S1E12: Establishing Human Accountability for AI Agents (ft. Adrienne Moore & John Almasan, Banner Health)
Healthy Uptime Podcast Adrienne Moore _ John Almasan (Banner Health) _ Jordan Cooper (Rackspace)-20261008_140112-Meeting Recording
October 8, 2026, 6:01PM
28m 42s
Jordan Cooper started transcription
Jordan Cooper 0:03
with John Almeson, the SVP and Chief AI Data and Infrastructure Officer at Banner Health, and Adrienne Moore, the SVP of Strategy and Growth at Banner Health. Adrienne, John, thank you for joining us today.
Almasan, John 0:17
Thank you for having us. Very excited about the discussion.
Moore, Adrienne C 0:19
My pleasure.
Jordan Cooper 0:21
So for our listeners, Banner Health, in case you don't know, and I'm sure you do, is a health system headquartered in Phoenix, Arizona, with 33 hospitals operating in six states, Arizona, California, Colorado, Nebraska, Wyoming, and Nevada. Now Banner Health is investing more than a billion dollars in AI robotics, telemedicine, and cloud-based technology to transform how care is delivered.
So today we'll be discussing how Banner Health imbues trust in BannerWise, Banner Health AI umbrella, through appropriate governance and oversight structures. So John, I'd like to start with you. Could you elaborate upon your sentiment that AI in healthcare must be governed before it is scaled and that organizations building trust and human accountability in their technology
are the ones that genuinely improve lives.
Almasan, John 1:07
Thank you, Jordan. I'm glad you're starting with this question. I define AI in a very simplistic way as very intelligent and bright teenage kids coming to work. They can really process thousands of documents in seconds, analyze them in ways that people can because of the volume of the speed.
They can actually create outcomes. They can be very powerful in the hands of the humans. And they can also make decisions and act. We have to think about these agents. They are built to be autonomous. They are built to have agency. And therefore, like any other kid, while they do fantastic job, they impress us every single aspect of what they do.
they also can make childish mistakes. So we have to have the oversight in place. Now, that oversight is critical, especially when you're dealing with healthcare, when you're dealing with regulations. And therefore, I strongly believe that in healthcare, but also beyond healthcare,
We have to build the governance first. We have to have the process, the protocols, and the people fully embedded into the process to have what I call the human-centered AI adoption. I also wanted to call out at Banner, we take this very seriously. And while AI and these agents, as I mentioned earlier, they
tend to be autonomous and to make decisions. We deliberately engineered human in the loop in ways that we have full control in every aspect of the agent. When it comes to designing these agents, thinking of the agents, how they actually engage within each other, how they think, reason, reflect.
Everything, what I call the accountability part, human decides and human signs off, and more importantly, human is the one to make the final decision on the outcome. Agents can make recommendations, and they should, because that's why I want to use them to be pretty much the
multipliers and amplifiers of our work to do more with the same people, but not the decision maker. Humans are the only ones authorized to make decisions.
Jordan Cooper 3:22
So I love how you use this metaphor of AI agents as bright teenage kids who are excel in certain capacities, but also are prone to making childish mistakes. Now, the health system and the humans in the health system in this metaphor are like parents who must set boundaries. How can healthcare systems like
banner health, set these boundaries in order to build trust in AI.
Almasan, John 3:47
Yes, this is something that, you know, I spent the last 25 years in the industry of AI, from qualitative analytics to quantitative analytics to data science, cloud AI, and now, of course, generative and agentic AI. And I can tell you that none of the companies in any industry really figured out how to do
this very well. So in order to ensure the full control of the human, as I mentioned earlier, it literally takes deliberate engineering. And you have to start with the infrastructure. We have infrastructure today. Most companies, I would say almost everyone, is building infrastructure or used to build infrastructure for people.
how people actually drive the use of technologies in a way that's meaningful and useful for the people. Now we are talking about a technology, specifically agentic and agents technologies, AI technologies that are different than traditional technologies. So we really have to redesign the technologies to be successful on the long term.
