AI voice agents: how to evaluate calls, cost and risk

What it does on a call

• Calls after an approved trigger, such as an opted-in form submission, and within permitted hours.
• Collects the qualification fields the business has defined and sends uncertain cases to a person.
• Answers only approved, low-risk questions and avoids regulated or contractual advice.
• Offers available calendar slots when the integration and permissions are working.
• Applies a documented retry and suppression policy, including an immediate stop on request.

What it looks like in practice

Consider a hypothetical gym running paid ads, with more leads than staff can phone promptly. A lead submits a form and the agent calls after the approved trigger, asks the gym's qualification questions, offers a trial slot and follows the approved retry policy if nobody answers. This illustrates the workflow, not an Ace Digital case result or a promised conversion rate.

Consent and jurisdiction come first

Calling rules depend on the country, call purpose and number type. In the United States, the FCC has confirmed that AI-generated voices fall within TCPA restrictions on artificial or prerecorded voice calls. That is not a universal rule for every market. Record the required consent, disclosure, suppression and recording policy with qualified counsel before launch.

Where voice agents are the wrong tool

The following conditions can make a human workflow safer or more economical.

• Complex consultative sales. If the first conversation genuinely requires judgement about a bespoke problem, an agent qualifying it is fine but an agent selling it is not.
• Low lead volume where the implementation, monitoring and compliance work exceeds the measured value of faster follow-up.
• Anything where a wrong answer is expensive. Regulated advice, medical guidance, contractual commitments. The agent should be routing those, not answering them.
• Calls without a documented lawful basis or required consent. The applicable rule depends on jurisdiction, number type and call purpose.

Cost, including the part that is not fixed

Build cost depends on the integrations, qualification logic, languages, testing and handover. Running cost often includes telephony and voice usage linked to call minutes. Model connected minutes, average duration, answer rate, platform fees, monitoring and human review. A quote that omits those assumptions is not comparable.

The things that decide whether it works

• Speed. The value is in calling while the person is still thinking about the enquiry, so the gap between form and call is the number to watch.
• Knowing when to stop. An agent that will not hand over to a human is worse than no agent.
• Being straight about what it is. Callers work it out anyway, and pretending otherwise costs you the trust you were calling to build.
• Retry logic that follows the consent, frequency and stop rules approved for the market and audience.