Sponsored by: Domo | From Prediction to Action: Agentic AI for Ambulatory Care on Databricks

Summary

Domo and Databricks combine to address the ambulatory care no-show problem, with Databricks powering real-time predictive intelligence from appointment, weather, and traffic data and Domo operationalizing it through agentic AI workflows that trigger SMS outreach, overbooking recommendations, and staff alerts.
This video states that patient no-shows and late arrivals cause $8–9 million in annual revenue leakage per health system, and demonstrates how a closed-loop system can identify at-risk appointments and initiate intervention within a critical three-day window.
The architecture includes human-in-the-loop review and approval for agent decisions and is fully governed and auditable, enabling healthcare organizations to achieve measurable ROI within weeks rather than months while maintaining compliance with enterprise oversight requirements.

Chapters

FAQs

What is the no-show problem in ambulatory care and why is it so costly?

Patient no-shows and late arrivals disrupt clinic scheduling, create idle provider time, and generate lost revenue that this video quantifies at $8–9 million annually per health system. The problem is compounded because most systems only report the issue after it has already occurred rather than identifying at-risk appointments early enough to take preventive action.

How do Databricks and Domo complement each other in this healthcare AI solution?

Databricks provides the real-time predictive intelligence layer—ingesting appointment, patient, weather, and traffic data and scoring no-show risk using AI models. Domo operationalizes those scores by building agentic workflows that decide the next best action and execute it, such as triggering SMS outreach or alerting clinic staff. Together they form a closed-loop system from prediction to action.

What data sources does the no-show prediction model use?

This video shows the prediction model drawing from live appointment data, patient history, weather data, and real-time traffic conditions. Combining these signals allows the system to assess no-show risk dynamically as conditions change rather than relying on static historical patterns, improving the accuracy of risk scores close to appointment time.

How does human-in-the-loop approval work in the Domo-Databricks agentic workflow?

When the Domo agent identifies a high-stakes action—such as authorizing overbooking or sending patient communications—it surfaces the proposed decision to a human reviewer for approval before execution. This video frames the human-in-the-loop step as both a governance requirement and a practical safeguard, ensuring that autonomous decisions align with clinic policies and that staff retain accountability for patient-facing actions.

