Every AI voice agent demo sounds the same. A pleasant synthetic voice handles an objection, books a meeting, and the founder on stage says “and it never sleeps.”
What the demo skips is the part that decides whether the thing is an asset or a lawsuit: who you are allowed to call, when, with what disclosure, and what happens when someone says stop.
I build and run one of these — an agent that dials newly registered FMCSA broker authorities from a daily feed for a US carrier’s sales team. Here is what actually has to be true before you turn one on.
I build software, not legal advice. The rules below are the ones that shaped how I built this system, and they are the ones your counsel will ask about. Get a telecoms lawyer to review your specific setup before you dial the first number — this area moves, and the penalties are per-call.
Why AI voice is treated differently
The critical thing to understand: an AI-generated voice is not treated like a human making a call.
In February 2024 the FCC ruled that AI-generated voices in calls count as “artificial” under the Telephone Consumer Protection Act. That single word does a lot of work. Artificial and prerecorded voice calls sit under stricter consent requirements than a human dialling manually — and TCPA damages are assessed per call, which is what turns a configuration mistake into a five- or six-figure problem rather than an apology.
So “we’ll just point it at a list and see what happens” is not a strategy. The consent basis has to be settled before the system exists.
The five things that have to be built in
1. Suppression and DNC scrubbing
Your agent must check every number against, at minimum:
- The National Do Not Call Registry
- Your own internal do-not-call list
- Anyone who has previously asked to stop — captured automatically from call outcomes, not from someone remembering to add them
The internal list is the one that gets neglected, and it is the one that matters most. If a prospect tells your AI to stop calling and the same agent dials them next week because a human never transcribed the note, you have a documented, repeated violation.
Build the suppression write-back into the call outcome handler. Not a spreadsheet. Not a process. Code.
2. Calling windows in the prospect’s timezone
Calls are restricted to reasonable local hours — generally 8am to 9pm in the called party’s timezone, with state rules that are stricter in places.
The failure mode here is subtle and very common: your system knows the area code, and assumes the timezone from it. Mobile numbers move. A 212 number can sit in California. If your window logic keys off area code alone, you will eventually place a 6am call and have no defence.
The system I run treats the window as a hard gate evaluated per call, with a conservative default when the timezone is uncertain — it would rather skip a call than place a bad one.
3. A consent basis you can point to
Before dialling, you need an answer to: what is our legal basis for calling this number? Different bases carry different obligations, and B2B calling to a published business line is not automatically exempt — the analysis depends on the number, the relationship, and what you are calling about.
Whatever your basis is, record it per record, at the time of ingestion, with a timestamp and a source. When someone asks, “why did you call me?”, the answer has to be a database row, not a recollection.
For the FMCSA feed, that means the ingestion pipeline stamps every authority with where it came from and when — because a defensible answer six months later depends on data you captured on day one.
4. Disclosure that the caller is AI
Do not try to pass the agent off as a person. Beyond the growing set of state rules requiring disclosure of artificial voices, it is operationally self-defeating: prospects work it out, and the ones who feel deceived are not the ones who book.
Disclose early and move on. In practice it costs less than people fear — a broker who is going to engage will engage anyway.
5. A human transfer path
There must be a way for the person on the phone to reach a human, and it must work. This matters for compliance, and it matters for revenue: the calls that most want a human are the ones closest to buying.
The operational guardrails
Compliance keeps you legal. These keep you solvent.
Budget caps. An AI agent bills per minute of voice and per token. A prompt loop or a bad list can spend a lot of money quickly. Set a hard daily cap that stops dialling, not one that sends an alert.
Capacity planning. Booking more demos than your sales team can take is not a win — it burns the leads you paid to generate. Pace the dialler against real calendar availability.
Recording rules. Several states require all-party consent to record. If you record, your consent language and your storage both need to reflect the strictest state you operate in.
Outcome logging. Every call needs a durable record: number, timestamp, consent basis, disposition, and the recording or transcript if you keep one. If you cannot reconstruct what happened on a specific call eight months ago, you cannot defend it.
What good actually looks like
For the campaign I run, the guardrails are not a settings page — they are gates that run before every single call:
- Suppression and DNC scrubbed against a list that updates from call outcomes automatically
- Calling window evaluated per call, conservative when the timezone is uncertain
- Consent basis recorded at ingestion with source and timestamp
- Budget cap armed, with dialling stopped at the ceiling rather than alerted on
- Human transfer available on every call
- Full outcome log written for every attempt
It lives inside the same TMS the sales team already works in, so a booked demo lands where the rest of their pipeline lives instead of in a separate tool nobody opens. You can read how that system was built.
Questions to ask any AI calling vendor
If someone is pitching you one of these, these five questions will tell you quickly whether they have run one in production:
- How does the internal suppression list get updated? If the answer involves a person, walk.
- How do you determine the called party’s timezone? “Area code” is the wrong answer.
- Where is the consent basis stored, per record? It should be a field, not a policy document.
- What stops runaway spend? A hard cap, not an alert.
- Show me the outcome log for a single call from six months ago. If they cannot, you have no audit trail.
The honest summary
AI cold calling works — the agent I run books real demos from a cold FMCSA feed. But the part that makes it viable is not the voice model, which is close to a commodity now. It is the boring layer underneath: suppression, windows, consent, caps, logs.
That layer is roughly where all the engineering effort actually goes, and it is exactly what gets skipped in a demo.
If you are considering one for your operation, tell me about your list and your consent basis — that is the conversation that determines whether this is worth building for you, and I will tell you if it isn’t.