An AI voice agent is a system that places or answers phone calls with a synthetic voice to do one narrow job: call an inbound lead within seconds, confirm an appointment, reactivate a past customer, or qualify and transfer to a human. It is not a better robocall blaster, and in the United States it is not outside the law: in February 2024 the FCC issued a declaratory ruling stating that when AI technologies such as voice cloning are used to emulate human voices on a call, those calls are "artificial" voice messages under the Telephone Consumer Protection Act (FCC; Mayer Brown).
This article is general information, not legal advice. Telephone rules change often, they differ by country and state, and the consequences of getting them wrong are financial. Before you dial anyone, get advice from qualified counsel for your jurisdiction and your use case.
In short
- The interesting question is no longer whether a machine can sound human. It is which call is worth making, whether you may make it, and when a human takes over.
- Compliance is a design input, not a launch checklist. Consent, identification, calling hours and opt-out handling shape the script.
- One agent, one job. Agents that qualify, book, upsell and object-handle do four things badly.
- The handoff is the product. Most business value comes from a warm transfer to a prepared human, not from the AI closing anything.
- QA is a weekly habit, not a launch task: listen to real calls, score them, fix the script.
- Measure outcomes in the CRM (conversations that led somewhere), never in dial counts.
Where AI voice actually earns its place
Salesforce's 2026 State of Sales report, based on 4,050 sales professionals in 22 countries, found that sellers spend about 40% of their time selling and that 54% already use AI agents; it also describes agents contacting 130,000 previously untouched leads in four months, producing 3,200 opportunities (Salesforce). That last figure is the honest case for voice: the biggest wins are usually on work nobody was doing at all, not work a good rep already does well.
Ranked roughly from easiest to hardest:
- Instant inbound callback. A form arrives; the agent calls back in under a minute, confirms the request and offers a time or a transfer. Closest thing to free money in most funnels, and the reason speed-to-lead is usually the first row on a leak map.
- Appointment confirmation and rescheduling. High volume, low ambiguity, easy to verify.
- Reactivation of past customers. You have an existing relationship and a record of it. Still requires checking your consent basis.
- Qualification and warm transfer. The agent gathers two or three facts and transfers to whoever is available, with the context attached.
- After-hours coverage. The call that would otherwise go to voicemail. Careful: outbound calling hours are restricted, and "after hours" is defined at the called party's local time.
Where voice agents are a bad idea: price negotiation, upset customers, anything requiring a promise about a contract, anything where a wrong statement creates liability, and any conversation where the person has already asked for a human.
US compliance basics (general information, not legal advice)
The rules below come from the FCC's ruling and from the TCPA implementing regulation at 47 CFR § 64.1200. Read them as the shape of the problem, then have counsel apply them to your business.
| Requirement | What the source says | Where it lives |
|---|---|---|
| AI voices are "artificial" | AI-generated or cloned voices used on calls are artificial voice messages under the TCPA; the existing prohibition and consent framework applies | FCC Declaratory Ruling, FCC 24-17, CG Docket 23-362 (FCC) |
| Consent for marketing calls | Calls that include or introduce an advertisement or constitute telemarketing, using an artificial or prerecorded voice, to wireless numbers or residential lines generally require prior express written consent | 47 CFR 64.1200(a)(2)–(a)(3) |
| Identification | The message must state clearly, at the beginning, the identity of the business responsible for initiating the call, using the registered business name | 47 CFR 64.1200(b)(1) |
| Callback number | During or after the message, state a telephone number for that business (not the dialer's), which for telemarketing must allow a do-not-call request during business hours | 47 CFR 64.1200(b)(2) |
| Calling hours | No telephone solicitation to a residential subscriber before 8 a.m. or after 9 p.m., local time at the called party's location | 47 CFR 64.1200(c)(1) |
| Do-not-call | Honor the national registry, and maintain your own internal do-not-call list with a written policy and trained personnel | 47 CFR 64.1200(c)(2), (d) |
Two more things worth knowing. First, the rules are moving: the FCC's "revoke-all" consent-revocation provision has been delayed twice and, per a January 2026 order, is not scheduled to take effect until January 31, 2027 (Burr & Forman). Anything you read about telephone compliance, including this article, has a shelf life. Second, states add their own layers, including laws about recording calls, and other countries have entirely different regimes. If you operate across borders, your compliance map has as many rows as your calling map.
