Q&A: Inova Forges Forward on Its AI Journey With Mindful Moves

HEALTHTECH: How does Inova determine which AI use cases will deliver the most value for the organization? How are you choosing solutions?

MCMANUS: We’ve stood up an AI ethics and oversight committee that has a robust structure. We have a mature AI product management function; a team of five AI product managers who help us with idea management, the front door to governance, business proposals — essentially, separating the noise of the AI market and vendor space from the real concerns we need to make decisions on. Through that process, we look at the risk and liability of a particular feature, any policy changes that may need to happen, adequate training and visibility, cost, and value-creation components. Are we taking it, shaping it or making it? Those are three very different disciplines across the AI portfolio. What is the expected return? Are those soft or hard savings? And there are many things that we do in healthcare that affect quality that don't really generate a financial return, and that’s OK. It’s important to treat the whole scope as a balanced portfolio.

We look at those components and then make a simple decision: Is this OK or not to do? We need to determine whether it is ethical, safe, reliable, and all the mechanics with it. 

Today at Inova, we have 67 different AI features in production. Of those features, we’re tracking value, safety and reliability and doing continuous monitoring.

HEALTHTECH: Tell us more about key stakeholder involvement and the multidisciplinary nature of adopting an AI solution. How do you align users’ interests with organizational aims?

MCMANUS: Even as Inova’s chief data and AI officer, I don’t believe in centralizing the implementation of AI. What we’ve centralized is answering a simple question from an interdisciplinary perspective: Is this OK or not to do on behalf of Inova? But we encourage our organization to really explore opportunities around AI and many other technologies.

Our AI ethics and oversight committee is chaired by one of our physicians. We have a subcommittee structure where we focus on certain cohorts of the business. That provides the space to have dedicated conversations about what our physicians and advanced practice providers care about relating to AI. What are their training needs? What are their concerns? What technology areas do they really want to focus on? That can be very different from the business support services cohort. It is a priority for Inova to see and hear the different needs our team members and caregivers have regarding AI.

I think what a lot of people miss out on is a focus for AI training with leaders, because that’s where you need to teach how you can orchestrate with AI. How do you identify and evaluate a problem that AI is appropriate to solve? Can that be done responsibly and safely? There are different learning objectives associated with that.

Our AI ethics and oversight committee has 24 members, and they’re from all different swaths of the organization. We’ve got seven different physicians on there, covering different scopes of practice; we have key administrators, leaders from nursing, supply chain, legal, privacy. We also have a component for bioethics, as well as the voice of our community and our patients. We bring all of those things together. I know that can sound complicated, but we’ve streamlined it so it’s an easy-to-understand process.

HEALTHTECH: How did Inova prepare its foundation to deploy these AI solutions?

MCMANUS: At Inova, we have evaluated over 400 different AI features to date. It’s not uncommon for health systems to have over 1,000 vendors in their portfolio. When I look at all of that activity, I return to these questions: What do we take? What do we shape? What do we make? 

When I say we take something, it’s usually an approved medical device or clinical decision support algorithm approved by the U.S Food and Drug Administration. We don’t touch those things, from a technology and configuration standpoint. We implement a vendor product, and we’re doing a level of vendor management and clinical use oversight to ensure its use is appropriate.

What we shape is actually most of what we do. We buy something from a vendor partner, and we’re given some level of low-code UI to configure and shape the product to fit Inova to whatever degree that vendor allows us to. Almost all Epic AI features are in the shape category. That’s not really full-scale data and AI engineering — it’s a low-code way of shaping a product to fit whatever workflow it’s serving.

The final category is make, which is really where the modern data stack — your infrastructure, your tech stack — comes in. This is where we actually make solutions that are truly unique and custom for Inova. Altogether, 10% of our inventory we currently take, about 80% we currently shape and about 10% we currently make. When we walk about the discipline of AI product management, that is an example where we have chosen to invest in the people, process, governance, roles and skills that 90% of our AI inventory requires, which is taking and shaping.

I wanted to highlight this because a lot of people think they just need to hire a bunch of AI prompt and industrial engineers and be done with it. But actually, it is important to invest in the skills needed for the type of AI that you are engaging with.  

On the making front, I think, first and foremost, you can’t do AI if you can't connect your data to these models. So, we invested in a modern data stack. We have our stack run on Microsoft Azure. We have positioned our data primarily in Databricks. We use FiveTran ingestion, dbt Labs and Databricks for transformation and Astronomer for orchestration of those data procedures. We have Microsoft Fabric and Power BI for delivery. We have DataHub for metadata management, DataRobot for data science, GitHub for code versioning. It’s really an entire CI/CD pipeline for analytic development and deployment.

Click the banner below to sign up for HealthTech’s weekly newsletter.

Because we’ve built that in a modern way, we are able to orchestrate, fuse and layer AI all over it. So, we are seeing dramatic capabilities return. We have a series of AI engineering platforms like Signal 1 that allow us to orchestrate and govern agentic capabilities. We’ve built products to help us distribute these capabilities out. We’re doing a lot of work right now on an enterprise knowledge management platform, because we don't want to redo the concept of doing retrieval augmented generation for every single AI feature; we would like to orchestrate our institutional knowledge once in a unified appliance, almost like a data warehouse for knowledge, so we can support continuous API utilization every time an AI feature needs a knowledge-based component to help it perform its duties and services well.

You need a very robust monitoring strategy: not only how do you monitor the quantitative and qualitative elements of accuracy, performance, drift and safety, but you also must have institutional or industrial level monitoring for logs in the network. We have partnered with companies like Signal 1, Rubrik and CrowdStrike to truly monitor this ecosystem in depth. How are you on the lookout for agent use that’s outside the scope of your approved protocol? There's a level of work that has to happen there. 

HEALTHTECH: How is Inova upskilling its workforce? How are you getting users excited to use these AI solutions?

MCMANUS: You have to make AI accessible and consumer-friendly at scale. I mentioned earlier, 90% of our AI portfolio, we’re taking or shaping, meaning we’re buying from a vendor. That creates a different problem. Yes, there's a lot of embedded AI in a lot of the products we’re using, but it’s in a lot of different places, and it looks different. That creates a literacy problem. How can you be literate when there’s 100 ways to do things?

What we’ve focused on is purposefully creating Inova proprietary and branded distribution portals so we can slow down the manifestation of change, even though we’re not slowing down the technology capability. We want folks to be able to have a chance to build literacy, which means things have to stay static a bit. The place I go to find an AI agent to help me do something needs to start feeling the same, so we get familiar with using it and trusting it, even if we’re changing out the backend frequently. We’ve invested in a consumer-friendly approach to distribute AI capabilities to build a layer of separation from the velocity of change.

We also have a detailed communication plan infused with an upskilling plan. That involves how we are communicating to senior and midlevel leadership, units and departments, and team members as a whole. It can be top-down through email, or bottom-up through AI town halls, where people can form communities to learn from others firsthand. We also have focused on starting to develop an AI champions network.

And how do people know where to go for support? That highlights structural things that organizations can miss. For example, we updated our patient safety reporting system to have a designation for AI, so our team members can anonymously share a patient safety event that they want to report, and they can list AI as a partial or primary contributor to that safety event. That way, we can facilitate root cause analyses to support effective responses, like we do for any type of patient safety-related space.

We also have formal training for AI continuously added to our learning and development program through the organization. These things all go hand in hand.