S1E14: Third Party Risk with AI (ft. Corey Zeigler, Helio Health)
Healthy Uptime Podcast Corey Zeigler (Helio Health) _ Jordan Cooper (Rackspace)-20261030_150331-Meeting Recording
October 30, 2026, 7:03PM
24m 39s
Jordan Cooper started transcription
Jordan Cooper 0:06
Hello, I'm here with Corey Zeigler of Helio Health. He's A Chief Information Officer. Corey, thank you for joining us today.
Corey M. Zeigler 0:15
Afternoon, happy to be here.
Jordan Cooper 0:17
So as background, Helio Health is a regional behavioral health and substance use disorder provider headquartered in Syracuse, New York, with 509 inpatient beds. Today, Corey, we're going to be discussing about third-party risk with AI, and specifically how Helio developed a robust onboarding and governance process pertaining
To that, so please, Corey, just give us a high level, give us background on, you know, what Helio Health is doing with third-party risk regarding AI.
Corey M. Zeigler 0:50
Well, we really tried to rethink our whole third and actually 4th party risk when we started thinking about how to govern AI. And so from a strategy standpoint, we put those two together with the assumption that pretty much all the applications, services, relationships that might handle
protected health information are probably have AI embedded. And whether that's a closed system particular to that solution or a vendor, or whether that could potentially be an open system, you know, we want to govern that entire chain of custody for our patients and make sure that their information is not exposed. So
Jordan Cooper 1:19
Mhm.
Corey M. Zeigler 1:34
We re-engineered our entire third-party risk onboarding and governance process to include AI or make it.
part of everything that we do.
Jordan Cooper 1:48
Excellent. So AI is ubiquitous and you have to make the assumption that it's there and proactively implement governance procedures to ensure that you're preventing PHI from being exposed.
Can you say specifically some of your interventions or the particular changes to your governance structure that you implemented that you hadn't had before and what specifically you did and then kind of what the outcome has been since then?
Corey M. Zeigler 2:21
We really tried to get kind of a systematic process in place with branching in the very beginning. So I'll just take an example applications. So somebody wants an application to do single sign-on. So we use
Microsoft Entra for our identity and they want to sign use their sign Microsoft sign in to access that application. So that we fast track that, you know, that's we have a track for low risk. We have a risk rubric. So they put in a request, it comes into the risk rubric, we look at it.
Jordan Cooper 2:49
Mhm.
Corey M. Zeigler 3:00
We say, okay, that's a low risk, low impact. We get it in our catalog for applications and we approve it and the individuals and anybody else that wants to use Entre for that application to sign in using SSO can do so.
Jordan Cooper 3:04
Mhm.
Corey M. Zeigler 3:20
we have an application that say it's a new finance or ERP or something like that that comes in. That goes through the full process. So, you know, there's, we evaluate the application itself, you know, how it's stored, where the data resides, you know, we have regulatory requirements that as
required to stay in the United States for any data that is Medicare or Medicaid data.
you know, evaluate whether AI is by default enabled or whether we control whether or not AI is used. And of course, AI is a very broad definition. It could be anything from spell check to natural language processing to full agenic.
you know, act on its own, you know, type automation and then figure out what that, what they're doing with that AI model. Is it closed and contained within their system or whether or not, you know, they're contracting in a third or fourth party arrangement where they're using.
somebody else's tools to process that inside their solution. Once we kind of understand that, then it goes, if it's dependent on the risk level, we put it through various levels of governance, starting with our IS steering committee. You know, they have their authority for medium to.
high risk. If it's very high risk, it goes to the full executive team. And if it's extremely high risk, it goes all the way up to the board. And then we have some mandatory things like an open AI model that automatically escalate the risk to the board level. So anytime those things we've decided that we're absolutely not going to do,
from a governance standpoint, like expose our data to an open model.
that automatically kicks it up to the board level. So we have to basically make that value statement that says, this is so important to what we do, we're willing to take that risk.
Jordan Cooper 5:19
So...
When your vendors incorporate third-party language, large language models or kind of different kinds of AI, as you said, from spell check to NLP to full agentic automation, how do you assess where patient data goes and who ultimately has access either to that patient data or just even to the application itself?