Full transcript

[00:07] How's everyone doing? Is everyone exhausted yet? I'm pretty impressed. I see some of your swag bags are are bigger than your children. Um I'm going to give you an insider tip. I've been coming here for about the past 3 years. Wear the swag to the airport.
[00:20] Not only will you get it all home, you will be the most interesting person on the flight. Um but thanks again. Thanks so much for coming and spending some time with me. Um today we're going to talk about how healthcare organizations can move from
[00:32] prediction to action using Databricks and Domo. Specifically, we're going to look at how AI agents can help solve one of the most expensive operational challenges in ambulatory care, patient no-shows and late arrivals.
[00:44] So we'll show how Databricks provides the intelligence, Domo operationalizes it, and together they create a system that can proactively identify risk and take action before revenue and patient outcomes are impacted.
[00:57] I'll just do a quick intro. My name is Dan Wentworth. I lead a team of solution engineers at Domo focused on helping customers maximize their investment in Databricks. And over the past year we spent countless hours helping customers
[01:09] and prospects move beyond dashboards and reporting into AI-powered applications, workflows, and agents that can actually drive business business outcomes. And today I'm going to show you a healthcare solution that does just that.
[01:22] And for those of you who aren't focused in the healthcare sector, um after the presentation ends, please march over to the Domo booth and ask the team to show you solutions across other verticals, manufacturing, retail, CPG, fintech,
[01:34] hospitality, or even horizontally across the business. So, operations, finance, sales, supply chain. I promise you won't be disappointed. We've built a ton of really cool solutions leveraging the infrastructure and and AI uh assets in
[01:48] in Databricks. Um so for the agenda real quick, I'll just start with a brief introduction to Domo and how we both complement and amplify your investment in Databricks. We'll about the ambulatory care no-show problem and why it's such a costly
[02:01] challenge for healthcare organizations. And from there, we'll look at how a genetic AI can help address the problem. So, at Domo, we describe ourselves as the genetic platform for the intelligent enterprise.
[02:13] And that's because we're focused on helping customers operationalize AI, not just analyze data. We help customers move beyond analytics by building intelligent applications, workflows, and agents that create measurable business value.
[02:26] But honestly, the challenge isn't creating intelligence. The challenge is getting that intelligence into the hands of the people who need to act on it. And that's where Domo comes in. We help organizations connect data, AI, applications, workflows, and agents into
[02:39] a single platform that can drive action across the business. Because ultimately, intelligence by itself doesn't create value. Value is created when that intelligence changes a decision, triggers an action,
[02:51] or improves an outcome. So, we think about this journey to becoming an intelligent enterprise in three phases. And the first is foundation. Before AI can create value, you need trusted, governed, AI-ready data.
[03:03] Domo helps customers connect, integrate, and transform data across the business with the lakehouse serving as a governed foundation. Second is activation. This is where most organizations struggle. They have data,
[03:15] they have AI, but they don't have a way to operationalize it. The goal isn't just to analyze information, it's to create intelligence that can actually be used by the business. The goal is to drive action that results in business outcomes.
[03:28] And finally, distribution. Insights don't create value sitting in dashboards. They create value when they're delivered into the workflows where people actually do work. Intelligence has to reach the people
[03:40] making decisions. Mobile users, schedulers, nurses, operations team, frontline employees, the people closest to the outcome. And the most important part, all of this runs on Databricks, on your cloud, on
[03:53] your cloud governed by the controls you've already established. So, if there's anything you take away from this presentation when you think about Domo moving forward, it should be the following. And the first message is simple. Domo is the fastest path to getting line
[04:06] of business data into the lakehouse. And I want to be clear here. This isn't usually a problem with core systems. Most organizations have already connected the ERP, CRM, and major enterprise platforms. The challenge is
[04:18] the long tail. Departmental systems, scheduling platforms, marketing tools, operational applications, spreadsheets, the dozens if not hundreds of systems the business relies on every day.
[04:30] And as AI becomes more important, this challenge becomes even bigger. Because AI is only as good as the data behind it. If the data isn't in the lakehouse, your agents don't see it. Your models don't see it. Your users don't benefit from it
[04:42] at all. Domo provides over a thousand pre-built connectors, low-code transformation capabilities, orchestration and reverse ETL that help customers move from silo data to a governed lakehouse quickly.
[04:54] But getting the data in only solves half the problem. The second message to take away is that Domo is the lakehouse activation layer for your business users. Domo deploys intelligence from the lakehouse to the edges of your
[05:07] organization. Many customers have already invested heavily in Databricks, but they're still asking the question, "How do I get these capabilities into the hands of my business users?" Remember that most people making operational decisions every day aren't
[05:19] always sitting inside Databricks. They're working in scheduling systems, CRMs, mobile devices. They're helping patients. They're serving customers. We help organizations deploy intelligence from the lakehouse into the workflows where business users are
[05:32] already working. We can employ intelligence to We can deploy intelligence into applications. Agents, approvals, automations, alerts, next best actions. Because intelligence only creates value
[05:44] when somebody acts on it. So, the easiest way to think about our partnership is the following. Databricks creates intelligence, Domo operationalizes it.
[05:58] Databricks is the intelligence engine. It's where organizations unify data, build predictive models, create AI agents, manage governance through Unity Catalog. Domo is the path to outcomes. We take that intelligence
[06:10] and we put it in the hands of business users. We orchestrate workflows. We support human-in-the-loop decision-making. And ultimately, we turn that intelligence into measurable business outcomes. Databricks tells us what is happening
[06:22] and what should happen next. Domo helps ensure something actually happens. That the recommended action is actually executed. So, let's transition to our healthcare use case.
