AI voice agents for lead follow-up: what works, what's risky

What an AI voice agent actually is

An AI voice agent is software that places or answers a phone call, listens with speech recognition, decides what to say with a language model, and speaks with a synthesized voice. The good ones sound like a calm person on a decent headset. They can hold a short conversation, ask questions in a sensible order, handle interruptions, and write a summary afterward. They cannot read a room, negotiate, or know what they do not know.

In lead follow-up, the job is narrow: call a person who just filled out a form, confirm what they need, ask the qualifying questions a good appointment setter would ask, and book a time. That narrowness is why it works.

Where it works

Speed

Harvard Business Review's audit of 2,241 companies found the average response to a web lead was 42 hours and 23% of companies never responded. Companies that responded within an hour were about 7 times more likely to qualify the lead. The 2007 Lead Response Management study (vendor data from InsideSales.com, often mislabeled an MIT study) found contact odds 100 times higher at 5 minutes than at 30. A voice agent makes the call at about one minute, for every lead, without a dispatcher noticing the form email. All the response-time numbers with sources.

Sources: HBR, 2011; Lead Response Management study, 2007.

Evenings and weekends

Podium reports that 40% of home-service calls arrive after hours. A voice agent does not need an on-call setter. It calls inside quiet hours (we use 8 am to 8 pm in the lead's own time zone) and emails the rest instantly with a promise of a morning call that it then keeps.

Source: Podium, home services.

Consistency

It asks the same eight questions in the same order to every lead, scores every answer on the same rubric, and writes the same five-bullet summary. Human setters have good days. The agent does not have days.

Volume spikes

A hailstorm produces fifty roofing leads in an hour. The agent calls all fifty at minute one. No queue, no overtime.

Where it fails

• Complex objections. "My brother-in-law says you guys overcharge." A person handles that. The agent should acknowledge it, note it, and move on to booking, not argue.
• Pricing. Unless you give it a range, it must not invent one. An agent that guesses a price costs you the job.
• Anything regulated. Insurance coverage, tax credits, medical suitability, VA benefits. The right answer is always "the consultant will go over that."
• Bad audio. Speakerphone in a truck, a crying baby. Agents degrade faster than humans in noise. Script a graceful exit: "I'll have a team member call you back."
• Leads who hate robots. A minority, but real. The fix is honest disclosure and an easy path to a person, not a more human-sounding voice.

What's risky: the law

On February 8, 2024, the FCC ruled (Declaratory Ruling 24-17) that calls using AI-generated voices are "artificial or prerecorded voice" calls under the Telephone Consumer Protection Act. In practice: an AI voice call to a cell phone requires the called party's prior express consent, and if the call is marketing, prior express written consent. The one-to-one consent rule that would have required naming each seller on a lead form was vacated by the Eleventh Circuit in January 2025 and later repealed by the FCC, but the base requirement did not move.

What that means for a contractor:

• The lead must have agreed, on your form, to be called by an AI or automated voice. An unchecked-by-default checkbox with clear language is the standard.
• You keep proof: the wording, the time, the page.
• Purchased lists, aged leads and "we had a conversation once" do not count.
• Quiet hours (8 am to 9 pm local under the TCPA; we use 8 to 8), opt-out handling and Do Not Call rules still apply.

Sources: Hunton on FCC 24-17; Pierce Atwood on the Eleventh Circuit decision; Womble Bond Dickinson on the repeal. Not legal advice. The TCPA in plain English.

Our rule is simpler than the statute: the agent calls only people who just asked this business to contact them. No lists, ever. It is the one design decision that removes most of the risk.

What's risky: trust

The second risk is reputational. An agent that opens with "Hi, this is Jessica from Summit Roofing" and gets caught being software has cost you the lead and probably a review. Ours opens with the business name, says it is an AI assistant, and says the call may be recorded. Conversion does not suffer. People are used to talking to software; they are not used to being lied to by it.

How to script one

1. Open with the business name, the AI disclosure and permission. "This is Lily, an AI assistant for Summit Roofing. Is now an OK time for two quick questions about the leak?"
2. Ask six to eight questions, one at a time. Urgency, what and where, ownership, timeline, decision-makers, access, best time. The trade decides the set; see the roofing or bath remodel question lists.
3. Score A, B or C as you go. Decide in advance what gets booked automatically and what gets a human callback.
4. Book, repeat it back, confirm the channel. Day, time, who will be there. Email confirmation now; text when carrier registration allows.
5. Write the five-bullet summary. Need, urgency, decision-makers, timing, anything unusual. Send it with the recording.
6. Give it a list of things it must not say. Prices you did not provide, coverage, guarantees, medical or legal opinions.

What to measure

Metric	Target	If it is off
Median seconds to first call	Under 90	Webhook or consent capture is broken.
Contact rate	Above half of consenting leads on the first attempt	Wrong caller ID, bad numbers, or calling outside the lead's quiet hours.
Set rate on Tier A	Trade-specific; compare with your human setters	Questions in the wrong order, no calendar slots, or the script argues instead of booking.
Opt-outs and "give me a person"	Low single digits	Disclosure is weak or the voice is pushy. Fix the opening.

If set rate does not move within a month, the technology is rarely the problem. The script is. That is why we listen to real calls and tune monthly on the Done For You plan. How AI appointment setting works here.