What this means practically for a builder: consent status, calling window and opt-out state are data fields, and your workflow must check them before dialing. A system that can technically dial anyone in the table is a system that will eventually dial someone it shouldn't.
The MitHub voice agent spec
Fill this in before you write a single line of script. If you can't complete it, you are not ready to dial.
| Field | What to define |
|---|---|
| One job | The single outcome of the call, in one sentence. |
| Who may be called | The exact filter: consent basis, relationship, time zone, opt-out status, retry limits. |
| Opening disclosure | Business name at the start, why you're calling, and the callback number path. |
| Facts the agent may state | A short, fixed list. Everything else is "I'll have a colleague confirm that." |
| Facts the agent may collect | The two or three fields that actually change the next step. |
| Transfer condition | The trigger for a human handoff, including "the person asks for one." |
| Escalation and failure path | What happens if no one is available: voicemail policy, callback, or silence. |
| Hard stops | Prices, legal terms, medical or financial advice, anything with liability. |
| Logging | What is written back to the CRM on every outcome, including refusals. |
| Review loop | Who listens to which calls, how often, and who can change the script. |
Notice how little of that is about the voice. The voice is the easy part now; the operating constraints are the engineering.
The QA loop
A voice agent is the one automation your customers experience directly, so it needs the tightest review cycle of anything you build. Weekly is a reasonable rhythm.
- Sample. Pull a fixed number of calls per week, stratified: some transfers, some refusals, some hang-ups, some "no answer". Never review only the good ones.
- Score each call on five dimensions, pass or fail: - Identification: did it say who is calling, at the start? - Accuracy: did it state anything not on the approved list? - Respect: did it stop when asked, and log the opt-out? - Outcome capture: does the CRM record match what actually happened? - Handoff: did the human receive usable context?
- Fix one thing. Change the script, the data filter or the transfer rule, and note the date.
- Re-listen next week to check the fix held.
Any failure on Respect or Identification is a stop-the-line event: pause the campaign, fix, then resume. Everything else is iteration. This is the human-in-the-loop principle applied where it matters most.
How to measure it
Your call platform will offer you a dashboard full of numbers that feel like results. Most of them are activity. Build your own short list instead, and compare against your own baseline, since published benchmarks rarely match your industry, list quality or hours:
- Reach rate: conversations started ÷ people attempted.
- Qualified transfer rate: transfers a human judged worth taking ÷ conversations.
- Outcome rate: meetings, payments or resolutions ÷ people attempted.
- Cost per qualified transfer, all-in (platform, telephony, build and review time).
- Complaint and opt-out rate, watched like a smoke alarm.
- Data completeness: share of calls that wrote a correct outcome back to the CRM.
Run a baseline period before launch, or you will have no way to tell whether the agent helped. And do not accept "the campaign ran with no errors" as proof: open the records, as we argue in the revenue leak map.
What this looks like at scale
MitHub's pioneers have built AI voice campaigns that ran across 28 live branches of a multi-location lending business, including a 10-branch pilot with 13,159 AI calls. The lesson from work at that size is not that the technology is magic; it is that operations decide the outcome. Phone number provisioning, consent data, transfer availability during business hours, branch-level ownership and weekly call review determine whether a voice system produces revenue or noise.
If you want to learn to build systems like this properly, start with the free foundations in MitHub's Faculty of Revenue Reverse Engineering, especially Prove value fast for choosing the smallest useful system and Operate for the review loop that keeps it honest. And before you build anything: map where your revenue is actually leaking. A voice agent pointed at the wrong step is just a faster way to reach the wrong people.