Corey M. Zeigler 5:47
Well, with the...
The safest, most conservative path to that or strategy is just assume that any PHI is all PHI. And so we're covered under CFR 45, or 45 CFR, which is HIPAA, and we're also covered under something called
42 CFR Part 2, which is specific to substance use data. So there's some additional requirements there to protect that data.
It's all intermingled. So it's no longer that we can segregate data or tag the data. We treat it all as 42 CFR Part 2 data. And so we apply the same protections for everything. If it's has the patient identifiable,
information, even if it doesn't have diagnosis or anything like that, the fact that they're receiving treatment at one of our facilities basically makes it protected data. And so we basically treat it all that way. That actually simplifies our process because we don't have to worry about, you know, having five, 6, 7, 8 different decision trees.
Jordan Cooper 6:46
Mhm.
Corey M. Zeigler 7:02
to try and figure out how to treat that data. We have one way of treating it, regardless of what system it's in. And the only downside to that is potentially sometimes we have to say no to something due to how we treat that data as a collective.
Jordan Cooper 7:18
So as AI becomes more integrated into clinical workflows, how are you thinking about business continuity if one of your external AI provider experiences an outage or security incident? So again, this is talking about third party and 4th party risk with AI.
Corey M. Zeigler 7:35
Well, not only do you have to consider the risk of it going down, but you also have to consider the risk of just the dependency. So we're starting to think about tokenomics and the fact that, you know, I'm sure everybody's seen in the news that Anthropic is
you know, going through their IPO process, and they're reporting huge losses. So I think the move afoot in a lot of this is to get organizations kind of hooked on this, fully dependent on it, and then potentially that we're going to start to see the
The cost start to elevate, and so...
As an organization, you have to think about not only if it goes out, but what if you can't afford it anymore? You know, what if you have to make hard choices about the competitive market and how hard is it for you to move off of potentially one model or one vendor of the AI onto another?
And so is it hard-coded into the solution that you're buying? Is it able to be separated from that solution? Are your processes hard-coded to utilize that particular vendor? Or can you easily kind of pick your stuff up, your toys, and put them in another playground?
And so just thinking about that resiliency of your own organization, not only dependency for it going down, but also what if you have to make a change due to cost or, you know, other adverse actions by the vendor that, you know, you need to switch partners.
Jordan Cooper 9:18
So on the topic of kind of supply chain and switching different vendors, I'm wondering if you see an opportunity to consolidate AI infrastructure and governance across the enterprise rather than having individual departments contract independently with AI vendors. So is there
Only one procurement process, or is it possible that this department or that department can have small little pilots and you have multiple different competing kind of contracts going on with different, or is it all streamlined through your governance process?
Corey M. Zeigler 9:52
We do allow experiments, and so, you know, we are pilots, and because those sometimes uncover great value to the organization for the folks that really understand their needs and their workflows to be able to, you know, test something. So, we do allow those, those are...
highly controlled.
and you know have a definite start and end date with deliverables and all of that. But once that pilot has been deemed ready for production or ready for a decision, then it comes through a centralized governance process and we centralize the procurement of that, the configuration of that,
and we try and set it up so that we almost set up like a little bit of a store where we have these pre-configured solutions that departments or programs can tap into with their tool sets to get access to this technology. But yet we can control it centrally so we can manage our risk, manage our auditing.
make sure that we have all of those controls in place and not expect, you know, program leaders to figure this out on their own.
Jordan Cooper 11:07
So as AI, I want to pivot a little bit to the underlying infrastructure that supports these kinds of applications. As AI adoption accelerates across Helio Health, how are you thinking about the infrastructure needed to support those applications, particularly when balancing security, performance, cost, and control over that patient data?
Corey M. Zeigler 11:29
Well, at Helio Health, we are very much a cloud-first organization. We're very much distributed geographically. And just by the nature of the care that we provide, we're in the community. We have mobile methadone, these huge buses that travel to communities and provide care in the communities. We have folks that are
driving around in their vehicles, trying to find folks and make sure that they have access to care. So we're very much a highly mobile care delivery model. And so we don't have a castle with a moat around it where we can kind of protect that castle. We're protecting the farmlands that everybody
Jordan Cooper 12:08
Mhm.
