One of the big challenges is that we start with a problem and then figure out the right technology or solution to address it. This started around opportunities that we felt were big, around improving the patient experience for colonoscopies in particular. And the challenge around colonoscopies is that there’s a pretty involved process. Actually, there are all sorts of caveats if you have other diseases or symptoms. There are different drugs and different patterns you need to go through in the preparation process, which starts seven days in advance. And what was happening was that a very large percentage of procedures had to be canceled before the procedure, because we found that people hadn’t followed the protocols and the right processes leading up to preparation. Twenty percent of all colonoscopy procedures were getting canceled, and that’s sort of like the financial equivalent for us of a plane traveling with an empty seat.
So there were a number of steps we took, with the Artera solution being part of that, with the idea being to help educate patients better, make sure they were following the right protocols, and answer their questions. How do we make sure that the process, which is complex and confusing, can be as simple as possible for them, so we can increase adherence and reduce the number of last-minute cancellations due to preparation issues?
How was the AI customized to your needs?
Eric Steinberger: I wouldn’t say that what we employ is agentic AI, in the sense that an agentic AI is a self-learning process. We have an agent-based AI, and part of the solution and part of what allowed us to overcome and get our broader organization really comfortable from a quality, safety, compliance, and adherence standpoint was that we created and trained this AI in a walled garden. We mitigated the risk that it would go off, learn something on its own, and then go tell patients about that thing we didn’t fully understand or fully control.
Tom McIntyre: This is the start. It was an AI that was basically trained on Atlantic Health’s own knowledge base, approved FAQ, and library. Their clinical and operations team defined what the AI can and can’t say. It wasn’t an agentic AI in the sense that it completed the action. It didn’t take an action and actually schedule the patient; it was there, trained to be able to answer the questions, walk them through the procedure that is complicated and can be customized and nuanced, as Eric mentioned. It’s just the starting point. And as we build this, we would look to add agentic capabilities so that it can actually complete the action itself. We always like to start with our customers in these spots where we can basically extend staff capabilities.
Eric Steinberger: I feel like Tom’s underselling it. Even the capabilities that we had in these initiatives were great for us, in the sense that, as an AI agent that could speak with tonality, it could handle questions that are phrased in all different types of ways. The mitigation is on any ability of the AI to provide information or facts it would have concluded on its own, which we may not trust for accuracy because it might have pulled information from outside the wall. It was a great partnership that we could get to that solution.
Could you talk about the outcomes?
Eric Steinberger: The goal here was to improve the customer experience. But even more importantly, can we save some hard costs here? In the first six weeks of the pilot, 44 percent of patients answered the AI agent’s calls. Each of these is a confirmed appointment. The first objective was just getting the appointment confirmed, which reduced by 40 percent the number of outbound calls that agents needed to make. In terms of answering questions, 18 percent of the people are speaking to the agent, asking questions, and getting answers.
Different patient offices and doctor-physician offices across the state of New Jersey were making a handful of phone calls manually to these people. Now we have recorded scripts of all of that interaction. Now we add that to the AI agent’s scripting.
What challenges did you come across?
Eric Steinberger: I think the only challenge is the newness and the uncertainty that come with being the first of an initiative. You’re discovering things that you know you hadn’t necessarily thought about. I think because of the rigor of how we approached it, that stuff came up. When you’re doing a pilot, you always have some protections in place.
What future developments are you considering?
Tom McIntyre: This is an example of, kind of a broader model of what we would love to continue to do work with Atlantic Health on. We extend the clinical capability. We don’t replace it, but we’re doing this across use cases and scheduling. AI, especially agentic AI, is so great because it removes parts of people’s jobs so they can actually spend time on care, giving the human touch that people want and crave. It’s finding those other areas to work with, with our partners, whether it’s scheduling, the procedure prep, follow-up, or running other additional campaigns for marketing and outreach to the patient population.
We’re going to be at a point where people are going to actually prefer to talk to AI because they have answers. They have it right away, and it’s on their time.
