The AI Conversation CTOs are Having Behind Closed Doors with Robert Duffy, CTO at HealthEdge

(Intro Narrator at 00:00:01) Today, we're talking to Rob Duffy, CTO at HealthEdge, about what CTOs are saying about AI behind closed doors. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:16) I'm a big fan of the conversations behind the scenes. I like behind-the-scenes content when it comes to movies. And when Josh pitched this to me, he's like, "Hey, we're gonna have an AI conversation about the conversation CTOs are having about AI behind closed doors." I said, "Yeah, yes, sign me up. I want to do that with Rob." So I'm gonna leave it to you because I'm not actively a CTO. I was previously, but for the past seven years, I've just been running the podcast, and that's nice. So what's happening? Give me the insights too.

(Rob Duffy at 00:00:48) What's happening? I'll give you the insights too. I went to a conference a couple of months back, and it was a health care technology conference. And there was a lot of presentations, and as you can imagine, a lot of presentations on AI. And the remarkable thing about that conference is I could have created a little time machine and teleported you to that conference last year and teleported you to the conference this year and challenged you to tell the difference between the demos and the presentations and the talk tracks and the narratives between those two years. And I think that even though the pace of AI development is incredible and moving very, very quickly, the adoption and the things that the use cases that people are talking about and the demos that people are giving isn't keeping up with the same pace. And I think that's largely because everyone now has demo fatigue, and everyone has seen all the demos, and they've seen all the flashy use cases. But people are struggling to take that flashy demo and flashy use case and actually put it into production environments and scale it and get it into a system that's delivering real ROI. And that is what everyone is talking about right now. It's like, we see the potential, we've seen the demos, but what are the actual real world hard ROI use cases that we're able to bring to production? And a lot of us are thinking about that, and a lot of us are talking about that. And I think one of the most interesting things that I've started having a conversation with my leadership team about and my board and our investors is that the migration of work from human cognition to AI and agentic systems is probably about as large as a cloud migration would be in terms of moving things from legacy on-premise infrastructure to cloud infrastructure. But if you think about those two projects inside any organization, the cloud migration has a project management office. It has a cloud center of excellence. You create a cloud factory. You put things in these waves. You construct a wave plan. There's the books and blogs, and AWS has guidance on its well-architected framework, and it has this method of assessing what you're gonna do with all your on-premise infrastructure. You have these seven Rs, and it's like rehost, refactor, replatform, et cetera. And there's this whole big motion around it, and everyone knows how to do that by now because most of us have participated in one, if not more than one cloud migration at this point in our careers. And we sort of know how to do that. But then you look at how people are driving the migration of work to AI and agentic systems, and there's none of that. There is none of the process and the systems and the tools to look at work the way that you have tools to look at on-premise and catalog all the infrastructure. And I think that's why we're seeing lots of demos but not much in the way of hard migrations and hard ROIs to these hard migrations to AI and agentic systems. And what we're thinking about and we're putting in place now is that kind of motion and that kind of structure around the project because it really is as large. When you think about the impact it can have to your organization, it is as large an ROI as a cloud migration might be.

(Joel Beasley at 00:04:10) Wow. Yeah. As you're saying that, I'm thinking of about 15 years ago, I built this real estate type software. We integrated it. It only went into companies with over a thousand agents, so big software. And when we first started going around and talking to these companies, it was so fascinating to me because I was just a software engineer who had built something, right? The only thing they really focused on, like 90% of the conversation was on implementation, customer service, rolling out behavior change. And I was blown away by it. I was like, "We built it. It's there. We just swap the data. We're good." But to get the people to change and do this new behavior turns out to be one of the most challenging things. So we ended up building that company out largely around the implementation and the transition of that project.

(Rob Duffy at 00:04:56) It is one of the hardest things in engineering leadership: changing your team and changing your culture and changing the way work is structured. And moving stuff away from humans and into AI and agentic systems is probably one of the largest scale changes that we're gonna have to levy on the team in the next century. And we need to focus on that and have tools and systems and processes on how to do that. And I think, you know, you asked what the conversation behind closed doors is. That is the behind-the-scenes conversation. Like, how do we move this organization forward? And in this case, there is a really, really strong headwind, which is people are terrified. They're worried that if they do move some of their work over to an AI and agentic system, that means they don't have a job anymore.

(Joel Beasley at 00:05:49) For some of them, that's true. Yeah, it is. But it was that way with SaaS software. You would go into, we would go into teams that had a hundred people in the team, and we would leave after implementation. They'd have 10 people that learned how to use the new tool that did the work of a hundred people. I mean, this isn't new. This has been going on for a long time.

(Rob Duffy at 00:06:07) I know. It's just that tech has always been on the other side of that equation. Right? If you think about software and software developers, it's always been the software developers who are installing software as a service or the software developers who are creating a new tool that's gonna automate things. And the engineering teams are the ones that are building and creating and not the ones that are always seeing the benefits from automation and optimization. And then now they're in this position of, "Well, hold on. If 10, 30% of my work goes away to an AI and agentic system, what does that mean for me? What does that mean for my career? What does that mean for software development?" And we're strongly of the opinion that what it means is that a lot of the work that software engineers are doing is gonna be supervising AI and agentic systems. We still need software engineers, and we still need people to review the code that comes out of these tools. And we need people to think about architecture and design, and we need people to get really, really great at prompt engineering and creating tools and creating MCP tools. And that's the work. The work is gonna shift, and we're gonna see the emergence of these AI-enabled developers or agentically enabled developers. And I think those are the teams that are gonna win, are the teams that can make that shift quickly.