Adopting AI quickly can actually trigger maybe immediate recognition or some benefits, but certainly in the long term can trigger risk and problems for every company. That's why at Banner we take it very seriously. We focus on the big picture and in the long run. We want to build a foundation first, and it starts with
everything on the stack. We want to, and we have infrastructure that we fully control. Everything that it comes to the data, our patient data, our employees data, every data, it is properly protected and all the safety nets are in place. So we control the data, the pipelines, the perimeters,
inside of our private cloud, as well as on the private cloud providers that we are using. Second, the data itself, the platforms for the data, we recognize having the data in multiple locations is not efficient, not just to manage the data itself, but also it's prone to errors, it's prone to cost increases,
And that's why we are doing a tremendous effort to consolidate all the data into one single place. And last but not least, we want to bring the data, AI, and BI together. Again, that creates a specific focus on accountability, identity, and access control management, authorization, authentication,
all in one single place. And last but not least, while we have the stack build to adopt agents at every single level, at infra, at data, and BI, we can also build cybersecurity by design, bottom up. If you build it at bottom, you can inherit it to every single layer in the stack, which makes it significantly safer
by design rather than silo AI adoption, which usually creates less safety nets and difficult opportunities for managing properly and controlling these ecosystems of agentic and AI systems.
Jordan Cooper 6:47
I appreciate that explanation, John. Adrienne, now you're the SVP of Strategy and Growth. So I'm wondering if you could help our listeners ground this AI infrastructure conversation that John's just elaborated upon within Banner's overarching clinical financial and operational strategy, possibly offering insight into how you've been calculating ROI,
Obviously, a lot of what John is talking about is very resource intensive, right? He's talking about doing things right from the very beginning, building things from the ground up with cybersecurity. And doing things right sounds like the best interest in the organization in the long term, but requires somebody to put a lot of upfront investment of resources in order to enable
this long-term solution architecting. So can you talk about how you're calculating how to allocate an AI budget, prioritize AI requests, consider AI tokenization, and how everything John is talking about fits into Banner Health's overarching strategy?
Moore, Adrienne C 7:43
There's a lot to unpack there, but yes, first I want to acknowledge John's comments about governance and exercising caution are foundational for Banner. I think AI strategy is also based on a moment in time. It's constantly evolving. Every health system has an AI strategy, but
The moment in time that we're facing is this time where healthcare is facing unprecedented headwinds. We've got regulatory challenges ahead of us, coverage losses for patients and members, payment cuts. And at the same time, we have these exciting AI tools that are poised to accelerate the way that we change.
Maybe we can bridge some of that gap. Can we increase access to care, make life better for patients or team members? So when we talk about prioritizing, and I'll get to ROI after that, when we talk about prioritizing, we need to start with the problem you're trying to solve and understanding the value of that problem for your organization.
What return do you get out of it? Is it more work-life balance for your team? Is it more capacity to take care of more patients or members? Is it a cost savings, improvement in quality? Once you understand the type of return you're getting, you can try to calculate an ROI. But you also need to understand if it's a problem you can solve with AI or...
If you can't solve it completely, can you see an improvement with AI? And I think we think about prioritization the same way that we always have. AI isn't new, maybe agentic AI is new, but we've had machine learning and automation, lighter forms of AI for a long time. And you prioritize that.
by identifying repeatable tasks, high volume binary decision-making, things where a set of rules can guide a decision, so you don't necessarily need a human. Those are the same rules a human would use. So we're prioritizing this agentic AI the same way. The thing is, the capabilities are higher.
But you still need the same basics for its payoff. We need enough scale to make the investment in developing or buying an AI agent. It's not worth the cost if you don't use it. So you can't automate something that, you know, three people use twice a year. That isn't worth the investment. Outside of.
prioritization, you need to consider the process too. It is possible to automate a bad process and do the wrong thing really fast. So it's worth having process engineers evaluate a process before you're training an AI agent to replicate it. And you can identify and eliminate failure points and redesign them prior to automation.
There's not a hard and fast way to create an AI budget. Right now, I think everyone's AI budget includes foundational building blocks that organizations need to invest in. But we all have this expectation that AI budgets will decrease over a longer time horizon after you make these core investments.
Let's say your AI budget needs to reflect your priorities, and it is an area where costs can escalate. So you need to keep an eye on that spend and really understand your return, knowing that sometimes your return, it may be an improvement in work-life balance. You know, an example like ambient listening, improving documentation time so that providers
are doing less pajama time, as they call it, with their documentation. Some of those returns are not as easy to put on paper, but really understanding the benefit that you're getting, whether it translates directly to dollars or not, is important. And I will come back to what John has been sharing. AI governance is critical, and you have to have that in place.