Corey M. Zeigler 12:15
that's out in the wild. And so we have to basically protect that information, you know, from end to end, from the keyboard of the individual that's working with it, all the way to where the data is stored in transit and at rest.
So, given that requirement, the cloud just makes sense for us, so we're not hosting any of this onsite, and...
Jordan Cooper 12:37
Mhm.
Um...
So continuing on along these veins, how are you determining which, so when you talk about the cloud, there's public cloud and private cloud. How are you determining which AI workloads can safely operate in a public cloud environment versus those that should be private or controlled infrastructure? I think you mentioned or alluded to sovereign cloud earlier when you mentioned having to have the data centers in the United States.
So how do you determine public versus private when you talk about the cloud?
Corey M. Zeigler 13:38
We think of it more of open AI versus closed AI. So when we're utilizing these, you know, Anthropic has an enterprise that allows them to...
Copilot, Microsoft has a BAA with Healer Health. If we have a dedicated ambient system for the electronic health record, you know, we would get a BAA with them. And in that, you know, it has the stipulations and these requirements. So we very much look at these as
you know, software as a service versus, you know, a private cloud or, you know, hosting dedicated servers, you know, we're trying to the best of our ability, rent these as a service, and that way we can be much more flexible with the levers that we pull. You know, do we need more resources? Do we need...
you know, additional capabilities, you know, we can add those on and building blocks without having to worry about infrastructure.
Jordan Cooper 14:45
Let's see. So in terms of the economics that we were talking about earlier, how are you weighing running these AI workloads? I guess you said you're running them, you're the, you're running, you're working with vendors on a software as a service model.
I mean, obviously you have to have some infrastructure that you're managing or that you're outsourcing to somebody to manage beyond the application layer, the more the base like CPUs and GPUs. What are, how are you evaluating how much you're internally relying on hyperscalers or third party API providers? Where do you see opportunities to improve efficiency, I guess is what I'm trying to.
to ask.
Corey M. Zeigler 15:32
Yeah, that's where I was talking about. We may end up.
you know, potentially reversing that decision to basically crunch the data locally if it becomes the cost for renting that.
becomes cost prohibitive and we end up having to host that locally. You know, we're seeing some market forces out there where, you know, these data centers and trying to keep up with the demand, could it, you know, inflate the cost such that we have to make a choice to put in, you know,
infrastructure locally and use that to.
process the request. Right now, you know, the hyperscalers, the GPU, and all of that stuff is almost transparent to us because it's kind of behind that software as a service. So we pull the levers of what capabilities we need. We let the vendors manage, you know, the actual infrastructure that.
Jordan Cooper 16:27
Mhm.
Corey M. Zeigler 16:34
it takes to accomplish those tasks.
Jordan Cooper 16:36
So when we're talking about AI, throughout this conversation, I'm hearing 2 general themes. I'm hearing a lot about cost, price sensitivity, and security. And those are the two primary kind of considerations. And then obviously the utility of the application, what clinical outcomes or business processes are you supporting?
I'm wondering, on the security side, how you're preparing for outages or disruptions involving third-party AI providers, and if your current infrastructure provides sufficient flexibility and redundancy in case, you know, either your application becomes too expensive or if it's compromised by a cybersecurity incident.
Corey M. Zeigler 16:57
Mm.
Big.
The tough answer to that is we rely a lot on agreements, you know, with the third party vendors. You know, we have service level agreements, we have, you know, the requirements. So in the event something bad happens, evaluating the contract is kind of an after the fact.
right? You have to get it up and running first, take care of your patients first, you know, secure your systems first, and then all of the dollars that went into, you know, that recovery process. So we have to argue about who pays for that later, you know, whether it be a cyber insurance,
If we have, you know, vendor liability in that because of the, you know, lack of availability, things like that. So, you know, what we look at is, you know, we're sitting down with our vendor partners and talking about, you know, their RPOs, the recovery point objectives, and their RTOs, the recovery time objectives.
and making sure that those align with where we see our needs and our acceptable level of risk to operate. At the same time, you know, it's our responsibility to have good downtime procedures. You know, to have, like what we do is we have workstations that are identified
that have local copies of the forms that are needed to collect information on the folks we take care of. And also on there is the kind of an overwritten file that has, for the outpatient, has all the folks that are scheduled to come on site in the next two weeks or scheduled to receive care.