(Joel Beasley at 00:07:31) Well, my friend, good friend of mine, we programmed, you know, 20 years ago, and we still keep up. We were just having this conversation the other day about these, you don't wanna say 10x developer, but basically using these technologies correctly. And so we started both playing with them, just side projects and, you know, just hanging out. "Hey, I did this. Check it out." "Hey, I did this. How did you do that?" We send each other videos and stuff. We've been doing that the past two or three months now. And I'll tell you, I have hard data on it now. A very specific process I would go through that would take me eight weeks and about $7,000 and involve roughly three different freelancers. So eight weeks, $7,000, three different freelancers, I was able to achieve in about three hours. So that's my compression that I'm looking at because for me, that's real. That is real hands-on, and it involved just me working with an AI. And so I said, "Okay, this is happening. It is here today." And it did it, in my example project that I did, it did it better than, I'm a tough person. I have really high expectations. Okay? The bar is high, and it met them. I was like, it had just the perfect amount of incompetence to where I was like, "I can see this is good. This is really good."

(Rob Duffy at 00:08:59) And we're seeing that. We are seeing those proof of concepts and those point demos, and we're seeing things that can be automated and produced in five minutes where it used to take two sprints. We just recently had a, internally, we have to keep up with regulatory updates. How that works is that the CMS, Center for Medicaid Services, updates their regulatory website. We read it, and then a subject matter expert has to kind of codify what that means. And then the subject matter expert delivers a Visio chart showing the logic graph of that update to an engineering team, and then the engineering team has to write that. And it takes them, you know, a sprint or a couple of sprints depending on the complexity. We're now going straight from some of those updates and some of those Visio charts straight to code, fully tested with full test automation based on some of our other regulatory edits. And we're doing that in, you know, an hour and then, you know, plus some review time with the review time baked in. And that's incredible. You're taking something that used to take two sprints, or one or two sprints, and turning it into, you know, one hour, two hour review task. This is really where we're gonna get a huge amount of lift. I think the other thing that we're noticing is that you can get AI or you can get generative AI tools or an agent to write really, really good unit tests. And then you can take those unit tests and get it to write some code to make those tests pass. The thing that we're seeing is that getting an LLM to write really, really good unit tests relies on there being really, really good acceptance criteria and really, really good user stories, and in a consistent format and in a way that that LLM can interpret it and then create those unit tests. Unfortunately, not every development team is blessed with perfect user stories and perfect acceptance criteria. But an LLM can create really, really good user stories and really, really good acceptance criteria, you know, in conjunction with their product manager and can have it in a consistent template that can then be consumed downstream. Then you see the upstream problem is that, well, getting an LLM to produce really good user stories and really good acceptance criteria is that they need a really good product requirements document and really good requirements. And again, not every development team is blessed with perfect requirements and perfect PRDs. However, you can get an LLM to write that. And you can get an LLM to provide it in a consistent way that it can then be consumed downstream by the user story and acceptance criteria writing LLM that can then be consumed by the test automation LLM that can then be consumed by the code writing LLM or agent. So what we're actually seeing is that, you know, point solutions, you can get really good demos in. Right? You can show the agentic system or the LLM writing automated unit tests for the user story. But to scale it, you actually have to embed this AI and these agents all the way along the software development life cycle, right from the point of ideation all the way through to the point of testing. And when you do that, that's when you get scale, and that's when you get real ROI. And that was a discovery we had on one of our pilot programs that we ran a couple of months ago. And the regulatory example that I just gave you is a perfect example of that because it is using LLMs upstream to then hand off information to downstream LLMs in a format that they can understand, and that's delivering, you know, that one day or one hour from two sprints worth of work.

(Joel Beasley at 00:12:39) And these materials, like these frameworks for understanding and implementing this, are you creating internal versions of it as you go and sharing it with your peers? Or is McKinsey creating stuff or Deloitte?

(Rob Duffy at 00:12:51) No. We have a largely internal, grassroots LLM adoption process. And, you know, it started with this pilot, 53 users. We got onto Claude by Anthropic and used their Claude for Business offering. And, you know, we put all those people in their channel. We had weekly leaderboards. We had show backs on data. You know, we created some gamification. One of the things that I did was offer a jar of spicy pickles to a person who produced the best tools. Right? So it was a fun, kind of convivial atmosphere, but we put in a huge amount of energy to it. And, you know, we got people sharing and showing what they had completed and showing what they had done using the tool. And then that provides this sort of collective intelligence. It sparks other people's imagination. Claude has a great project facility where you can sort of create an almost like a mini RAG with the project knowledge and then a mini system prompt with the project instructions. So that then provided a path to production. Right? You have a user who can go in, somewhat motivated by the chance to win a jar of spicy pickles, and they create a project. They put in their information, then they share it and show other users that. And, you know, we have the videos and recording sessions. So you have to put in that much energy and that much momentum to get people really thinking about, like, how can this be injected in my everyday? And how can I personally use it? And how can I share it? And, you know, now we've got AI champions on every team. We give them some swag, and they are the people who are sort of going into those teams, listening for all the updates from all the other teams and sharing that knowledge. And it's a real grassroots movement for us. And for me, you know, I want us to be the most AI-native and AI-literate company in all of health care technology. And I don't think we can do that by relying on external partners to lead that transformation. We have to be the change, you know.