[06:34] The challenge in healthcare isn't a lack of data. It's not a lack of dashboards, and increasingly, it's not even a lack of AI. The challenge is turning intelligence into action quickly enough to change an outcome.
[06:46] Specifically, in the case today, reducing patient no-shows, minimizing provider delays, and improving clinic throughput. This is where Agentic AI becomes really interesting. Not because it predicts something, but
[06:59] because it can actually do something about it. So, let's start by looking at the size of the problem. At a first glance, a no-show doesn't seem like a major issue, but at scale,
[07:11] the impact becomes enormous. Across the organization we partnered with to build this solution, we were shown that roughly one out of every five patients either arrives late or doesn't show up at all. That creates significant variation in
[07:23] clinic throughput. Providers get delayed, schedules get compressed, and now the staff spends their day reacting instead of executing. And when you aggregate that across dozens of clinics and hundreds of providers, the financial impact becomes
[07:35] very real. We're talking about 8 to 9 million dollars in annual revenue leakage. The key point is that by the time someone notices the problem, it's already happening. The missed appointment has occurred, the provider schedule is already
[07:47] disrupted, the revenue opportunity is already gone. So, why is this still such a difficult problem to solve?
[07:59] Well, first, no-shows aren't random. They're influenced by weather, traffic, patient history, appointment type, travel distance, and dozens of other variables. Traditional reminder systems don't account for any of that. Second, providers often don't know
[08:11] they're falling behind until they're already behind. By the time someone realizes throughput is slipping, the backlog is already forming. Third, data is fragmented. The organization we partnered with had grown through acquisition, different
[08:23] systems, different definitions, different operational processes. And finally, the response is typically manual. Staff spend hours every day pulling reports, calling patients, coordinating
[08:35] schedules. It's expensive, reactive, and impossible to scale. What's interesting is that none of these are really technology problems. They're decision-making problems. So, instead of asking people to react
[08:47] faster, what if we let an agent do the reasoning for them? So, this is the solution we built. It's an AI Well, it's a Domo-powered AI agent running on Databricks. The agent continuously evaluates live
[09:00] appointment schedules, patient history, weather conditions, traffic patterns, and operational data. Its first responsibility is understanding risk. Who is most likely to miss their appointment? But, prediction alone doesn't create value.
[09:13] The The next step is action. The agent determines the most appropriate intervention and automatically initiates outreach. Some patients receive SMS, others receive email, others may trigger waitlist notifications or overbooking
[09:25] recommendations. At the same time, providers and clinic managers receive real-time visibility into emerging bottlenecks. The key different The key difference is is this just isn't an AI model, it's an operational system. It
[09:38] senses, it reasons, and then it acts. And when you turn prediction into action, the economics become very, very compelling. So, the value isn't coming from one massive improvement, right? It's coming
[09:50] from lots of small improvements across the operation. Recovering a portion of missed appointments, reducing idle provider time, smoothing throughput variation, eliminating manual administrative work.
[10:02] When you add those together, the organization is addressing roughly 8 to 9 million dollars in annual operational impact. And what's more important is that this doesn't require replacing existing systems. It sits on top of Databricks, uses the
[10:15] data already available, and turns that intelligence into action. So, again, I just want to reiterate, the breakthrough isn't that we can predict the probability of a client to no show. The breakthrough is the action that is
[10:27] taken in response to that probability. So, I'm going to jump over here to the application. Um, I know we don't have a ton of time to explore explore all the features of this solution. Um, it's something
[10:39] that we we built with a customer, and obviously we've done, you know, several iterations. Right? They see something really cool, we use the agent to streamline this process. Let's think about using this in another area. So, it's kind of built over time.
[10:51] Um, but I do want to highlight a few key areas where we're leveraging, right, AI within Databricks uh, to power the solution. Um, let me switch over. Let's go here. Um, keep in mind this healthcare the
[11:04] healthcare system we worked with operates about It's like 25 plus clinics and hundreds of providers. And obviously this is, you know, we we mix modalities here. There's a very bread-and-butter analytics experience.
[11:16] So, at any given time, leaders can see things like clinic throughput, provider efficiency, appointment risk, and operational bottlenecks. But we've combined that with different AI modalities, where you have things like conversational analytics that you can
[11:28] embed within the platform. Our dashboards are and applications are very composable. Not only can you add the pretty charts and graphs that tell you give you the insights, you can add in a conversational analytics piece where you're chatting directly with Genie
[11:40] asking it questions. You can add in agents that will summarize data and give you next best recommended actions. What should I be thinking about? Why are some of my metrics dipping? So in this in this particular example,
[11:53] I'm looking here at basically a provider comparison hub and I've narrowed it down to a few of our different clinics, Lakeview and Midtown. And I can also look at different metrics. So I could start with you know visit counts. See my top five providers
[12:06] by highest visit counts. Maybe I want to see medium provider time, percent of breach, high no-show risk, composite throughput score. If I go to no-show risk and I notice Dr. Hannah Cooper at the top, I may say, you know, I want to
[12:18] come down here to to Hannah Cooper and I want to run an AI insights analysis, which will actually hit a model within Databricks and it's going to look at all the data provided about the appointments and it's going to basically give you information on how
[12:30] that provider is performing. Obviously, you can tailor the model to give you summaries and outputs that that make sense that that you're trying to capture. In this one, it simply sent me an email, right? I set up a workflow
[12:42] that when I trigger that AI insights, the model runs, it gives the output, generates the text, and then sends an email to the user. So in this one you can see there's key insights, protocol breach concerns, no-show risk management, throughput opportunity,
