The pitch, and the only number that matters
AI voice vendors sell on cost per minute. That is the wrong unit for a partial. A partial is not a support ticket. It is a person who already hesitated once in front of a monthly health subscription, and the call exists to turn that hesitation into a payment.
The number that matters is cost per closed sale. A bot that is ten times cheaper per minute and closes a fraction of the patients is not cheaper. It is burning the warmest leads you own. So the real question is not “can a bot make the call?” It is “will the patient stay on the line, will the call be legal, and will it ring at all?”
Wall 1: Americans say they don’t want AI on the other end
The survey data is consistent, and it is not new:
- In a Gartner survey of 5,728 customers, 64% said they would prefer companies didn’t use AI in customer service, and 53% said they would consider switching to a competitor if they found out a company was going to. Their top concern: AI makes it harder to reach a person. (Gartner, 2024)
- In a 2026 Gartner survey of 3,566 customers, 87% said that when a company uses GenAI for service, it is essential to offer a way to reach a human agent. (Gartner, 2026)
- In a Gartner survey of 1,539 U.S. consumers, 50% said they would prefer to give their business to brands that don’t use GenAI in customer-facing messages, ads and content, and 68% said they frequently wonder whether what they see is real. (Gartner, 2026)
Notice what these surveys measure: service and marketing, situations where the customer is usually the one who started the conversation. An outbound sales call is the hard version. The brand calls, uninvited in that moment, and asks for a card number for a weight-loss, hormone or sexual-health program. If people push back on AI when they reached out, expect more resistance when AI reaches out to them and asks for money.
Wall 2: the phone is already a trust problem
Before any voice speaks, the patient has to pick up. That was hard long before AI:
- Pew Research found that only 19% of U.S. adults generally answer cellphone calls from unknown numbers; roughly eight in ten don’t. (Pew Research Center, 2020)
- Hiya’s 2026 State of the Call survey of more than 12,000 consumers in six countries found that nearly 1 in 3 received an AI deepfake scam call in the past year and 34% said their trust in phone calls fell. Americans reported an average of 10 unwanted calls a week. (Telecompetitor on Hiya, 2026)
That is the room a synthetic voice walks into. When a patient hears a voice that is almost human, the first question in their head is not “is this my GLP-1 program?” It is “is this a scam?” And the default answer to a doubtful call is a hang-up.
Wall 3: the law treats AI voices as artificial voices
This is where AI calling stops being a product decision and becomes a legal one. A short tour (not legal advice; run your program past counsel):
- The FCC ruling. In February 2024 the FCC unanimously confirmed that AI-generated voices count as an “artificial” voice under the Telephone Consumer Protection Act. Because it clarified existing rules, it took effect immediately. (FCC; Cooley analysis)
- Written consent. Under the FCC’s TCPA rules, telemarketing calls that use an artificial or prerecorded voice to cell phones or residential lines generally require the consumer’s prior express written consent. (47 CFR § 64.1200)
- Per-call damages. The TCPA lets consumers sue for $500 per violation, and a court can triple that for willful or knowing violations. (47 U.S.C. § 227(b)(3))
- The FTC’s rule. The Telemarketing Sales Rule requires an express written agreement before prerecorded sales messages, and requires every sales call to disclose promptly who the seller is, that the call is a sales call and what is being sold. (16 CFR § 310.4)
- Bot disclosure. States are moving the same way. California already makes it unlawful to use a bot online that hides its artificial identity to drive a purchase; clear disclosure removes the liability. (Cal. Bus. & Prof. Code § 17941)
Put those together and look at the per-call math. At $500 a call, a consent gap on a list of 10,000 partials is a theoretical exposure of $5 million before trebling (illustrative). And the safe path, telling the patient up front that they are talking to a machine, sends you straight back to Wall 1.
Wall 4: the network filters before anyone says hello
FCC rules let carriers block calls without liability when the blocking is based on “reasonable analytics designed to identify unwanted calls” that take caller ID authentication into account. (47 CFR § 64.1200(k)) In practice, the same analytics engines put labels like “Spam Likely” on screens.
Those engines look at how numbers behave: volume, call length, complaints, authentication. The headline feature of an AI voice agent, thousands of parallel calls with no one getting tired, is exactly the kind of pattern you have to manage carefully to keep numbers clean. A flagged number doesn’t just lose one call. It drags down every call on that number until the reputation is repaired.
What about conversion rates?
Here is the honest answer: we found no independent, public study of AI voice agent conversion rates on telehealth partials. There are vendor case studies, chosen by vendors. Treat any number without a method, a sample and a control group as marketing.
What you can do is measure the gap that matters. An illustrative model, not data:
That is one month, first payment only. On a subscription, a few points of conversion outweigh almost any saving on minutes. If a vendor pitches you AI calling, ask for one thing: a split test on your partials, against a human team, reporting answer rate, conversion of answered calls, refunds and cost per closed sale.
Where AI does belong
None of this means telehealth sales should ignore AI. It means AI belongs behind the call, not in place of the voice:
- Deciding who to call first and when, inside each state’s calling hours.
- Watching number reputation and answer rate in real time.
- Guiding the agent live with the brand’s approved answers.
- Scoring every call for compliance and coaching.
Let software do the parts that need speed and memory. Let a trained person do the part that needs trust.
The bottom line
AI voice agents are cheap per minute and expensive per sale. American consumers say they want a human, the law treats synthetic voices as artificial voices with per-call damages, and the network is built to filter exactly the high-volume automated traffic bots generate. For a partial, the patient who already hesitated once, the voice on the line should be a person who knows your category, sticks to your script and closes on your payment page.