(Joel Beasley at 00:14:55) It has to be baked into the culture because all the small details matter in the time of execution. Like, in the runtime environment of your day-to-day job, you need people that understand what is possible, what the limitations are, all of that. Congratulations on that. So would you recommend this process of getting a bunch of different, were they, let's talk about these 53 users. Different people from different teams? Like, what was it? How did you pick these people?

(Rob Duffy at 00:15:20) We picked five teams. So cross-functional teams. Right? Product managers, subject matter experts, QA folks, developers, managers. We created the team captains.

(Rob Duffy at 00:15:33) We had a kickoff. So basically, the premise of the pilot was if we go all in on AI, how far can we get? We mapped the whole software development life cycle and the whole process. And again, this goes back to my earlier comment around we really need to understand work before we can understand how work can be migrated to AI and agentic systems. We mapped the whole software development life cycle, identified every step, every handoff, how many times teams were doing that, how long it took, and then could that be offloaded to an AI and agentic system. And then identified a proof of concept, AI prompt in just the Claude web UI, and then thought about how can we extend that.

(Rob Duffy at 00:16:15) So the teams were picked because we wanted end-to-end cross-functional reimagination of the software development life cycle. And we ended up with 53 people in there. The energy around it was incredible. As soon as you give people permission to just go end-to-end and tear up everything that they thought they were doing and then point an LLM at it instead, you get huge results out of it.

(Joel Beasley at 00:16:43) Yeah. And so many companies, it's all company culture, right? A lot of them pull back. No shadow IT. But you found these constraints that they could have fun within and then let them kind of do their thing. Now, curious, did any custom software come from that? Like, if you've got 53 people all looking at workflows and processes, identifying what's repetitive, how did you organize all that information into a here's the punch list? We've got 800 things we could do, but these top 80 things are the ones that'll create the most benefit for the time investment. How did you manage all of that knowledge coming out?

(Rob Duffy at 00:17:21) So we left that up to the teams. I mean, we could have put in big program management around it. And, you know, as we start to scale that pilot out, we probably will put some kind of structure around the project and program management. But we basically had a kickoff. We said, this is what we want you to do. Here are the constraints. Here are the parameters. Here's some of the tools that we want you to use. Here's how we want you to approach this. And here's the commitments that we want you to make around going all in on this and really helping share and helping communicate when you have success.

(Rob Duffy at 00:17:49) And then we let the teams organize around their particular part of the software development life cycle. And the there's two modes that we observed. The first is experimenting and just, I'm going to pull up Claude, I'm going to upload a source file into the web UI, which is probably not how you'd ever do anything in production. And then I'm going to ask it a couple of questions and see what I can do. Right? What is the art of the possible here? Here's a Visio file. Is it possible to get from this Visio file to code? I don't know. Let's try it. Right? So there's this experimenting thing. And you have to get people to work in that mode. But then there's this weaponizing—we call it weaponizing of prompts—where you turn them from this experimental chat into a system prompt.

(Rob Duffy at 00:18:32) And in that system prompt, it has things like control of the facts must come from the source file that has to follow this. You put in all of your restrictions and constraints, and you take that experimentation mode into the creation of something that can be used as a tool by other teams. And then you put that into Claude as the project instructions. We share that tool. People hand it out to others. Other people use it.

(Rob Duffy at 00:19:00) And then after that, we take that and we put it into our own internal agentic system that's something that we've developed on top of AWS Bedrock. So there's this path from low-cost experimentation, which gives people the freedom to be completely creative. And it's inside a tool that you can't over-engineer because it's literally a chat interface with the ability to upload some files. So that gives them the ability to experiment. That gives them the freedom to really kind of let their imagination run and try things.

(Rob Duffy at 00:19:32) And then you have a very, very clear path to, okay, let me put this into a tool that I can share with other people. And if they're getting the advantages that I'm getting out of it, then we can talk about how to get from that tool into something that's running in production and being used by the entire software development life cycle.

(Joel Beasley at 00:19:51) That is awesome. So you kind of designed that as a document essentially explaining this process, and then just let them run with it and figure it out. So if somebody has something—let's say one of the teams, give me something real. Let's use that Visio example. They find something and they're like, okay, it can do that. We can get from the what was that word that you used? The changes from the medical company or whatever.

(Rob Duffy at 00:20:18) Yeah. The regulatory changes.

(Joel Beasley at 00:20:20) The regulatory—sorry. There we go. Got it. Regulatory changes all the way through to production. How did that actually roll out?

(Rob Duffy at 00:20:32) So we take that, you take the early prompts, which was literally upload a Visio document into the Claude web interface, and then start asking a bunch of questions around it. And then I think after that, they uploaded a source file example to say, this is what it should look like. And they got to a good place. So you take that, then they created the tool, which is a project as a tool, which has some system prompt information in the project instructions, and then some reference files uploaded to the project knowledge. And then they shared that tool in our channel. We had a chat channel. Shared that tool with the rest of the team. The rest of the team started to use it. We tweaked the system prompt a little bit, tweaked the knowledge, asked it to make sure that it complied with our coding standards, et cetera. And then got to a good place where we felt like, this is good.

(Rob Duffy at 00:21:25) And by the way, if you share a tool and you get two other people to use it, guess what you get? Jar of pickles. So surprising motivational tactic.