[12:54] right? Dr. Cooper's composite throughput score 77 places her in the middle tier of the network and within her specialty. You can look at peer comparison with the ortho specialty ranking, visit volume analysis. And then down here you have a spot for recommendations.
[13:07] And obviously this one's a little bit lengthy. Again, you can tailor this to be very specific with what you want the the agent to summarize, what you want to know about the provider. But it shows you down here, right? What are the different priorities? What are the things I should be focused on with
[13:19] this specific provider? Where can we focus to increase that throughput composite score? Or to decrease the number of no-show appointments. And then the second place I want to focus on is is really the agentic piece
[13:31] we talked about. It's the ability to not only leverage intelligence within Databricks to give a probability to a no-show or a late arrival, but it's the ability to build a Domo
[13:44] workflow on the tail end of that that says, "Okay, I have the prediction. They may no-show. What action do you want to take? What action should I take as the agent?" So, I'm going to jump over here into kind of our workflows interface.
[13:59] It looks a little little complex at first, but I'm going to show you here at the top. There's this agent AI. This is kind of the scheduling piece. I have it open up here on another tab. Let me see if it it pops up. Okay, maybe it's not open there. So, I'm just going to I'm going to open here. We're going to edit the agent task.
[14:12] And what you're going to see in here, we have several different pieces that help this agent do exactly what we needed it to do, which again is scoring kind of the probability of someone to make it to their appointment,
[14:24] and then also taking action via the workflow. Okay, so that's not I'm just going to I'm just going to scroll through here. So, what you can see, it says here, you and these are instructions to the agent, right? It says, "You have access to three inputs. You have internal appointment data.
[14:36] That's risk score, no-show, age, age band, SMS opt-in, lead time hours, and clinic zip. You have today's live weather conditions for each clinic. You have today's live traffic conditions." So, one of the cool things you can do with these agents is
[14:48] you can provide them with tools. In this case, I've given it an API package where it can query these data sets directly for appointments, weather, and traffic data. So, we're not just scoring patients in isolation. This isn't just rule rules-based scoring. We're
[15:01] reasoning across operational context. I can also add knowledge in here. So, I can add documentation or file sets. These could be proprietary best practices. This could be information
[15:13] that lets the agent know how we act in certain scenarios. If there is a severe no-show, what actions do we want it to take? So, it can read through those files. It's almost like training like a rag model where it now knows to respond within the guidelines of your best
[15:26] practices at the at the healthcare provider. So, you give it the tools, you give it the knowledge, you give it the prompt what you want it to do. So, if I scroll down here at the very end, right? And I can even constrain the the context of the schema that it looks at. So, I know
[15:39] that it's only going to be querying data that I wanted it to have access to. But down here at the end of the day, right? It's it's evaluate each appointment by considering the weather, traffic, patient profile, model risk scores, and exist a recommended action.
[15:52] And down here I even have decision guidelines. I have send an email when the proactive confirmation or alternate scheduling is needed. Send a text when quick same-day reassurance or reminder would suffice. No action only if risks, weather, and
[16:05] traffic appear stable and low impact. So, I kind of simplified this. The one we built with the customer that it's a little bit more complex. There's more outcomes that can happen. I kind of simplified this example down to send a text, send an email, or do
[16:17] nothing. And so, if we look at the actual workflow itself, what's great is again, as I mentioned earlier, the intelligence is great, but it doesn't mean anything. It doesn't add value unless someone's acting on it. In this case, the agent's going to act on it because after the agent task, I've built
[16:29] a I've built a workflow down here. So, depending on what the agent determines, you can see a workflow branches here. Is it SMS or is it email? If it's SMS, send it, we're done. If it's an email, come down here. If the patient doesn't respond, let's out let's overbook and
[16:41] let's reach out to the waitlist to get someone else scheduled. If they do respond, if it's a yes, then we're good to go. If it's a no, send a list of upcoming available appointments so they can reschedule. So again, the real piece here
[16:55] is that you have an an agent making a decision, but based on that decision, you now have a workflow that's taking action. And again, these can get more complex. We can have more outcomes, but this is very simply showing I don't need a human in the loop any longer. And for
[17:07] the example with our customer, we did have them actually in the loop, so we we provided a stopgap at every point where the agent was making decisions, and we spot-checked it for a few weeks before we decided this is actually working
[17:19] really well and we could take that human in the loop out. But you can input human in the loop actions in here if you wanted someone to approve something within a queue, a decision that an agent made or someone reviewing actually say yes, that's the right decision. Let's
[17:31] move forward with that. You could email that approval, you could send a text message, you can notify them via an alert and they can actually go in and and approve that. So this is really the big thing we wanted to show. The the main point here is agents are great, intelligence is
[17:44] great, but how do I automate the action? How do I save people time? And how do I save money doing that? Um I'll kind of close here, right? The healthcare industry it really it
[17:56] doesn't have a prediction problem. The predictions are easy, it's an action problem. Right? Most organizations already know no shows are happening. The challenge is identifying them early enough to intervene. And by combining Databricks for prediction, Domo for operational
[18:10] execution, we're able to move from insight to action before that bottleneck occurs. And that's how we turn an 8 to 9 million annual problem into measurable operational improvement. So if you'd like to see the full
[18:22] architecture, we're kind of running out of time here, or the implementation approach, you want to see the full application, please stop by the Domo booth. We can we can talk to you how this pattern would apply in in your industry. Really the pattern of an agent
[18:34] decisioning and then actually building a workflow that takes action. And we'd love to you know, continue the conversation with all of you. So thank you so much for coming today.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.