For the inpatients and the residentials programs, it's everybody that's on the census, you know, that's occupying a bed. And it has the key information there, the demographics, their allergies, medications, the things that you really need in case of an outage to take care of them. So they've got that to marry it and then marry it up with the forms to continue to document the care that they're receiving.
until such time as we can restore the system. So it's kind of a blended model of, you know, check but verify with the vendors and then internally taking responsibility for, you know, worst case scenario, you know, how are we going to continue to provide care.
Jordan Cooper 19:44
Given that Helio Health does have that 42 CFR substance use, you know, very sensitive data, are you seeing an increase in the cadence of AI generated attacks?
Corey M. Zeigler 20:01
It's almost scary to say not yet, because you don't want to jinx yourself, you know, knock on wood. We are having a lot of conversations about, you know, one of my engineers said the other day, you know, as in regards to security is.
We can't operate at human speed. You know, when the vulnerabilities would be announced and potentially they had been found in a lab or there was a potential, you know, for this or it was released in the wild, you know, we had some time to shore up our systems to protect ourselves against that.
And then, you know, zero-day vulnerabilities that came in, and that compressed that time. Now, you know, with AI, we're really having to react to things at machine speed, not human speed. And so we almost have to have our AI work.
Jordan Cooper 20:50
Mhm.
Corey M. Zeigler 21:02
against their AI to try and, you know, be reactive, you know, proactive where we can, but also reactive at machine speed so that we can close vulnerabilities or close gaps, or at least be made aware of those gaps as fast as the bad actors AI's are.
probing and trying to compromise our systems.
Jordan Cooper 21:28
So as we approach the end of this podcast episode, Corey, a final question. Many of the listeners of this podcast are at large academic medical centers or integrated delivery networks with thousands and thousands of beds and many hospitals. I'm wondering,
Do you have any messages for them? How should they work to enhance, what should they demand from their AI vendors that would benefit the whole community? How should they better coordinate care with their community partner organizations like Helio Health? What message and what unique insight do you think
coming from a regional behavioral health and substance use disorder provider, could you have two larger health care delivery systems who are dealing with the risk and costs of dealing with AI in this era?
Corey M. Zeigler 22:26
Well, I get to talk to a lot of them. I sit on the policy steering committee with Chime. We meet every week and we get to talk through a lot of these challenges. And I get to hear from their perspective, you know, the challenges they're dealing with. You know, it's they have a much bigger
team and much more resources, but they also have much more exposure and a much broader area of responsibility. I would say that if I had one thing to ask would be to set the precedence. You know, they have the buying power and the influence in these large systems to
Jordan Cooper 22:46
Mhm.
Corey M. Zeigler 23:09
ensure that the vendors are doing the best thing for the industry. So to the extent that they can hold them accountable and make sure that we set, you know, a standard of security, a standard of care, then the little folks like me, you know, when we enter into these agreements, it becomes the defect.
standard that these controls are in place. So it's the big guy looking out for the little guy would be great. From A behavioral health standpoint, I think that we've, you know, as a nation, as a culture, have made great progress towards reducing the stigma of care.
that, you know, an addiction diagnosis can be treated just like an appendicitis. And, you know, it's healthcare. And our transitions of care between the different health systems and working with your local behavioral health and community-based organizations, we want to
see that care extended outside of the hospitals, outside of the IDNs, and into the community, and have it seem to the individual that's receiving that care that we're all one health system with their best interests in mind.
Jordan Cooper 24:24
Well, I appreciate this conversation that you've shared with us today. For our listeners, this has been Corey Zeigler, the CIO of HelioHealth. Corey, thank you so much for joining us.
Corey M. Zeigler 24:37
Thank you.
Jordan Cooper stopped transcription