(Joel Beasley at 00:21:39) I would be donating them. I would definitely be donating them.

(Rob Duffy at 00:21:43) So then you have that tool. You have your jar of pickles. And—

(Joel Beasley at 00:21:46) Can they be garlic pickles? Because I'm down if it's some garlic pickles.

(Rob Duffy at 00:21:51) You don't like spicy pickles?

(Joel Beasley at 00:21:52) I don't know, man. That's a whole other podcast, I think.

(Rob Duffy at 00:21:58) So then we basically have an AI and ML platform tool. We call them the Work Transformation Team because we think about this as work transformation. And what they do is then they take that system prompt that's inside the Claude tool and turn that into a tool that you can basically take Visio charts and it will produce GitHub pull requests that a human can then go and review based on that upload. So that kind of path to early experimentation or path from early experimentation through to a shareable, reusable tool through to production is super critical because you stop people trying to create the production system and over-engineer it before they've even proven that they can get value out of this use case and this flow. And then when they do get value out of it, they go to the Work Transformation Team and they say, here's the value I got from it. This is a really great thing for you to turn into a production tool and a production system. And that Work Transformation Team, who are a bunch of AI and ML experts and our AI platform team, create the most high-priority use cases that are going to get the largest amount of ROI for it.

(Joel Beasley at 00:23:15) Okay. So the Work Transformation Team is interfacing, and they're part of this group of 53. This group of 53, they're kind of figuring out stuff and having a lot of freedom and flexibility. Then once they feel like, oh, we got something, then they come to the Work Transformation Team and say, hey, this is what we've got. Then they're making sure everything's aligned from technology, security, all of that, and they turn it into this final endpoint that people can use in real-world scenarios of the company. Okay. That's the—

(Rob Duffy at 00:23:43) And they figure out what is the interface to it? Is the interface something running on top of GitHub? Is the interface a command-line interface? Is it going to be used by business people? Does it need to upload user stories to Jira? Does it need to interface with other internal systems? So they figure out the work of taking that thing that probably needs a lot of copy and paste and prompting and turning it into a tool that's integrated with the rest of the system. And that team is responsible for the care and feeding of that tool, too. So you get model drift, you get model updates. Sometimes a prompt that was working really well stops working. And that team is responsible for monitoring the output and then continuously refining it.

(Rob Duffy at 00:24:25) Also monitoring usage. So when they monitor usage, we know just from the intake process how many hours did this save. And when you monitor the usage, you can just multiply the number of times the tool is being used by the number of hours that it saved, right? So now you have some really cool metrics on what is the ROI of individual tools and what is the collective ROI of our agents that we are creating.

(Rob Duffy at 00:24:49) And our long-term goal—and when I say long-term, I mean six to nine months—is to leverage the same platform that we have built internally for our internal users for our customers. So our customers can go through the same process, the Work Transformation Lab. We can monitor their workflow or at least document their workflow, identify areas where agents could do the work, proof-of-concept them, go through the experimentation phase, and then move them into production. And our long-term goal is for those two use cases—external customers on top of our platform and internal customers on top of the platform—to be indistinguishable from a technology perspective.

(Joel Beasley at 00:25:32) Okay. This is great. So that's what I was thinking. I was like, why don't—the next logical progression is for this to spin out into some consultancy that helps people do this. But tell me what HealthEdge does. Are you already consultants or what?

(Rob Duffy at 00:25:47) So HealthEdge is a software-as-a-service company for health insurance payers. Now we say health insurance payers because it's not always a health insurance company. And what we do is we provide three core pieces of capability. So HealthRules Payer is our primary product, and that is a core administrative processing system. So when you go to a provider in the United States—and a provider can be a clinician or some physical therapy or a doctor or GP, whatever—that provider packages up all of the services and consumables that you received during that visit into an electronic medical record. And then that electronic medical record gets priced and coded and sent as a claim through an intermediary called a clearinghouse to the health insurance company.

(Rob Duffy at 00:26:39) And then the health insurance company has to adjudicate that claim, i.e., figure out what portion you need to pay, what portion is going to send to the provider for payment, and sends you an explanation of benefits. And the way that they do that is by codifying all of the rules and contractual obligations and pricing and rate cards that they have with all of the providers, codifying that into a rules engine. And that's what we run. HealthRules Payer is effectively a rules engine that takes all of the provider contracts, all of the plan documents, and then allows insurance companies, health insurance companies, or health insurance payers to adjudicate that claim as it comes in electronically and do that in an automated fashion very, very quickly, and enable those claims to obviously be paid quickly and then the insured person to get their explanation of benefits quickly. And we get up to a high watermark of 95% auto-adjudication of those claims. And then we have the system that the manual claims processors will go in and manually adjudicate ones that didn't get through the auto-adjudication process. We have another product, which is a pricing and editing product. And pricing and editing is sometimes the providers incorrectly price or incorrectly code things that are sent to the insurance company.

(Rob Duffy at 00:27:51) And we have a pricing and editing or payment integrity product called Source, which goes in and corrects the claims, edits—it's called editing, edits the claim. And then we have a care management product called GuidingCare and a member engagement product called Wellframe. So we basically help health insurance companies stop focusing on technology and payments and try and focus more on the consumer and try and deliver that sort of Amazon level of consumer experience to their insured population. And we do that by providing best-in-class, AI-enabled platforms that they can build their business on top of.