A step one before you can develop an AI strategy.
Jordan Cooper 11:33
So we've seen organization, thank you, Adrienne, for elaborating on kind of the business approach to what tech is doing at Banner Health. We've seen organizations across the country, notably potentially the example of Uber, but there have been many other organizations that have blown through their full year AI spend of tokens
of AI tokens in Q1, but by the end of the year, you know, they don't have anything left. So I'd like to, since you just mentioned about kind of the budget decreasing and the high initial spend, and potentially kind of, you know, I guess, how do you think about who gets access to what?
what use cases get prioritized and authorized, and how do you kind of stabilize usage so that it doesn't blow through Banner Health's budget?
Moore, Adrienne C 12:25
Not sure which of us is taking that. Maybe I'll start and John, you can fill in. The governance process, I think, is key to that. You have to know where you're spending, what you're spending on, who is utilizing the tokens, and what kind of return you're getting out of it. If they're automating a low value task.
Almasan, John 12:30
Lazier.
Moore, Adrienne C 12:45
It may be something that you have a conversation about, John.
Almasan, John 12:50
Yeah, so I wanted to continue on Adrienne's point. I think she made some very important statements when it comes to both governance. And when I say governance, we are not talking only specifically to privacy, hallucination, quality, but also governance on the cost. And how do we maintain that focus, sharp focus on the business value?
And I would say to me, as Adrienne mentioned, it's all driven by the business. It starts with the business, as the business is going to bring either a potential opportunity for improving the patient care or improving the, I would say, support for clinicians to take the burnout from
those very difficult and time-consuming type of tasks that everybody has to perform. But at the end of the day, AI, it is a technology that can be used in many ways. And at Banner, we are very mindful that this is a technology that can make a huge impact, a positive impact, but also can have
risks. So we deliberately manage the use of AI in a very narrow and very controlled way. That's why we put BannerWise, you mentioned Jordan in the beginning, BannerWise, our front door for everything AI at Banner. So we use BannerWise to channel all of the AI
through that specific front end as well as a back end. The front end is literally the user experience. Imagine a very well designed.
UI, user interface, where our clinicians and our financial people and HR people can go in one single pane of glass to access the data they need and the knowledge they need in one simple place. Means the knowledge, the agents, and the work is done in one single area. But now everything has to be done
with that cost in mind. So we are now narrowing the scope and allowing those people, for example, our clinicians at the Banner MD Anderson Oncology Center, instead of going through during the night with the scribes and nurses and doctors to spend hours through
hundreds of pages of patient medical records. Now we can use our agent to really summarize those for the doctors and the nurses so the scribes can go validate, ensure full quality of the correctness, completeness, and ensuring that everything is accurate.
And then the doctors, in a matter of an hour maybe or less, they can create the best personalized care for our patients. So they can have quick access to meaningful information rather than spending, it used to be maybe 6, 8 hours of work, maybe even more sometimes, depending on the volume of pages.
which could be very high. 300 pages could be a normal number of pages for a medical record for a patient. So again, our scope is to be very narrow in scope, very targeted, to bring the value only where it matters most, and allow AI to be used in those areas that really have a strong and very well-defined ROI.
Jordan Cooper 16:07
So, John, I understand that BannerWise is your front door for everything AI at Banner. And so there's some in-house models of agentic models that you've been developing, but I think you also leverage AWS and Azure. So my question is, does moving from frontier models to in-house models actually lower costs once you account for GPUs, ops, and people?
Almasan, John 16:31
So that's a great question. The truth is that we want to make the best use of everything that exists on the market. And I can say that BenderWise has that capability to enable, if needed, frontier, maybe the latest frontier models from any of these labs that are available today.
But also, when it comes to the quality, as you mentioned earlier, now with the open weights model, or with the ability to build in-house, you can actually provide significantly higher security, privacy, and cost if we build in-house. So we have a blend of both, and we are making those decisions on how to use AI based on
what Adrienne mentioned earlier, what are the outputs the business want? And if you can do it with a lower cost, if you build in a house, then we'll certainly adopt that way. If we are seeing an opportunity to use something from a frontier model that is not available or it's not easy to accomplish, they will certainly use something from our vendor partners, like you mentioned Amazon or
Oracle and others, yes.
