Agent-Led AI Pilots Help Contact Centers Move From Tool Testing To Real Efficiency

Contact center teams are spending heavily on AI, and many are still waiting for the efficiency gains to arrive. The operators seeing results work a particular way. They run small, low-commitment experiments and expand the ones their agents adopt on their own. Most rollouts that stall do so in the change management and frontline habits around a tool, well before the technology itself gets a fair test.

Jeremy Hyde is the Senior Director of Customer Service at Sun Country Airlines, where he runs a fully remote contact center supporting travelers across voice and SMS. He is also the Founder and President of the WFH Alliance, a community where contact center leaders compare notes on remote staffing and AI adoption. With nearly two decades in the field, including customer service leadership in health insurance before the airline, he reads new tools as an operator by watching what shifts on the floor once his agents are using them.

The hardest part, he has found, is rarely the tool itself. "The gap is not in the technology. It's in the company's ability to give the technology what it needs to operate," says Hyde.

Stalled savings

He ran into that early, on an auto-summary tool his team rolled out to write agents' after-call notes, a change they expected to trim handle time the moment it switched on. The savings didn't arrive, and the reason had nothing to do with whether the tool worked.

The notes themselves were fine. What the tool couldn't do was change how agents spent the time it freed up. For many agents, after-call time was simply time they always got, and they kept spending those seconds the way they always had, so handle time held steady. Hyde's team reset expectations, presenting the saved minutes as a reason to move to the next contact sooner. "You don't have to think about note-taking anymore, so use those notes and move on to the next thing," Hyde says. Once the behavior caught up, the efficiency the team had projected began to show.

That correction points to how Hyde runs AI more broadly. He favors short, low-stakes pilots that his team can stand up quickly and judge by what agents do with them. Testing several ideas this way shows what works before the company invests further. The method surfaces problems while they are still cheap to fix. The episode taught him that the agents who live with a tool belong in the room when it gets chosen.

Empathy on tap

One pilot has cleared that bar. His team connected an internal chatbot to the company's knowledge base, and it has become the way agents find answers. An agent asks the bot a plain-language question and gets a plain-language answer back, with a link to the source, replacing the keyword search that used to return a long list of articles to dig through. The friction it removes is the kind that agents feel on every shift, which is why they take to it without being pushed.

The use that surprises Hyde most is one his team never designed for. Agents turn to the bot in tense moments, asking it how to word a reply to an upset customer. "They'll explain that a customer is really upset, give the answer, and ask what they should tell them, and it gives them some nice, empathetic-sounding language," Hyde says. A tool built to surface knowledge has quietly become a coach for the conversation itself. A separate pilot in its first weeks extends the same idea to SMS, letting agents dictate a reply and drop in approved policy language without retyping it.

Written for humans

The chatbot earns its keep because a clean knowledge base sits behind it, and that dependency is the part teams tend to underestimate. When a customer-facing system comes up short, the cause usually traces back to the material feeding it. Most contact centers write their procedures for people, leaning on the judgment a trained agent already carries and skipping the steps too obvious to spell out. A system reading those same procedures cannot supply what they leave out, so it inherits every assumption a person would have filled in. "With technology, it's like training a new person. You can't make those same assumptions," Hyde says.

Closing that distance takes more than installing the software. It means rewriting reference material in full, mapping the decisions a seasoned agent makes on instinct, and building the APIs that pass the right information to the system at the right moment, work that few operators have spent the past decade preparing for. That readiness decides whether a tool that impresses in a demo can carry live volume. Teams that build it into the deployment from the start, before the need surfaces halfway in, are the ones whose pilots hold up as they expand from a handful of users to the whole floor.

Off script, on message

As agents lean on these tools, they raise a question every service leader knows. When AI helps two agents work through the same issue, do their answers still sound like one company? Hyde sees an opening more than a risk. Lightweight prompts can steer an agent toward the right framing at the right moment while leaving the human conversation intact, the consistency that heavy scripts once promised without the rigidity that made agents resent them.

What Hyde keeps returning to sits further out, in resolution shaped to the individual customer. Most service models still run on one-size-fits-all rules, even when the same event hits each person differently. He expects AI to keep moving closer to the customer as teams earn trust in it, eventually weighing who someone is to the business and what makes their situation unique before pointing them to an answer. "Everyone on that flight might be experiencing the same delay, but the impacts on the individuals are different. If I'm going for a leisure trip and can't check into my hotel until 4 p.m. anyway, a delay is inconvenient, but it might not be the end of the world. If I'm going for a job interview, that's a very different scenario. Designing flexible, contextual experiences is where the really interesting opportunities are," Hyde says.