The linguist who designs how business phone bots talk, and why less polish works better

The failure behind the zero key

Anyone who has called a company and heard a cheerful, overly tidy greeting knows what happens next: you start pressing zero to reach a person. As poly.ai reported in a profile of its own design lead, the agent has often given the right answer in these cases. It simply sounded like it was reading from a brochure.

Oliver Shoulson is the person PolyAI pays to worry about that gap. His title is Agent Design and Engineering Lead, and he designs conversations that large companies put in front of millions of callers. He spoke about the work on the Digital Project Manager podcast.

His background is in language. At Yale he studied syntactic variation, meaning the small differences in how people build sentences within the same grammar rules. He also works as a visual artist. He says a linguist's habit is to question his own ear: why does this sentence sound right and that one sound off? Then you write the rule clearly enough that a machine could follow it.

The goal is less effort, not a perfect fake

Shoulson doesn't define success as realism. "No one's getting gold stars for tricking anyone," he says. What he wants is to spare the caller's mental energy for the actual problem.

His example is the phone menu. When you face a list of options, you end up guessing how the designer would file your issue. Is it an account problem or a billing question? That guessing burns attention before you've even described what is wrong. A well-designed agent lets you just talk the way you normally would.

That doesn't mean copying every human habit. He notes that natural speech has limits, and an agent should be able to stand firm when a caller turns abusive. The aim is to support good conversational patterns, not to imitate all of them.

Why a "helpful" bot sounds patronizing

The large language models behind modern agents (AI systems trained on huge amounts of text) tend to be agreeable, careful and thorough. On a phone call, those traits backfire. The agent explains things nobody asked about, walks through steps the caller already knows, and confirms everything twice.

Shoulson ties this to how cooperative conversation works. We expect the other person to give us what we need and no more. If someone says more than the task calls for, we read extra meaning into it. Asking an adult to tell you once they have clicked an obvious button hints that you doubt they can find it.

The same goes for opening a simple request with a long explanation of why you need an account number. The agent comes across as distant rather than as a partner. When it trusts the caller with context both sides already share, the exchange feels more like teamwork and less like a script.

Timing details nobody notices

Spoken conversation is fast. People interrupt, correct themselves and overlap, and the gap between turns is often only a few hundred milliseconds. An agent that takes three seconds to answer breaks that rhythm, and the caller senses it before they can say why.

Shoulson describes one early problem he worked on at PolyAI: phone numbers. Callers read a US number in pieces, area code first, then three digits, then four. They expect a small sign of acknowledgment after each piece. If the agent stays silent after "one two three," the caller may stall, unsure anyone is listening. Worse, the agent might treat those three digits as the full number.

The fix was to teach the model to give that quick "I'm following" response at the moment people expect it. This kind of back-channel (the little "mm-hm" that tells you to keep going) is automatic for humans and hard to engineer. When it works, nobody notices. PolyAI says the payoff is more calls resolved without a person and more satisfied customers.

A cloned voice and a day off

PolyAI points to two customers. Fogo de Chão, a Brazilian steakhouse chain, wanted its phone line to feel as hospitable as its restaurants. PolyAI cloned the voice of Selma, a longtime leader on the company's customer experience team, and built it into the agent that handles reservation and rewards calls.

PolyAI reports that this agent achieves 95% customer satisfaction and completes 88% of bookings. Keep in mind these figures come from the vendor's own write-up, not an independent test.

The profile also leans on the idea of social presence, the feeling of being with someone in real time. It says research links that feeling to greater trust, more willingness to follow advice, and more comfort sharing personal details. A voice, the argument goes, can create that feeling more easily than a text chatbot.

The second customer is Carter's. After adopting PolyAI, it could give contact center staff Black Friday off as well as Thanksgiving. Shoulson frames the work as taking on repetitive questions so people handle the calls that truly need them.