800TB, Zero Pipelines: How Ensemble Health Partners Built an AI-Ready Revenue Cycle on Lakebase

Summary

Ensemble Health Partners manages revenue cycle operations for 30+ health systems, processing 25 billion+ transactions across 800TB of healthcare financial data previously siloed across 15 individual SQL Server databases with separate schemas.
By consolidating onto Databricks Lakebase as their AI-ready transactional layer, Ensemble eliminated brittle pipelines, enabled self-service pods to materialize views directly from harmonized Delta tables, and achieved a 3–5% net revenue yield improvement worth over $30 million per $1 billion health system.
The architecture supports agentic AI workflows for denial prevention, accounts receivable follow-up, and payer-behavior prediction, governed by a healthcare ontology and HIPAA-compliant Unity Catalog access controls.

800TB, Zero Pipelines: How Ensemble Health Partners Built an AI-Ready Revenue Cycle on Lakebase

Ensemble Health Partners manages revenue cycle operations for 30+ health systems, processing 25 billion+ transactions across 800TB of healthcare financial data. Legacy SQL Server sprawl —1 5 individual databases with separate schemas —c reated silos that blocked cross-client insights and slowed AI innovation in an industry where providers operate on 1% margins. this video shows how Ensemble consolidated onto Databricks Lakebase as their AI-ready transactional layer, eliminating brittle pipelines and enabling self-service pods to materialize views directly from harmonized Delta tables. The results: 3–5% net revenue yield improvement ($30M+ per $1B health system), 20% organizational efficiency gain, and 10x reduction in failure rates — while powering agentic AI workflows for denial prevention, AR follow-up, and payer-behavior prediction at scale.
Talk By: Grant Veazey, CTO, Ensemble Health Partners ; Surya Sai Turaga, Field Engineering, Databricks ;

Chapters

FAQs

What is Ensemble Health Partners and what problem were they solving?

Ensemble Health Partners is a full end-to-end revenue cycle management company serving 30+ health systems, processing over 25 billion transactions across 800TB of healthcare financial data. Their legacy architecture consisted of 15 individual SQL Server databases with separate schemas that created silos, blocked cross-client insights, and slowed AI innovation in an industry where providers operate on 1% margins.

What is LTAP and how does it work on Databricks?

LTAP (Lakehouse Transactional and Analytical Platform) is an architecture pattern that unifies OLAP analytics in Delta Lake with OLTP transactional workloads in Lakebase under a single lakehouse roof, eliminating the need for separate data movement pipelines between the two systems. In Ensemble's case, this allowed 800TB of healthcare data to be available for both real-time AI agent queries and analytical reporting without building or maintaining ETL pipelines between systems.

What results did Ensemble achieve after migrating to Databricks Lakebase?

Ensemble achieved a 3–5% net revenue yield improvement worth over $30 million per $1 billion health system, a 20% organizational efficiency gain, and a 10x reduction in failure rates. These results came from eliminating brittle pipelines, reducing the 18% first-pass claim denial rate through agentic AI, and enabling self-service analytics across all 30+ health systems from a unified data layer.

How does Ensemble use AI agents for healthcare revenue cycle management?

Ensemble built agentic AI workflows on Databricks for denial prevention, accounts receivable follow-up, and payer-behavior prediction, grounded in a healthcare ontology that gives agents domain-specific knowledge about coding, contracts, and payer rules. All agent access is governed by HIPAA-compliant Unity Catalog controls with full traceability of which agent accessed which data and why.

