AI Voice Cloning: What Businesses Can and Cannot Do

A minute of someone speaking is now enough to generate new speech in their voice saying anything you type. Over a phone line, where the audio is compressed anyway, most listeners will not notice. That opens a small number of genuinely useful things for a business and one significant problem, and the problem does not require anyone to target you specifically.

What it is actually good for

The marketing suggests endless applications. The honest list is short.

Training and internal video narration. Consistent across modules, and updating one section no longer means rebooking whoever recorded the original.

Phone prompts and hold messages. Changeable without a recording session, which is why these are usually years out of date.

Translation with a recognisable voice. In a bilingual market this is the one with the clearest return, because producing a French version of your content used to mean a second voice and a second session.

Accessibility. An audio version of written material, which is cheap to produce and genuinely useful to some customers.

Every one of those involves your own voice or a hired voice with a signed agreement. That is not a coincidence, it is the boundary.

Consent is not a formality

Get it in writing, before you record, and make it specific.

The consent should say	Because otherwise
What it will be used for	Scope drifts from greetings to adverts
For how long	It runs indefinitely by default
What happens if they leave	A former employee is still answering your phone

That last row is the one businesses stumble into rather than choose. Someone recorded a greeting in 2024, left in 2025, and their voice is still speaking new sentences on your behalf. Under Canadian privacy law a voice is personal information, so purpose limitation and retention apply exactly as they do to anything else, and the general approach in keeping AI use PIPEDA-compliant covers the rest.

Somebody may already have what they need

Here is the part worth sitting with. Cloning your voice does not require your cooperation, and it does not require anyone to target your business specifically.

If an owner or manager has spoken on a podcast, appeared in a video on the website, or posted anything with audio, the raw material exists and is public. That is enough to generate a convincing phone call, and the usual pattern is an urgent request to whoever handles payments: change these bank details, approve this transfer, the client is waiting.

The defence is not technical and no tool will sell it to you. It is a rule: payment details and transfers are verified through a separate known channel, always, regardless of who appears to be asking and how urgent it sounds. Make it a policy rather than a judgment call, because judgment is exactly what an urgent call is designed to bypass.

Tell your team the rule out loud

One conversation, five minutes, and it is the highest-value thing in this post.

Tell whoever handles money that a voice on the phone is no longer proof of identity, and that they will never be penalised for calling back on a known number to check. That second half matters more than the first. Most of these frauds succeed because a junior person did not want to appear to distrust the boss, and the fix is making it explicitly safe to verify.

Say when it is synthetic

Where a customer would reasonably assume they are hearing a person, tell them. A phone system greeting callers in a synthetic version of the owner's voice feels perfectly normal internally and lands badly when someone works it out.

For narration and training content it matters less, though it costs nothing to mention. Regulatory direction across several jurisdictions is toward labelling synthetic media, so this is moving from courtesy to expectation, and the reasoning in disclosing AI to customers applies here with the volume up, because a voice is more personal than a paragraph.

Where the cost went

The reason this is suddenly a question for ordinary businesses is price. Studio narration was a real budget line, which meant most small companies simply never had audio content and never updated the phone prompts.

That constraint has effectively gone, as we covered when professional voiceover got dramatically cheaper. The practical consequence is that the audio backlog you have been ignoring for years is now affordable, and the thing standing between you and it is a consent form and a decision about disclosure rather than a budget.

Frequently Asked Questions

How does voice cloning work now?

You supply a short sample of someone speaking, often a minute or less, and the system can then generate new speech in that voice saying anything you type. Quality has improved to the point where a casual listener will not notice, particularly over a phone line where the audio is compressed anyway. The cost has collapsed alongside the quality, which is the part that matters commercially: what used to require a studio session and a voice actor now costs pennies and finishes in seconds.

What are the legitimate business uses?

Narrower than the marketing suggests, but real. Consistent narration across training modules and internal videos, so updating one section does not require rebooking anyone. Phone system prompts and hold messages that can be changed without a recording session. Translating your existing content into other languages while keeping a recognisable voice, which matters in a bilingual market. And accessibility, giving written material an audio version. Notice all of those involve your own voice or a hired voice with a signed agreement.

What consent do we need?

Written, specific, and revocable, and get it before you record anything. It should say whose voice it is, what it will be used for, for how long, and what happens to the model if they leave. A staff member who agreed to record a phone greeting has not agreed to have their voice generating new sentences three years after they resigned. Treat a voice as personal information under Canadian privacy law, which means purpose limitation and retention rules apply to it just like anything else.

What is the fraud risk to my business?

Real and increasing, and it does not require anyone to target you specifically. A short clip of an owner or manager speaking, which is often available on a website, a podcast, or social media, is enough material to generate a convincing phone call. The usual pattern is an urgent request to a finance person to move money or change payment details. The defence is not technical: it is a rule that payment changes are verified through a separate known channel, never approved on the strength of a voice.

Should we disclose synthetic voice to customers?

Where a customer would reasonably assume they are hearing a person, yes. A phone system that greets callers in a synthetic version of the owner’s voice without saying so is the kind of thing that feels normal internally and lands badly when discovered. For narration and training content it matters less, though it costs nothing to say. Regulatory direction across several jurisdictions is toward labelling synthetic media, and disclosing early is easier than retrofitting trust.