(Joel Beasley at 00:28:26) So do you think this will spin out, or do you think it'll just become a different part of your offering?

(Rob Duffy at 00:28:32) I think the biggest part of or the biggest cost in claims processing is manual adjudication. So if you imagine a claim comes in, it can't—the rules engine that we have and the HealthRules language and the HealthRules Payer platform can't adjudicate that claim for some reason. A human has to go in and open that claim, look at all the information, look at all the stuff that's been submitted, and go through a standard operating procedure that the health insurance companies write. And there's thousands of these SOPs and hundreds of reasons why claims pend. And they have to manually go in and look at this and open up this, copy this field from here, put it in here, or put it in this other system.

(Rob Duffy at 00:29:15) And it can be a 20 or 30-page document describing this step-wise SOP. And one of the biggest things that we can do with AI and agentic systems is remove the burden on the human for finding and collating that information from all these different systems and putting that information right in front of them and allow them to execute that SOP without having to go to 10 or 20 or 30 different systems. And that means they'll do it more accurately. They'll deliver faster claim processing, lower the cost structure of the insurance companies. And when we can provide faster claim processing and lower the cost structure of the insurance companies, then the health care and the servicing of health to people gets cheaper and better, too.

(Joel Beasley at 00:30:05) So you're talking about when you describe this whole framework for how you do these transformations within your own company, you're talking about taking that, packaging it up, and then using it specifically to help with that 5% at the insurance—the customers already have. So you're going to go to them and say, hey, we've got this whole life cycle that we can run on this part of your business and execute. Is that what you're talking about?

(Rob Duffy at 00:30:29) Yeah. Totally. 100%.

(Joel Beasley at 00:30:30) So cool, man. Because that's such a specific measurable—you already have the proof that you're able to do it within. You already understand the problem set. You already are vendors. You're in the companies. And I think it's pretty cool.

(Rob Duffy at 00:30:44) Yeah. And we've also learned how to do care and feeding of these systems, right? And do it safely, securely, in a compliant way, in a way that can make sure that we're not increasing the risk of leveraging AI. So the fact that we're sort of scratching our own itch first and dealing with our internal systems and building that platform—

(Rob Duffy at 00:31:08)
And then once we feel confident about doing it ourselves and we have a good process and we have a good workflow, turning that out onto our customers, I think, is the right approach.

(Joel Beasley at 00:31:20)
It's still so narrow because the whole world's going to need this. It's like a CMS back in the day. Everyone's rebuilding WordPress from scratch because no WordPress exists. Then a couple will come out and you just hope you're not Joomla, you know?

(Rob Duffy at 00:31:37)
Right, exactly. But that's a massively real fear for me. Right? As a CTO, we are pretty advanced in terms of trying to get AI and agentic systems up and running and deliver some productivity to our workforce and to our customers. And to do that, I need to build things. And every time there's an update from Claude or, you know, it's the AWS Web Summit in New York next week, I'm terrified that what's going to happen is a bunch of stuff that I've invested in has just been built by the AI platforms, which, you know, it's like every investment decision and every time we look at building a piece of technology, in the back of my mind, the evaluation criteria is, do I think this is on anyone, you know, like Claude's roadmap or AWS's roadmap already? And will it be released in the next three to six months? In which case, we should probably just wait. So it's a really dynamic and rapidly changing environment, which is providing some very interesting prioritization and discussions around what we should build versus just wait for.

(Joel Beasley at 00:32:39)
Yeah. I like the idea, though, because a similar concept came up to us a couple months ago where we were looking at the companies that are in health care and insurance. They have much more monitoring over their employees as far as their security systems and tracking movements and all of that. And I was like, what if we just feed all this data into a central location so that we can figure out what the repetitive tasks are and then stack rank them by time savings? And we were looking at that, but that was just kind of like, yes, someone out there is going to do it and they're going to do it really well and they're going to have way more knowledge than we have of this specific problem set. But I really like what you did where you didn't go build some technology that didn't exist yet. You just gave permission to people, let them figure stuff out, and then now they're bought and sold in, which is that behavior part, which is the hardest part.

(Rob Duffy at 00:33:30)
Yeah, exactly. And they're asking for new tools, and they're asking for different things. And that's what you want, this energy. You just need to put enough momentum into the flywheel of change to for it to start to become self-directed. And it does kind of grow up. It graduates out of the need for constant attention after a while. You know? And it starts to become you get your community leaders emerge, and you sort of elevate them, and then it gets to the point where it becomes a kind of self-sustaining initiative.

(Joel Beasley at 00:34:01)
So how does it work outside of your company? What are, like, just general employee count for technology? And you said you took 53 of these people. What's your total? What was that, like 10% of your—

(Rob Duffy at 00:34:12)
Yes, about that. We've got about 1,200 people all in, all said and done, in R&D.

(Joel Beasley at 00:34:18)
So that's a very small percentage then.