Full transcript

[00:07] Welcome to the session. Uh the session topic is um 800 terabytes zero pipelines. Um quick show of hands to the folks who have joined uh who were there for the last 20 minutes yesterday in the
[00:19] keynote. Awesome. Now you should be able to tie back to the um topic today how um Ensemble is able to bring together 800
[00:31] terabytes of data on your lakehouse and the lake base which is your OLAP and OLTP systems um under the single roof of lakehouse with zero pipeline build right
[00:45] um so it is also called as the LTAP um so um we are we are fortunate uh we do have um Grant who is the CTO of Ensemble. Uh my name is Surya. Um I lead
[00:58] the life sciences or the field CTO um at data bricks. I've moved into this role pretty recently as well. Um with that said um uh I would like to um bring up
[01:10] the agenda what we're going to discuss today. Um and we we would look into uh more on the revenue cycle management. Uh we will have Grant on the stage who will introduce himself. um talk about the
[01:22] company, who who who Ensemble is, what they do. Um and also we will get into some of a very technical details. Um we always celebrate the result. Uh but this
[01:34] session we will actually walk you through the journey. The journey um how how as a company we were able to innovate fast, how did we evolve? Uh what are some of the humps that we have seen? uh and we also will talk and
[01:47] showcase some of the examples around deep details when you're starting your journey towards agentic AI right uh and we will cover a little bit of history as well um how we were able to evolve um and help our customers throughout the
[02:00] board now that said uh I would love to welcome Grant on the stage um who's uh who is the CTO for Ensemble >> thank you Seria um glad to be here today
[02:12] glad to really be able to talk about our journey uh from a data platform to an AI platform. Uh, and what we've been able to do at Ensemble. Uh, just a little bit about myself. I've spent over 25 years
[02:25] my entire career in healthcare technology. I've been on both payment integrity side for payers for about 15 years and I spent about 10 years on the revenue cycle side uh with two different companies. I've worked with companies
[02:38] and four companies like Optum, UHG and now the last five years uh at Ensemble Health Partners. So excited to talk a little about that. Um just a little bit about Ensemble. We are a full end to-end
[02:50] revenue cycle management company. Uh our mission is to be able to take something that we believe is not core to a health system uh the revenue cycle and take that on and bring extreme expertise to
[03:03] that to do it at the highest level so they can focus on their mission and their mission is to provide the best care they can to their communities and their patients. um in in those communities
[03:27] systems. And without our services, they could be facing bankruptcy, merger. It's a very tough time financially for hospital systems. Uh today, one of the things that is kind of an overarching crisis in the health care
[03:40] system is the denial crisis that's going on. And for us, that means that's reimbursement that our clients aren't going to receive for the care that they provided to their communities. Uh and
[03:52] just some of the stats here from that perspective is about 18% of all claims are denial on first pass or first submission. That means 18% of all charges that a hospital sends to a payer
[04:04] for reimbursement uh get denied. And what you'll see from that, you guys can see a whole lot better than I can, is 65% of those never get uh challenged, if you will, from that perspective because
[04:16] it's a it's a laborious process to get around each one of those things. So really, we find ourselves in a place to help our clients be able to to recoup some of these losses from a denial
[04:28] perspective. and what's going on in the industry. Uh, and especially in the day of of AI today, if any of us looked at kind of the last earning calls from UHG, they touted that they're going to spend
[04:40] $1.5 billion this year on AI. Onethird of that will be directed towards their optim insight business is which where their payment integrity business sits. So $1.5 billion
[04:53] is more than the net patient revenue of over half our clients. So how are they supposed to compete in that world by themselves u from that perspective? So we sat around about five years ago to
[05:08] really build out our data platform. Uh so we built the foundation. So that's important part of our story is what we did to build out our data platform. We says here on the slide 30 plus systems we have 35 different hospital systems.
[05:21] They have multiple EMRs. We're talking about 60 plus EMRs. We're talking about hundreds of payers. We're talking about 10 plus clearing houses. That's all the data it takes to to bring together to be able to understand what are the challenges we're trying to do to help
[05:33] our clients through uh through this this battle they're in with payers from that perspective. So we set aside to bring that all together in a unified data model to bring that to bear so that we
[05:45] could build that foundation. And in all honesty five years ago when we started that journey that was a lot about analytics insights how do we look at it from that perspective. So we started about that journey and again most of our peers and somewhat
[05:59] ourselves are very siloed in this industry and specifically you look at rev cycle you can break that up into different pieces and so the silos are data sometimes around financial clearance coding uh backend denial
