TL;DR

  • Get written, specific consent from the team member before recording.
  • Record clean audio, test against a checklist, and tell callers it is AI.
  • Limit use by campaign and hours. Let the owner revoke anytime.

Get written consent from the team member first

Get a signed, specific consent form from the voice owner before anyone records a second of audio. A cloned voice is closer to a fingerprint than to a work product. Ask counsel whether biometric data rules apply where you operate. Your employment contract probably does not cover it, so use a separate form.

The form should name:

  • The voice owner and the company that will hold the model.
  • Purpose: which campaign types, such as surveys, reminders or win-back calls.
  • Languages the clone may speak.
  • Duration: a start date and an end date.
  • Pay: a fee, or a plain statement that there is none.
  • Revocation: how to withdraw, and how fast you will stop.
  • Exit: what happens to the model and source audio when they leave.

A buyer at a fintech investment platform asked how to make an AI agent sound like its own reps instead of a generic robotic voice. The wish is reasonable. It also means a real colleague's voice will reach thousands of customers.

Then check the law where your callers live:

  • US: The FCC ruled in 2024 that AI-voice calls need the called person's consent under the TCPA.
  • EU: Article 50 of the AI Act has applied since 2 August 2026. AI that talks to people, including phone systems, must say it is AI from the first interaction.
  • Colorado: From 1 January 2027, public-facing conversational AI must disclose that it is AI.
  • Other US states: Ask counsel whether any state you call has right-of-publicity or voice-likeness rules.

Keep two layers of consent apart. The employee's form covers using their voice. The person you call has their own layer: do-not-contact lists and an AI disclosure.

Record clean, varied source audio

Record in a quiet room with one speaker, because the engine copies everything it hears. ElevenLabs' docs say the model mimics speed, accent, breathing, mouth clicks and background noise. If the room echoes, the clone echoes.

Set up the recording:

  • Pick a small, soft room with no music, fans or echo.
  • Use a decent mic held at a steady distance.
  • Use the voice you would use on a real call.

Match the amount to the clone type. ElevenLabs recommends 1 to 2 minutes of good audio for an instant clone and 30 to 180 minutes for a professional clone. Clean audio beats long audio.

Read scripts built for your calls. Write lines for your campaign, such as a survey, a payment reminder or a win-back call. Include a greeting, a question, an empathetic line and a few numbers and product names. Record in every language the consent form lists. ElevenLabs warns that a clone cannot reliably speak a language the model does not support.

Keep a record of the session. Your QA manager should be able to answer five questions later:

  • Who recorded the audio?
  • When was it recorded?
  • Which consent version applied?
  • Where is the raw audio stored?
  • When will it be deleted?

Store the files in access-controlled storage and set a retention date. ElevenLabs notes you cannot export a clone as a file, so keep the source audio if you may need to rebuild it.

Test the clone with a quality checklist

Test the clone before any customer hears it, and give each check a clear pass or fail. Phone audio is compressed, so a voice that sounds great in a studio can sound thin on a call.

Run these checks on a real phone, not laptop speakers:

Check Pass means
Product names and numbers Every name, amount and date is said correctly
Pacing Speed matches the owner's natural pace, with no rushed digits
Clipped words No cut-off starts or ends of words
Accent drift Accent stays steady from the first line to the last
Emotion fit A reminder sounds calm, an apology sounds sorry
Latency No awkward gap after the caller speaks
Phone line Still clear and natural over a normal call

Work in three rounds.

  1. Talk to the agent in a browser. This catches script and pronunciation errors with no phone number or cost.
  2. Place a test call to your own phone. This checks latency and the compressed audio.
  3. Ask the voice owner to approve it. They should hear it before launch and have the right to say no. Then have three or four colleagues rate it blind.

Use ordinary listeners, not only the sensitive ones on your team. In one demo for a digital lender, a tester said he could tell the call was AI because he is more tuned to it than most callers. He still understood the agent and was understood.