(Rob Duffy at 00:34:21)
Very small percentage. Now we have to scale it. And we have to break this, we have to take that pilot group and figure out, like, how are we going to get that level of passion and fire out of a much larger organization. And, you know, we think about maybe we just run a sequence of 50-people pilots, right? But we don't do it for four weeks. We do it for two weeks. And we do three in parallel, and, you know, we turn through this. And eventually, you'll reach a tipping point. You don't have to do it with 100% of people. Or do we just in tranches, you know, bring people on and get them through the workflow and the process and get that sort of build their fire and their passion for it and make them feel that, make them take ownership of stuff. It's a very interesting problem because going from 53 highly engaged people, you know, all desirous of a jar of spicy pickles, to a thousand people is difficult. Like, if you just open the floodgates, what you'll find is 50% of the people that come through those floodgates won't have any fire or any passion, and they'll just have been given another tool. Right? So you have to really think through, like, what is the scaling of a pilot where you've turned a bunch of people into converts, and they are now all in on AI in a very high-touch way? And how do you scale that across the organization?

(Joel Beasley at 00:35:51)
So let's say I worked at your company, say, as a software engineer. And I wanted my other buddy that works at, I don't know, Uber is using Cursor, and I just have my regular IDE. And I want to use Cursor, and I want to use an LLM. Can I just do that? Or is there, do you put out as, like, a CTO, like, a list of approved services that they could use so they could plug their code editors into these LLMs? Or, like, how does that work there?

(Rob Duffy at 00:36:18)
So we work in health care, which is obviously a very regulated industry. And it deals with pretty sensitive information. Right? We have a custodial duty for customers' customers' health information. And so in that sort of environment, we have to think very, very carefully about security, regulatory, privacy. We have to make sure that anything we're doing from an AI perspective is safe. We think about ethics. We think about biases and models. We think about how we're going to get our data back. Where is our data going to go? Is it going to get trained, or is it going to get used to train an LLM? Is there a potential leak? So we have a governing team, and anytime we want to use a tool, so if you're a developer on our team, you say, "Hey, I want to use Cursor," it would go through that governing team. We would look at the company. We'd look at what they're doing. We'd ask them for a whole bunch of security information and a whole bunch of legal information, and we go through a vetting process, a fairly deep and intensive vetting process for anytime we want to use a new tool or partner with a new company. And, you know, part of it sometimes is, do we like this company? Do we like the leadership? Do we think they're the good guys? Or do we think they're the bad guys? And, you know, in addition to all of the security and regulatory checks that we go through, we also kind of just like that. Is this a good partner for us? Did they have the same vibe? You know?

(Joel Beasley at 00:37:40)
Do they like spicy pickles?

(Rob Duffy at 00:37:42)
Do they like spicy pickles, or are they more of a garlic pickle kind of—

(Joel Beasley at 00:37:47)
Those garlic pickle guys with their beards, stay away.

(Rob Duffy at 00:37:51)
Exactly. So, you know, we do that, and that provides us with some control over what tools we deploy. But we have 16 tools internally that we've approved. So we try and strike this fine balance between making sure that we're doing things safely, securely, and in a regulatory-compliant way, and also enabling innovation within the company and then enabling a developer who wants to try Cursor out to do that. And in fact, we have another pilot going with Cursor. I think it's about 20 people on there. We have another one, Amazon Q. We have Claude Code. So we let, enable people as much as possible while still doing it in a compliant, secure, and safe way.

(Joel Beasley at 00:38:34)
You guys sound like a great company to work for. Are you hiring any technology people? There's lots of people listening.

(Rob Duffy at 00:38:39)
Yeah. But only spicy pickle people.

(Joel Beasley at 00:38:42)
Only spicy pickle people. Sorry, Joel. You failed the personality test. It's like, oh, it's all right. It's all right. I get it. I get it.

(Rob Duffy at 00:38:49)
Yeah. We're, you know, we're hiring. We have openings. It is a good company. It's a fun time to be in health care technology because health care technology is one of the last, if not the last, industry to really go through this digital transformation. And, you know, we sort of have this cellular network in a third world country moment right now. You know, and for those who don't get the analogy, you know, a third world country kind of skipped multiple steps in the telephony process. They didn't have telephones, but they didn't create hardwired telephone infrastructure. They just skipped over that part of telephony development and went straight to cellular networks. And I think we're in the same position in health care tech where we can kind of, there's a little bit of skipping of some digital transformation that can happen because we now have AI, agentic systems, and we can drive the adoption there without some of that middle area that we have to traverse through in terms of technology modernization, which is exciting. Right? We can move quickly, but, obviously, we have to do it in a safe and secure way. But it's a very exciting and dynamic industry to be in right now.

(Joel Beasley at 00:40:05)
Yeah. I think you're doing a great job in such a regulated industry, still keeping it at a situation where it's fun and people can advance and they can try things. And I talk to a lot of people. I would say you're on the higher end of actually taking this head on and not drowning from it.

(Rob Duffy at 00:40:22)
I think you have to. I think, you know, you have to go all in. And the risk of that is that, you know, we might be wrong and the AI bubble might burst. But the risk of not going all in is that your competitors do. And then it is real, and they get all the leverage out of it, and they get, you know, they streak ahead, and you'll just never be able to catch up. It's very, very hard if you have a piece of transformational technology to catch up with people who have a head start on you because their returns compound and they accelerate. And your returns, you're behind on that compounding and that acceleration.

(Joel Beasley at 00:41:00)
That's when you've got to bust out the checkbook and buy them, and that's expensive. Yeah.

(Rob Duffy at 00:41:06)
A lot of space vehicles.

(Joel Beasley at 00:41:07)
That's too many. So I read your bio, and the most interesting part that I found was your work period of eight months where it was, like, protecting a tiny human. That is the most brilliant thing I've ever seen.