[06:11] management we were able to set off to make a unified view across a single view across all of revenue cycle and bring all those pieces together in one unified view from that perspective. So again your typical kind of lakehouse
[06:23] architecture that we'll talk about. So we went from the bronze level our our raw data ingest to silver where we harmonized that to a single enterprise standard our own ensemble schema and then lastly that gold layer we bring our
[06:36] enrichments to bear our intellectual property so that we can bring that and bring the best insights we at the time that we wanted from a revenue cycle perspective.
[06:50] But we solved the problem of analytics um and but analytics analysis wasn't enough and we wanted to move the world of action and how do we start to drive more and more of that action from that perspective. So we had a great
[07:02] foundation from that perspective but we started to see from our application layer and then specifically moved into agentic layer we needed transactional data uh from that perspective. So um little bit the
[07:19] so as we started to look what were the use cases we were seeing around that from from there is that while we had all the foundational data we needed when we started on the AI layer we started to see that hey when we started to hill climb through those processes that we
[07:32] necessarily didn't have the data we needed at the time we needed it and the freshness we needed it so we were creating all sorts of transactional data stores OLT TP on demand. We were building pipelines, custom pipelines one
[07:44] after another and it was creating a large engineering backlog from us. So enter into lakebase, right? And hopefully you guys have heard a lot about that over the last uh day or so,
[07:56] but that's our ability to create a synchronized version of that transactional data, be able to tap into that real time and bring that towards uh our ability to uh bring that to our
[08:09] agentic layer. So real-time freshness, zero pipelines and allow our agents to be able to consume that from the very beginning. So what we quickly found out is our data
[08:22] platform became our ensemble platform. So instead of just a data platform, it became the basis for our entire everything we were going to do in ensemble. So it became our agentic platform. So we had that base that Unity
[08:34] catalog at the bit but at the bottom so that we could have our categorization of data. We knew how we could start to bring in our annotation data, our classification data, bring it all into that layer, have that be consumable
[08:47] through lakebase into an OLTP structure and make that available for our agents in an API first uh process.
[09:01] So now we're moving from the a source stale snapshot analytic based to now an actionbased agentic form for us uh moving forward from that perspective. So if you can
[09:14] there we go perfect.
[09:28] meet our needs in a real time. And we'll go into a couple of use cases around that for us in real world activity. So we talked a little bit about denials and what we would historically do is take denials as a retrospective kind of view.
[09:42] So what's that mean? We would wait till a payer would make a denial. Maybe that denials on pre-authorization. We would do the discovery, bring that forward, understand that, and write an appeal letter and appeal that decision. But
[09:54] now, how do we start to prevent those things? So, how do we start to prevent denials? Well, now you have to start getting more real time. And that's where it really changes. So at the time of schedule registration now I can bring
[10:06] that same data that same intellect that same agent move that forward in the process to that moment of registration when that patient is sitting there about to face a clinical encounter and be able to run the analytics against that the AI
[10:20] against that to understand is a pre-authorization needed right at that moment and I can get the proper pre-authorization at that moment and possibly prevent that denial on the back end.
[10:32] Another great use case for us was around clinical what we call clinical reasoning. Our first step around clinical reasoning was about the ability again retrospectively. We've gotten a clinical denial. Clinical denial usually
[10:44] means a payer disagrees with a the medical necessity that was uh put forth by that physician and they have denied that claim. So, we were able to take our AI agent, make a um a clinical argument
[10:58] against that and do an appeal against that denial. But what about I want to move that more real time upstream into into healthcare. uh a lot of us may not know but if you are in an ED in
[11:11] observation and that physician believes you need to be upgraded uh into the hospital into an inpatient setting lots of insurance companies want to have a pre-authorization at that moment. So our
[11:23] ability to take that same AI, bring it up to instead of being a retrospective, bring that real time and help give clinical advice to that physician to make that clinical argument of why that upgrade from inpatient from observation
[11:36] inatient is necessary, provide that real time to payers while a patient sitting in a bed. So that's really really important for us to be able to have this turn this from an actionorient from a stale snapshot analytic view into a real
[11:51] time actionable view that we have that we can make it agentic. Our last use case is around something called DRG validation. DRG validation is DRGs are when you group episodes of care
