Guardrails, not rules - TruGreen's approach to agentic CX with Qualtrics

Last week at X4, Qualtrics was promoting customers live on its Experience Agents and as I wrote in my piece looking at some of the harder questions around Qualtrics' agentic AI strategy, it’s the implementations and customer use cases that will either validate or discredit what’s being proposed. 

With this in mind, I sat down with James Bauman, Senior Director of Experience, Analytics and Insights at TruGreen - one of the first companies to get Qualtrics’ ‘Experience Agents’ live, with some impressive results being seen. In conversation with Bauman, we get into the detail of what deploying Experience Agents actually looks like - the design philosophy, the first-day reality, and what the limits of the current system reveal about how far agentic CX has genuinely come.

Starting from scratch

A little context on TruGreen first. The company is the largest lawn care provider in North America, with close to $2 billion in revenue and 2.5 million customers. It is largely subscription-based - customers pay for the year, technicians visit on a set schedule - and it operates nationally and in Canada. The business has a significant number of customer touchpoints and, by extension, a significant number of ways in which things can go wrong.

When Bauman joined TruGreen, there was no centralised CX team. What existed was a transactional survey programme with data fed back to local branches. He built a central function from the ground up and, about two to two-and-a-half years ago, partnered with Qualtrics to build something more comprehensive. This new platform has surveys across the website and mobile app, post-service transactional surveys, a relationship NPS programme, and - notably - 100% transcription of phone calls across both sales and customer service. The full-call transcription gives TruGreen a real-time view of cancellation intent and the specific drivers behind it that surveys may find hard to capture.. Bauman describes it as giving the organization "a really useful mind map of what all the different things are that customers are talking about."

This foundation provided the base for a move to agentic AI. 

The 40% problem

The key driver for deploying Experience Agents was that TruGreen's traditional closed-loop process for handling customer feedback was manual and labour-intensive. But the company learned that education could play a helpful role. Bauman told me:

About 40% of the time, resolutions are just about educating the customer.

That means that nearly half of all the escalation and follow-up activity passing through TruGreen's human teams was not actually about resolving a service failure - it was about setting expectations correctly in the first place and providing support materials. If you can address that at the point it arises, you free up significant capacity and, arguably, do a better job for the customer, because the information is reaching them when it is top of mind rather than hours later via a phone call they may or may not pick up.

This thinking led TruGreen to five or six primary education use cases built into the Qualtrics survey flow - covering lawn care maintenance, watering guidance, and localized branch-level information. The agents draw on contextual data, including weather conditions, to personalize responses at a level that would be impractical at scale for a human team.

Education, education, education

The most mature use case Bauman walked me through involved weed complaints. Weeds are, unsurprisingly, a major reason people hire TruGreen in the first place. When a technician has just visited and a customer reports in their post-service survey that weeds are still visible, the agent steps in with context the customer might not be aware of. For example, when the last service was, what was applied, and - importantly - that herbicides typically take around ten days to work visibly. The agent commits to a follow-up check-in at that point.

Ten days later, an automated survey goes out. If the customer says weeds are still a problem, a high-priority ticket is created for the human team to investigate further. What the human agent receives at that point is not a cold alert, it is the full context of the interaction. The human agent knows what the customer said initially, what the agent responded, the fact that a follow-up was triggered, and the data point that the customer remains unsatisfied. Bauman said: 

It helps them be more relevant and more efficient in their outreach when it does get to them.

The helpfulness score for the weeds flow sits at around 60 to 65 percent. Early data on retention impact - with the caveat that it has only been running for about a month - shows something around an 8% uplift. These are early numbers and Bauman is appropriately careful about overstating them, but the directional signal is clear.

The first day was not great

Helpfully, Trugreen wasn’t shy of getting into the weeds of the implementation (sorry about the pun). TruGreen went live at full national scale from day one - a deliberate decision, Bauman explained, because it meant a meaningful sample size immediately and the ability to turn it off and revert quickly if needed. The first day's satisfaction rate with the agent was around 15 percent. The responses felt, as Bauman put it, almost bot-like: 

We tend to want to control things as much as possible. A lot of the power from these AI agents, I believe, comes from giving them a bit more freedom. But you have to do it carefully - you have to think carefully about your guardrails, and not try to over-programme the agent.

When we first launched, it felt almost like a bot response: the exact same line every single time, and a lot of the time that doesn't feel quite relevant - or there are pieces of it that feel like, oh, you didn't really read what I said. So that's a bit of the art of it: how do you create guardrails, not rules? That's my big piece of advice.

Within hours, Bauman's team was looking at live responses, editing agent instructions, and relaunching iterations - all on day one. By the end of the first week, performance had improved significantly, and the 30 percent-plus escalation reduction figure that Qualtrics cited publicly was real. But it’s helpful for others to know that this was not an out-of-the-box solution and did not arrive pre-packaged.

Also, it’s worth noting that Qualtrics was heavily involved. As I noted in my X4 piece, Qualtrics' own product team helped build this implementation, and the vendor acknowledges it was high-touch. This is not a criticism, it’s just useful for other buyers assessing the benefits Trugreen saw. The out-of-the-box narrative that tends to accompany agentic AI announcements does not quite describe what happened at TruGreen. 

The operating model question

I was also curious to ask Bauman how the use of agentic AI might be changing TruGreen’s operating model. He noted that the introduction of agents is already beginning to shift how TruGreen's customer service function thinks about its work. Bauman said: 

Our customer service team's work is shifting to be more proactive and focused on really engaging and building relationships with customers, rather than just reactively solving issues as they come in.

That is a meaningful shift - but, at this point, it’s a shift in direction of travel rather than a completed transformation. The most obvious next step would be allowing an agent to schedule a free service visit directly when a customer remains unsatisfied after the weeds follow-up. However, this has not happened yet. The CRM integration exists on the survey data side, but it does not yet extend to triggering actions in operational systems. Bauman said: 

We don't feel comfortable enough allowing the agent to schedule services for the customer. But that's the primary use case I see.

The barrier here is not technical, it is organizational. As Baumann said: 

 If it were just up to me, I'd have done it already. It's a very natural next step.