most business calls.
• Recordings and transcripts are the client’s, not a training set.
• Named US subprocessors, published, not hidden behind “the cloud.”
• SOC 2 and HIPAA are quoted when a buyer actually needs them.