A bot gets through a scripted call without sounding like a bank’s phone menu, and suddenly we’re putting microphones on pricing pages and inviting it to pipeline reviews. Somewhere in that meeting, somebody should ask what problem we just spent money on.
I build AI employees at Atonom, so yes, I have a dog in this fight. I also have to sit across from people who want to know what they’re getting for their money. “It sounds so human” is fun to hear in a demo. Try putting it in the ROI spreadsheet.
I wanted to look at this job by job, because “does AI voice work?” is too broad to be useful. Calling someone who asked for a quote and cold calling someone whose number you bought both show up as outbound calls. Anybody who’s carried a quota knows those are two very different conversations.
Where AI voice is working
1. Customer service
If I’m calling because my order never showed up, you’ve already got my attention. What I need is for somebody to look up the order, tell me what happened, and fix whatever can be fixed. If the AI can do that while your team is tied up, great. If it asks for my order number, reads me the help article I already found, and transfers me somewhere I have to give my order number again, you’ve built a phone tree with acting lessons.
Before you brag about how many calls the AI handled, pull the repeat calls and see how much work landed back on your team. The customer who hung up and tried again tomorrow probably looks like two successes in somebody’s dashboard.
2. Receptionist and front desk
A customer calls at lunch, the front desk is slammed, and after six rings they call the business down the street. Then Monday rolls around and somebody suggests spending more on lead gen.
This is about as straightforward a job as you can give AI voice. Give it your actual hours, a calendar it can book into, and enough information to route the caller or get an urgent issue to the right person. If someone leaves a message, make sure it lands with somebody who’ll actually call back. You can spend three weeks debating whether the receptionist sounds like your founder. I’d spend that time finding out how many calls you’re missing and what those callers wanted.
3. Inbound sales
One person calls to ask if you have a part in stock. Another fills out a form asking for a quote. Both are trying to give you money, and somehow we still make them work for it.
Pick up the sales line or call back while the buyer still remembers filling out the form. Answer the pricing question, check inventory, get what you need for a quote, or take the order. If they need a salesperson, pass along what they already told you so they can pick up where they left off. And if your AI asks for annual revenue before it’ll tell someone whether a part is in stock, you’ve trained it to be your most annoying SDR. Your CRM’s empty fields are your problem.
4. Recruiting coordination
A candidate applies on Monday. By Thursday, three people have emailed them about availability and there’s still no interview on the calendar. Everyone is doing their part, apparently, except the part where somebody books the interview.
An AI can call the applicant, get the missing information, confirm they have the required certification, and book the interview while they’re on the phone. It can explain what happens next and update the recruiting system so the next person has the details. If the job requires a commercial driver’s license, ask whether they have one. If you’re scoring “leadership presence” from a six-minute phone call, you’ve given a coin toss a headset. I’d measure how long qualified candidates wait for an interview and how much chasing the recruiter still has to do.
5. Appointment management
We ran a test with a sales team that was getting demos booked but only seeing 54% of buyers show up. We had an AI SDR call buyers before the demo to confirm the time and reschedule on the spot if they couldn’t make it. The show rate went from 54% to 63%. That’s nine more demos held for every 100 booked.
I like this one because it’s almost offensively boring. The buyer already wanted the demo. Their Thursday got screwed up, and now they can pick another time before your AE spends ten minutes staring at an empty Zoom room.
6. Outbound customer engagement
We had one team use AI voice to call existing customers and book quarterly business reviews. Another used it to reach out to contractors who already bought from them and see if they needed help with an upcoming job. If a customer says yes to a quarterly review, book it while you have them. If a contractor needs help figuring out what to order, get them to someone who can help and pass along what they already told you.
Same goes for onboarding or maintenance. I’m happy to get a call if I’m halfway through setting up something I paid for and you can help me finish, or if my car is due for service and I can book it while we’re talking. Even a feedback call can be worth picking up if you have something specific to ask and somebody’s going to do something with the answer.
The call should make sense to the customer before it makes sense on your activity report. Buying from you once does not mean I want a standing Tuesday call with your software.
Where AI voice is not working
1. Website voice agents
I understand why this looks good in a demo. I’m less clear on why I’m supposed to turn on my microphone at work to find out what your software costs. People browse websites while they’re sitting next to coworkers, half listening to a meeting, or comparing three vendors in three tabs. Reading and typing work pretty well in all three situations. An optional voice button can be useful when somebody wants it, but if your pricing page takes longer to get through because it now wants to discuss my needs, you’ve managed to give a bad page a sales quota.
2. Mass cold calling
Every pitch for AI cold calling eventually gets to the part where we hear how many calls it can make. The part I want to hear is what it’s going to say to the poor person who answers.
If your list is bad and your pitch could apply to anyone with a LinkedIn account, making the calls cheaper buys you more chances to annoy the wrong people. I’d look at qualified opportunities, meetings that actually happened, complaints, and revenue before getting excited about the dial count. A million calls is an impressive number to put on a slide when the pipeline number is too small to read from the back of the room.
3. Live meeting participation
I like AI in meetings when it takes notes and saves me from trying to remember who volunteered to do what. Giving it the floor is a much tougher job. It has to know who the buyer is talking to, whether a question has already been answered, and whether that little pause is an invitation to speak or just somebody taking a breath.
Imagine a buyer finally explaining why the last rollout went sideways and the bot jumping in to clarify your integration options. It might even be right. I’m still muting it. My test would be whether the people on that call ask to have it back, especially the ones who had to listen to it.
4. Voice-driven reporting and analytics
If I’m in the car and want to know what’s in the pipeline this quarter, I’ll happily ask out loud. If I’m trying to figure out why midmarket conversion dropped, I want the chart, the underlying deals, and a chance to see what somebody decided to exclude from the calculation.
By the time the AI finishes reading me the enterprise numbers, I’ve forgotten the midmarket numbers and opened Salesforce anyway. Voice is a perfectly good way to ask for the report. I’d like to see the report before we put the bot in charge of reading the whole thing to me.
5. Role-play and coaching
I love the idea of AI role-play. Reps can practice tough objections, support agents can rehearse angry customer calls, and managers can try giving hard feedback before they have to do it for real.
The problem is getting people to use it twice. Everybody tries it when their manager puts it on the calendar. Then Tuesday gets busy and nobody’s volunteering to have a pretend argument with a computer.
Closing Thoughts
I’d spend money on AI voice where there’s already a good reason to talk and the AI can do something about it. A customer needs an order fixed, a buyer needs to move a demo, or a recruiter needs to get an interview on the calendar. Those are jobs you can hand over and check afterward.
The shakier bets ask people to do something they had no interest in doing before the demo. Talk to a pricing page, sit through a cold pitch, or spend Tuesday afternoon arguing with a pretend customer. Getting the bot to sound more human won’t make any of that more appealing.
Before you renew the pilot, pull the numbers you cared about before the vendor showed up. Did more buyers show up? Did fewer customers have to call twice? If the only number that went up was how much your software talked, keep your money.