Jordan Cooper 17:34
So you mentioned, and especially we spoke earlier in this conversation about the concept of trust, and now we're talking about leveraging hyperscalers and pioneer models. I'm wondering what makes, and you mentioned that you host some of your environments of banner-wise in a private cloud environment. So what makes a cloud trusted for PHI, where you've spoken about security,
Almasan, John 17:40
Mhm.
Jordan Cooper 17:55
You spoke in. So what makes it a hardening an environment and making it compliant with all the regulations? What makes a cloud trusted for PHI and who carries accountability when it's a vendor's environment? And you know, some organizations, they're going to do private cloud, sometimes hyperscalers, sometimes they'll do a sovereign cloud.
You said that banner doesn't, is it cost, flexibility, or is it that hyperscalers controls aren't enough? There's a lot that I'm giving you here, but basically I'm asking what makes a cloud a trusted environment and how do you evaluate that accountability when you work with partners?
Moore, Adrienne C 18:31
Before you answer, John.
Almasan, John 18:31
Thank you. Yeah, please. Yeah, go ahead.
Moore, Adrienne C 18:33
I would say that even when we use a frontier technology, you ensure it's a trusted private environment.
Almasan, John 18:40
Thank you for saying that, Adrienne. Actually, that's the first thing, to Adrian's point. The first thing we do is to build a trustworthy relationship with all of our partners. They must be complying with all the regulations in the healthcare industry and certainly be HIPAA compliant. That's first and foremost. We don't deal with anyone by design.
That's part of the approval process, which includes legal, privacy, risk, compliance, cyber, AI data, you know, leaders that are obviously in charge of making sure that whatever cloud providers, hyperscalers, or platforms we are using, they all meet or exceed.
those requirements. Second, we also recognize that the progress, it is a hybrid. We cannot say we build everything in-house and then close the shop. That's not possible. We rely on the trustworthy vendors because they help accelerate the efficiencies and the outcomes that we want to do. So in that space, the way I think we are doing it is first and foremost, working
with our partners. We build trust by creating transparencies, by adding controls that, like I mentioned in the beginning, the ability to control our own infrastructure when it comes to the cloud, our private cloud, but also to control and be in partnership control with our trustworthy vendors.
I think this is the approach that we are putting in place. Second, we also do full isolation in the beginning. Our data is fully private, is fully secured, and we make sure that we have the highest standards for protecting our patient and our employee, virtually every single piece of data in our environment. Second,
Where we are engaging with these frontier models, we are doing it only into the private and to those channels that provide us the highest standards that are meeting or exceeding the industry regulations. And 3rd, when it comes to the controls, we are actually working to expand to our proprietary channels, the governance controls. We have them now building, we are
building and work in progress in building multi-hyperscaler and multi-platform governance that really is going to provide us what I call the window of transparency on how everything happens when it comes to data, when it comes to AI, when it comes to the whole process. So again,
We have multiple avenues to ensure from governance, from policy implementation by design, but also through innovation that we are putting in place to ensure that we have everything either meets or exceeds our controls expectations.
Jordan Cooper 21:19
I'd like to pose a question to Adrienne that follows up on what John just said and then maybe John, you'd like to add to it. Since partnership is so intrinsic to Banner Health strategy and Banner Health doesn't take the go it alone approach model, I'd like to ask Adrienne if you can help
listeners understand how Banner Health goes about drawing a line between what Banner chooses to run itself and what it chooses to hand over to a partner.
Moore, Adrienne C 21:48
So we've had a hybrid strategy in particular with AI and technology for as long as I've been at Banner. We have partners like Anthropic, like Microsoft, like Oracle, that have core competency that exceeds our own in specific work streams.
And then where Banner can develop something unique, where we have internal capabilities or where a product doesn't exist with a partner, we tend to focus on those areas for internal development. There are times where we may begin something and we have
the leading expertise in that, and then we seek a partner and we push it to the next level. So it really depends on who has the core competency or the subject matter expertise and how far we can take a technology or a workflow on our own. And if there's acceleration, we can get with a partner. We're open to
whatever strategy makes the most sense at that point in time to drive value.