(Rob Duffy at 00:41:26)
I get so many comments on that. That's why I've left it in. Every time I interview for a role or anything, people are always like, "Tell me about this period." Like, forget all of this other stuff. Tell me about this eight-month period.

(Joel Beasley at 00:41:38)
I don't care what you did at Amazon, Uber, or Salesforce. I want to know about this tiny human you were protecting. Were they royalty? Was it, like, who was it?

(Rob Duffy at 00:41:48)
It was my kid. It was a newborn kid. So I had this kid, and I had an opportunity to be a parent for a year. I was like, I'm just going to go and enjoy the kid and be a dad. And I put "protector of tiny human" because that's basically my job on my LinkedIn. And I was supposed to do it for a year. The idea was twelve months. And then eight months into it, I got a call from an ex-colleague and ended up at Salesforce helping that. But—

(Joel Beasley at 00:42:17)
You were ready?

(Rob Duffy at 00:42:18)
It was a fun, I was ready.

(Joel Beasley at 00:42:20)
I've got three. I can tell you. I know you're ready.

(Rob Duffy at 00:42:22)
Dude, when they're early, when it's, like, you know, months, like, years one to probably into the third year, there's, like, you are literally just keeping them alive. And there's, like, you know, they're cute and they smile, but they're, and they giggle and they cry. Like, you just, yeah. It's crazy. I don't know. I have friends who, we have a four- and a six-year-old. And I have friends who have a six-year-old, and they just had another kid. And I'm like, I feel like we're just getting over the hump a little bit at, like, four and six. And I'm like, I don't know why, they would go back to the beginning.

(Joel Beasley at 00:43:00)
Restarted the clock.

(Rob Duffy at 00:43:02)
Yeah. Exactly. You have three?

(Joel Beasley at 00:43:04)
Yeah. We have three. Seven, five, and two. One girl, she's the oldest, and then two boys.

(Rob Duffy at 00:43:09)
Nice. I have my oldest is a boy, and my youngest is a girl. And she's four. He's six. Yeah. It's great.

(Joel Beasley at 00:43:15)
It's beautiful because then you get to experience the true differences between boys and girls, which is beautiful. You know?

(Rob Duffy at 00:43:21)
And it's crazy how their personalities evolve and are different and, like, you know, you don't, it, one of my friends early on was like, "Oh, you can tell what the personality is going to be when they're, like, super young, like, one or two." And it's kind of true. Right? Like, these personality traits just carry them all the way through and develop and they grow and evolve. It's amazing. And she's, like, completely different. She's, like, feisty, like, strong-willed, super stubborn, beautiful, and, like, you know, fluffy bunnies, rainbows, unicorns. But then, if she wants to be, she's a total immovable object. And then he's, like, you know, just chill, more mellow. It's kind of fun. It's fun to see the differences.

(Joel Beasley at 00:44:00)
It is. It's 100%, it's amazing thing. And, honestly, I think it makes you a better leader. I think when you can watch the small humans develop, that gives you a better understanding of managing and working with people at work.

(Rob Duffy at 00:44:14)
Oh, yeah. Totally. And then the, you know, I always say that the collective intelligence, like, you know, every time you add 50 people to a group of people, the collective intelligence goes down by one year. So, you know, you have, like, a 2,000-person organization that can sometimes behave like a six-year-old too. It's a good, it really does teach you a lot. And, yeah, I agree. It evolves you as a leader.

(Joel Beasley at 00:44:43)
Oh, man. This is great. So, look, I love the protector of tiny human thing. I'd say that's my favorite thing about you. I will never forget. You can run into me in 10 years and be like, oh, protector of tiny human. That was beautiful. I think more people should do that. Do you ever see anyone else do it?

(Rob Duffy at 00:44:59)
Yeah. I mean, I see some people do it. I think the nature of work is changing a little bit and, you know, I think roles are changing as well. I see, I saw more people do it when we had a super buoyant tech hiring market. Right? Because I think a lot of people now are kind of afraid that if they stop, you know, their current job or they go and take a sabbatical, that, you know, they won't be able to find another job at the tail end. So I think there's a lot of fear in the industry right now about the job market for technologists. So, you know, pre-COVID, I think it was a little bit more of a hiring, an all-you-can-eat hiring buffet, you know, and people were doing more of that. But—

(Joel Beasley at 00:45:36)
You should be a little worried because things are advancing really fast. Like, I'm a little worried, you know, because it's really hard. The thing that makes me not worried, Rob, is how slow humans are to implement. I mean, we still have mainframes running out there. We have ancient technology that's still running out there.

(Joel Beasley at 00:45:54) There's a lot of work to be done, and so I'm out there looking at, okay, you know, for my day job, I get to sit here and look at the most advanced technologies in the world and talk to the most brilliant people. And I'm like, it feels like in my reality that it's here today, and it is. But there's this delay of actually understanding it, feeling comfortable with it, and then moving a ship that has 5,000, 10,000, 100,000 employees. I mean, that just takes a lot of time, and there will be a lot of opportunity in that.

(Joel Beasley at 00:46:24) So there's a lot of opportunity in that. If you get good at helping companies make the transition, you're set for life financially. Right? If you can get good at it and keep your eyes open and work hard, you can still make it happen. So I'm an optimistic person on it.

(Rob Duffy at 00:46:39) Yeah, I'm very optimistic. I think it's a great transformational period, and I think that people who really lean into it are gonna have a lot of fun, you know. I was playing around with some tools.

