AI voice agents on a sales line

Ask first which calls a voice agent may touch. How human it sounds matters much less. A team that answers that question narrowly gets a useful tool. A team that answers it broadly gets a system talking to buyers about money with nobody listening.

Inbound: only the calls nobody answered

Every sales team has calls that go nowhere. Sunday at two. Tuesday at 6:40, when both reps on shift are already on calls. The third transfer of a call that started in billing. The line that rings out because the person at the front desk stepped away. Nobody handles those calls today, well or badly.

That is the right place for a voice agent, and it is narrower than most deployments choose. The rule is simple: if a person on your team was free, the person takes the call. The agent covers overflow and after hours, and you should be able to show that from the routing settings rather than from a promise.

Scoped that way, the agent has a short job. Get the caller's name and number early, because everything else can be recovered from those two. Answer the question that was asked, if it is factual: hours, whether an offer is still available, what the next step involves, what information a quote needs. Offer a specific time for a callback or a meeting. Write what happened into the record so the person who follows up starts with context.

What the inbound agent should not do

• Take a call a person could have taken. Overflow is a routing decision, so set it in the phone system instead of assuming it.
• Negotiate. Price, payment terms, discounts, fees. Any figure that moves is a person's job.
• Commit. Holding a price, promising a start date, confirming an approval, guaranteeing a quote.
• Keep going when the caller is confused, upset or asking for a person. All three are handoff triggers.
• Collect payment or account details. Some businesses, such as mortgage lenders and finance companies, fall under federal rules on safeguarding customer information. Decide on purpose what any automated channel may touch.

Outbound: draft, then approve

Outbound is where voice agents go wrong, because volume is easy and judgment is hard. The design that holds up is draft and approve. The system assembles the call list and the reason for each call. A manager reads the list, removes the names that should not be called and releases the rest.

Approval costs a few minutes a day, and it is the only version where you can answer the question a customer will eventually ask: why did your system call me? “Because a rule fired” is a weak answer. “Because a manager reviewed the list on Tuesday morning and approved it” describes a process you can defend.

Approval also catches the cases rules never catch cleanly. The customer who signed with another company last week and told you so. The one in a billing dispute. The one who asked to be left alone, three days ago, in a text the calling system never saw. A manager scanning a list catches those in seconds.

What the federal rules say about an artificial voice on the phone

This is general information, not legal advice, and the details depend on the facts and the state. Have your own counsel review any voice deployment before it places a call.

The Telephone Consumer Protection Act, at 47 U.S.C. 227, restricts calls made with an automatic telephone dialing system and calls that deliver a message using an artificial or prerecorded voice. The FCC's rules appear at 47 CFR 64.1200. In a declaratory ruling adopted on February 2, 2024 (FCC 24-17, CG Docket No. 23-362), the FCC confirmed that those restrictions cover current AI technologies that generate human voices. Calls that use them need the prior express consent of the called party unless the call is for an emergency purpose or an exemption applies, and telemarketing calls of this kind need prior express written consent.

The rules also carry identification duties. Under 47 CFR 64.1200(b), an artificial or prerecorded voice message must state clearly, at the beginning, the identity of the business responsible for the call, and must state a telephone number for that business during or after the message. Telephone solicitations to a residential subscriber may not start before 8 a.m. or after 9 p.m. local time at the called party's location. The FTC's Telemarketing Sales Rule sets the same calling window at 16 CFR 310.4 and carries do-not-call obligations of its own.

Three consequences for a sales team. A customer who called you is a different consent situation from an outbound call you place to them, so treat the two separately from the first day. Whatever consent your web forms and texting programs capture is a question for counsel, not for a vendor's marketing page. And state law can go further than federal law on recording, on disclosure and on automated calls; the AI caller disclosure by state tool summarizes each state as a starting point.

Disclosure, in practice

Beyond what the rules demand, disclosure decides whether customers trust your company afterward. Three behaviors hold up.

1. Answer the question honestly, right away. When a caller asks whether they are talking to a person, the agent says it is an automated assistant, in plain words, without a deflection or a joke.
2. Do not pose as a named employee. An agent that introduces itself as a person on your staff creates a problem the day the customer asks for that person.
3. Say what happens next. A caller who understands that a rep will call back in the morning behaves differently from one who thinks the matter is settled.

Start by counting what you are missing

Teams often buy a voice agent before they know the shape of the problem it should solve. Spend a week finding out first, because the answer changes what you buy and sometimes removes the need to buy anything.

Pull your inbound call log and count unanswered calls by hour and by day of the week. Separate three groups that look the same in a report and differ in practice: calls that rang out when nobody was working, calls that rang out while people were free, and calls that were answered but dropped on hold or lost in a transfer. The second and third groups are staffing and process problems, and no agent fixes them. The first group is the window a voice agent is for.

You may find that most of your missed calls fall in a handful of hours each week. That is a smaller and safer deployment than answering everything, and it is the one to start with.

Permissions are the real specification

Whatever an agent is technically allowed to do, it will do, at three in the morning, on the call you did not expect. So write the permissions as a list and treat that list as the product.

• Which lines and hours it answers, and the routing rule that puts a person first.
• What it may state as fact: hours, location, what is available, how the process works.
• What it may never state: prices, payment terms, discounts, approval or availability guarantees.
• What it may write into your systems, and what it must leave for a person.
• Which conditions force an immediate handoff or a callback commitment.
• What happens when it fails: a person, a voicemail box someone checks, and a call in the queue the next morning.

What to do before you turn anything on

1. Count your missed calls by hour and day, so you know the size of the problem.
2. Set routing so a person is always offered the call first.
3. Write the permission list and the handoff triggers before you write the script.
4. Ask counsel about consent, disclosure and recording in the states you call.
5. Listen to every agent call for the first two weeks.