Insight7 can help with this testing and review. Its voice campaigns are in beta and use ElevenLabs voices. You can test a campaign's agent by talking to it in the browser, and campaigns have calling hours, retry rules and a test call to your own phone. AI call scoring evaluates 100% of conversations instead of a 2% manual sample.

After launch, score real calls against a weighted scorecard instead of sampling a few. Add voice checks to the rubric:

  • dropped calls
  • long silences
  • callers hanging up in the first seconds
  • complaints that the voice sounded fake

Read transcripts for the flagged calls. Any pattern should send you back to this checklist.

Set usage limits before any call goes out

Write the limits down and have the voice owner sign them before the first call. They protect your colleague, and they protect the people who pick up the phone.

Add these to the consent form or an annex:

  • Campaign types: list the approved ones, such as surveys or win-back calls. Anything else needs a new signature.
  • Languages: only the ones you recorded and tested.
  • Hours and retries: set calling hours and a cap on attempts per contact.
  • Banned uses: no payment authorisation, no identity checks, and no script the owner has not read.
  • Kill switch: name one person who can pause every campaign at once, and set a deadline for acting on a revocation.

Banned uses matter most. Scammers use voice cloning because a familiar voice makes requests for money or information more believable, the FTC says. If your clone never asks for payments or codes, a real fraud call sounds different to your customers.

Tell callers it is AI, up front. Set this per campaign. A customer who knows your rep may think they are speaking to the real person.

Give callers a way to a human. In a Gartner survey of 3,566 customers, 87% said companies using GenAI for service must offer a way to reach a human. Build a transfer line into every script.

Keep opt-outs clean. Mark consent on each contact and honour the do-not-contact list.

Review, audit and retire the voice

Review the voice every month, and retire it the day consent ends. Score a monthly sample of calls against your voice checks. Ask the voice owner to approve the clone again whenever the script or language changes.

When the team member leaves or withdraws consent, stop their campaigns. Delete the model and the source audio, and keep the consent log.

Related reading: AI Voice Agent Latency Explained and the Response Times Customers Will Tolerate, How to Answer "Where Did You Get My Number?" on AI Outbound Calls With Example Responses and 9 Design Principles for Natural-Sounding Conversational AI in the Contact Center.

Questions people also ask

Is voice cloning legal?

The rules differ by place, so ask counsel before you launch. Written consent from the voice owner is the starting point. The calls you make with the clone carry their own rules. In the US, AI-voice calls need the called person's consent under the TCPA.

Is voice cloning illegal without consent?

It is risky, and the answer depends on where you operate. Some places have right-of-publicity or voice-likeness rules, so ask counsel which apply to you. Never clone a voice without a signed form.

How does voice cloning work?

A voice engine learns from short samples of speech, then turns typed text into new speech in that voice. ElevenLabs' docs say it mimics accent, pace and background noise, so noise and echo carry over.

How much audio do you need to clone a voice?

It depends on the clone type. ElevenLabs lists different lengths for instant and professional clones. Clean, varied audio beats a long, noisy recording, so cover greetings, questions, empathy and numbers.

Do you have to tell callers a cloned voice is AI?

In the EU, yes, from the first interaction. The FCC's proposed upfront-disclosure rule was not final as of mid-2026. Disclose anyway, wherever your callers are.

Can an employee take back consent for their cloned voice?

They should be able to, and your form should say how. Write down the notice period. On withdrawal, stop campaigns, delete the model and source audio, and keep the signed consent log.

How is voice cloning used by cybercriminals?

Criminals copy a trusted voice to make requests for money or data sound real. The FTC says scammers use voice cloning for this. It is why your clone should never ask for payments or codes.

Check consent before you launch

A cloned voice is a person's identity, so treat consent, tests and limits as one launch gate. If any is missing, the campaign waits. Check that the voice owner has signed a consent form naming each campaign use.

Book a demo of Insight7