(Rob Duffy at 00:46:54) I was playing around with Amazon Q at the weekend, and, you know, I felt a feeling that I hadn't felt since, like, you know, Twitter was launching. And I remember when Twitter launched, every software engineer had all the tools at their disposal to launch something that looked like Twitter. You know? There was this amazing, incredible buzz in the technical industry and in the technology industry.

(Rob Duffy at 00:47:18) And people had crazy ideas. Everyone wanted to be an entrepreneur. Everyone's gonna be Mark Zuckerberg. Right? We had hackathons. Things were getting built in no time at all, and there was this creative energy in the industry about the ability to solve any problem that was thrown at you by a cool piece of technology that could take off and dominate in that space. And there's been a while since there was that level of energy in technology because, you know, because of that energy, thousands and thousands of companies were created to solve thousands and thousands of problems, and many of them are now market leaders. Right? No one's gonna start another Uber today. Right? It's just like you just wouldn't do that. But at that time, you could, and people did. And I think that that energy and that feeling is kind of returning to the technology industry when we see the power of AI tools and we see what you can do, and you see these agentic systems. And I see that spark of creativity in my engineering team's eyes, you know, when they realize what the potential is here, and then they have a thousand ideas about what they could do.

(Rob Duffy at 00:48:23) And, you know, that's invigorating. What a great time to be alive.

(Joel Beasley at 00:48:29) It is a great time to be alive. Hey, as we start to wrap up, I am watching the time. One thing I like to get when I have these conversations with brilliant people like yourself is I like to ask them one piece of leadership advice, and here's the constraints. So one piece of leadership advice that you've heard, you implemented, and it stuck with you through today.

(Rob Duffy at 00:48:52) I think the biggest piece of leadership advice I was ever given, I was running a program at Amazon, and it required many engineering teams to change something and do something in a different way. And I had a letter. We presented it to Jeff. We presented it to Jeff Bezos. We presented it to all the VPs or the S-team. And, you know, we got everyone's buy-in. And I had this letter, you know, in my hand that said everyone has to do this. And, you know, how I used that letter is I went and sort of beat teams over the head with it and said, like, you can't say no. Here's this letter. You're gonna do things a new way.

(Rob Duffy at 00:49:33) And as you can imagine, that was pretty disastrous. And my boss pulled me aside at one point, and he said, look, you can have a mandate. You can have a letter from Jeff Bezos. You can have whoever it is, is your CEO. But as soon as you pull it out and try and use it, it loses all of its power. And you have to go to teams, and you have to tell them why this is important and convince them and bring them along for the journey. And any of these large-scale changes, you have to go to the team, meet them where they are, understand what problems they're facing, how this stacks up in their priority, and do the hard work even though you might have a mandate. And I think that taught me about, you know, building movements and not using mandates and building communities and helping people along on the journey rather than just sort of yelling at them and telling them you have to do things a new way. And that stuck with me. It's advice I give to all my team all the time when we think about change and we think about getting humans to change.

(Rob Duffy at 00:50:34) Because, really, the thing that humans desire least in the world is change. And you have to help them through it and help lead them through a process rather than just trying to force them to do stuff. And that was a very empowering moment for me. I remember it very, very clearly.

(Joel Beasley at 00:50:50) Wow. That's some good insight. That's good advice. I like that story. It's a powerful story. Get the letter from Jeff Bezos, beat someone over the head with it, and then realize, oh, this doesn't work. They're humans. I gotta get them sold on the story and to follow me. Yeah. But that's an extreme example too because it's like, you know, you could use that at any other small startup. Be like, I got a letter from the CEO. It's like, great, it's Jeff Bezos, you know. Yeah. The one letter that would or should work in that situation, even if you're misusing it, still didn't work.

(Intro Narrator at 00:51:19) I love it.

(Rob Duffy at 00:51:19) Still didn't work.

(Joel Beasley at 00:51:20) That's beautiful. You're in, are you in New York? Is that where you are?

(Rob Duffy at 00:51:24) I'm in Boston today.

(Joel Beasley at 00:51:25) You're in Boston.

(Rob Duffy at 00:51:26) I'm based in New York.

(Joel Beasley at 00:51:27) Based in New York.

(Rob Duffy at 00:51:28) We have a couple of offices in Boston, so I'm up here.

(Joel Beasley at 00:51:31) Huge health community in Boston. Oh, that's one. Yeah. And Nashville too. That's where I'm at. Huge health in Nashville.

(Rob Duffy at 00:51:37) Yeah, I was just in Louisville, Kentucky.

(Joel Beasley at 00:51:40) Oh.

(Rob Duffy at 00:51:41) I went Vegas, Louisville, Boston, New York.

(Joel Beasley at 00:51:45) Experiencing all the cultures of the country.

(Rob Duffy at 00:51:50) Experiencing American consumerism at its finest.

(Joel Beasley at 00:51:54) Let's do it. Well, if you're in Nashville, hit me up, and I'll let you know if I'm ever around in New York. We've got family in Westchester, so we get up there about once a year.

(Rob Duffy at 00:52:01) Hit me up, man.

(Joel Beasley at 00:52:03) This is great. Well, Rob, we did it. We made a podcast. How do you feel?

(Rob Duffy at 00:52:06) I feel great. How do you feel? You feel like it's good content?

(Joel Beasley at 00:52:09) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.