The Most Successful Enterprise CX Teams Treat AI As A New Hire You Keep Coaching

Enterprise AI deployments are stalling out, and it’s usually not the technology's fault. Many organizations try to treat AI like a traditional software release that's built, shipped, and maintained. But an AI agent acts more like a new team member that needs continuous onboarding, coaching, and growth. The resulting friction leaves many enterprises stuck in small, isolated pilot programs. Making the jump to broader business impact often depends on shifting frontline workers from simple operators into active coaches.

Nicholas Galarneau is Vice President of Business Systems at Crash Champions, where he leads the engineering and technology strategy for a multi-billion-dollar enterprise operating within a traditionally non-tech industry: collision repair. Galarneau has spent his career building custom systems and developing engineering teams, and his current work centers on deploying AI into live customer-facing and operational workflows.

"With an AI system, it's not a one-time situation where you build the requirements and walk away. It's a consistent process of monitoring what's going on and continuing to push it. It's just like with people. You don't stop training people," he says. That framing of AI as a team member you keep developing rather than a project you finish runs through everything Crash Champions has built.

The AI agent as a member of the workforce

The clearest example is the customer service AI agent Crash Champions launched in January, which handles roughly 400 to 500 appointment-setting calls a day for its shops. The call center team, a major stakeholder in the project, treats the agent less like a tool and more like a new hire. "They listen to phone calls just like they do with the regular call center, and they coach it. They give us learnings: from this call at this time, here's what was said, here's where it should go. It's really that straightforward of treating it like another member of the team," Galarneau shares.

The agent has an advantage over a human trainee in some respects. It absorbs corrections consistently and applies deterministic rules quickly, so a fix like interpreting "Tuesday at 4" as next Tuesday rather than every Tuesday sticks immediately. But the trust barrier is real, and Galarneau frames much of the early work as helping both employees and customers get comfortable with a non-deterministic system.

The empathy challenge surfaced in ways the team didn't fully anticipate. Because the voice agent is convincing, some callers don't register that they're talking to AI and share personal, emotional details about their accidents. "We've had people share personal life stories with it. We don't try to fake empathy from an AI agent, but we train it to say, 'I'm sorry you had that experience. I'm glad everyone's doing well. Let's get your car repaired quickly and back on the road.'" That balance of acknowledging a human moment without pretending to be human is a deliberate design choice rather than an accident of the model.

Escalation built on severity and outcomes

The expansion of the agent into new use cases, like giving customers status updates on their vehicle repairs, depends on a governance and escalation structure that decides when the AI acts and when a human steps in. Galarneau's team runs a continuous evaluation loop keyed to severity and outcomes rather than raw activity metrics. "We give it a rating of A, B, or C," he explains. "In the A-B range we don't typically intervene. If it's in the C range, we put that under watch. Anything below that we consider a failing grade, and we go in and do a training cycle."

That training involves finding similar calls or emails and giving the agent more context and direction on how to handle comparable situations going forward. The escalation line isn't fixed. It shifts as the team gains confidence in what the agent can handle well.

Measuring sentiment and outcomes, not handle time

Crash Champions deliberately moved away from evaluating the agent on traditional operational metrics like handle time and first response. Those numbers still exist, but their role has changed. "Those traditional metrics are turning into operational support metrics rather than performance-of-outcome metrics. Wow, that call went really long, let's look at what happened. And often it's a bug or the system got into a weird loop, a technical issue, not an AI context issue," Galarneau says.

The primary measure is whether the customer got the outcome they needed, captured through a simple post-call survey and a sentiment model that reviews the interaction. When a customer indicates the outcome was not met, a human listens to the call, exactly as they would for a human rep, to understand the gap. Galarneau is candid that some gaps are just communication issues that occur with human agents too, or customers who were not actually ready to book.

The nuance he flags is that in a business built on accidents, sentiment is hard to read cleanly. It's not always obvious whether a customer is upset about their damaged car or about the interaction itself, which is why the team leans on human review rather than automated scoring alone.

The loop runs in both directions

The most compelling part of the model is that AI doesn't just get better through human coaching. Humans get better through what the AI surfaces. Crash Champions' SMART Review project turns an LLM against the company's own estimate data to find patterns where estimate-writing is causing performance gaps. "A lot of that project is based on finding where there are gaps in our human efforts and how we can train and improve those processes," Galarneau says.

Rather than only feeding fixes back into systems, the tool generates coaching for the service advisors who write estimates, flagging where additional evidence might be needed or where a repair time looks out of line. The behavioral change has been measurable within three months. "We're not finding as many of those recommendations happening with repeat offenders. They started building that muscle memory. Previously the only way we did that was pulling a couple of estimates for review every once in a while. Now we can review every estimate that's written."

That shift from periodic spot-checks to review of every transaction is what turns AI insight into durable human improvement.

Business owns it, IT enables it

The throughline across both projects Galarneau describes is ownership. He credits the success to engaged product owners who sit at the business unit level, not in IT, and who participate directly in the development lifecycle, including standups and retrospectives. "We're not getting there without the business owners being that engaged and taking the outcome and success on their shoulders as much as IT does."

He's direct that the projects that stall are the ones treated as IT projects. The winners are owned by the call center, the review team, or the training team, with IT as a heavily involved enabler rather than the accountable party. The business holds the domain expertise that defines the requirements, and when that ownership is real, the loop between machine improvement and human improvement actually closes.