Jordan Cooper 22:56
Great, so I'd like to make this concrete. You know, we have a few more minutes before we wrap up this episode. So I'd like to just take a concrete use case here and let's talk about, for example, your Genie Databricks application. I believe it puts natural language analytics in the hands of people who aren't data engineers, and correct me if I'm wrong.
So I'd like to ask who is using it today? Clinicians, finance, operations, all of the above. How do you decide what data they can or cannot ask about? And more generally, I'm sure listeners would love to hear about Banner's development journey with Genie. And again, this pertains to governance, access control, and data readiness.
Almasan, John 23:33
So I'll kick it off. The fact is that Genie from Databricks is just one of the capabilities that we are making use of. And that's because Databricks provides us a trustworthy and an environment where we have, again, full control when it comes to the data itself, the data flows.
and more importantly, the outcomes and how the data is actually made through the process of AI. I would say that BannerWise, the back end of BannerWise is our front door for all of the capabilities that we have, which includes for AI, which includes
data bricks. But we are, like I mentioned earlier, we are not limited or specific only to one. Now, we are creating synergies with all these vendors whenever is necessary or whenever we see a benefit, because we can see that, I'll give you the example earlier for the oncology, the Banner MD Anderson Cancer Center, where we are having an agent that really creates a major impact on
quality, time, speed. Also, we have similar prototypes and implementation production actually to support our patients. We already have working with another partner to create coverage for those times of the day when we don't have a person on the call for an appointment. So then an agent, an AI agent can actually go
and potentially help a patient to find the right doctor and then book an appointment. So again, this is AI technology that we are leveraging to really make our patients' access to healthcare easier and faster. And when it comes to internal agents, we have, like I said, narrow targeted but very aggressive to help
our clinicians in building visibility. And we are now expanding from oncology to cardiology and infectious diseases and everywhere else to really speed up the process. They will take away, you know, that time-consuming task that the doctors, the nurses, and
our scribes are doing and then letting them spend more time and giving better care to our patients.
Jordan Cooper 25:54
So, Adrienne, we have time for one more response for each of you before this podcast concludes. I'd like to present one of two questions to each of you, and you can choose which one to take. Obviously, these tools have been completely game changing at Banner Health over the last few years.
What does your data and AI team look like now compared to two years ago? Or question two, what kind of advice would you give to yourself two years ago based on the lessons you've learned over the last few years?
Adrienne, would you start?
Almasan, John 26:30
And...
Moore, Adrienne C 26:31
Sure. I don't know if you noticed my eyes lit up when you said Databricks and Genie. The advice I would give to myself two years ago is that the best users of Databricks and Genie are going to be the people who already understand the business and know how to write code. So think of your super users on your enterprise data warehouse.
your decision support, your process engineers. Databricks and Genie works best when you know what you're asking for, you know where the data lives, you know how to validate it, but it gives you such speed and iterative.
iterative processes that can take your analytics to the next level and you really optimize that when you invest in the people who know your data and your business the best and giving them the skills to work in that environment.
Jordan Cooper 27:24
Thank you. And John, would you like to take one of those questions?
Almasan, John 27:28
I will, I would take the one with, I will think, I will think about AI in a way that actually brings excitement. I know there are two positions here, some are skeptical, some are very excited. I will look at it somewhere in the middle and I would say that AI is here to stay and here AI is going to be
really the empowerment tool for everybody. What I wanted to highlight for everyone is that we better start using AI because it's going to make us faster and more efficient, but we also have to do it in a wise and thoughtful way because not something inexpensive and also it's something really risky. Now at Banner, we certainly build all the guardrails by design so we can protect
either one of the risks when it comes to hallucination or when it comes to the cost. But I encourage everybody to be curious, to start using it, and to make the best out of this technology as is really going to make a change in our lives.
Jordan Cooper 28:24
So for our listeners, this has been John Almussen, the SVP and Chief AI Data and Infrastructure Officer at Banner Health, and Adrienne Moore, the SVP of Strategy and Growth at Adrienne Health. Adrienne, John, thank you both for joining us today.
Almasan, John 28:40
Thank you.
Moore, Adrienne C 28:41
Thank you, Jordan.
Jordan Cooper stopped transcription