[12:03] or uh different services uh under a same episode of care together. Lots of times it's a big opportunity for people to miss reimbursement at that time. So what we're able to do is real time during the
[12:16] actual coding activity be able to run our AI now right at the time of coding to make sure that those primary and second diag secondary diagnosis are not missed at any time from that perspective. So this is the ability. So
[12:29] all these use cases is what we have done in real world to be able to take that great data foundation that we've built make it into transactional datas that can be consumed by our agents and have them real time not retrospective but
[12:41] real time in the point of care across the spectrum to service our customers. With that I'll talk a little bit about that this has been a journey and a great partnership that we've had with data bricks. So I'm going to turn it over a
[12:54] little bit to Seria again to talk about how we've been on that journey and some of the things that how the data bricks products have allowed us to do those things that we just talked about.
[13:07] >> Awesome. Thank you Grant. Um so so let's let's let's live through this journey right um how how ensemble is able to to reinvent itself again and again and time and again right um if if you look at uh I would divide the journey of ensemble
[13:21] into six eras. um especially after the uh Game of Thrones episode yesterday uh we're back into the era right so um what what you see here today is the era one which is we call the before the era 2
[13:33] where the rise of the data lakes you know structured and unstructured data stored in the same object storage in your cloud um the the rise of lakehouse you probably have all lived through this
[13:45] um and um machine learning which is trying to predict the outcomes of your business and so on and so forth and then finally a leap um a small step and and a
[13:57] leap for the humankind also called as AI right gosh it's been 20 minutes and I did not use the word AI right um all right cool and then finally we're moving into the world of agent um so we will try to a little bit slice and dice uh
[14:11] this whole entire journey and how they were able to execute this uh execution is really key but um here we are in the agentic era where we are looking for software who's going to act on your data
[14:23] right that is agentic and how we made it possible at ensemble that's really the journey that we are going to look um now now we have looked at different companies right they all need these AI agents um you know they want to they can
[14:37] operate and they need to be able to operate safely inside real business processes right and ensemble deals with um hospital care providers actual PHI information and so on and so forth very
[14:49] very sensitive data Um hence they they need to find uh the groundness the the the real core of it is governance. So how they govern data um how how complex rules are implemented
[15:02] um and then finally who is accountable for these workflows right um and then when they started to standardize uh in era 3 on the lakehouse uh really it was a pivoting moment really it's a pivoting
[15:14] moment for the entire company uh because they can now suddenly experiment fast right fail faster right uh and guess what the best part is there is no high
[15:27] cost of experimentation. It's not zero but it is really really low. And how is it possible? Because now suddenly those 15 platforms that you have has basically consolidated into one lakehouse
[15:39] platform. So you can iterate fast uh on specific wings of those lakehouse copies and structures also called as branches in the lakebased world. Right? Um so
[15:51] really we were able to bring that cost nearly to zero cost of experimentation nearly to zero and hence they were able to iterate and we're going to share some of those technical details in the rest of the 15 to 20 minutes we have today.
[16:03] Awesome. So let's look at and and Grant mentioned this earlier uh in his talk. Um how did the pattern basically harmonized? What what happened? Uh and you can see there are gold data sets that were curated. So we are now
[16:15] basically raising ourselves ourselves up the stack. Um there is a lakebased sync pipeline as early adopters. Now that is gone with LTAP. Uh you probably have heard that yesterday. Uh and then uh
[16:27] they have a fully managed Postgress and OLTP which is actually backing their flagship product EIQ ensemble IQ right you can just look it up and you will you will land in that. Um and and Grant also
[16:39] mentioned that hey really what they suddenly started to realize is their data platform is actually the the platform itself that they are offering to their customers. So this is how they started to uh elevate themselves up the
[16:52] stack uh up to the application to the presentation layers right. Um now Ensemble basically migrated their data foundations of Ensemble IQ their flagship product into the data bricks
[17:05] platform right um so that naturally gave them the evolution naturally gave them the ability to evolve them bring in machine learning models bring in agentic AI and build those workflows right so
[17:18] now let's look into the actual agentic data platform so this is the recipe uh this is how and I can see um phones coming out. So I I'll take a second um so they can take photos. But um this
[17:30] this is what really they were able to execute through and through, right? Um and you can see the enterprise software stack uh is is long separated as a system of record. So you can you can find system of records in this
[17:43] architecture. You can find systems of intelligence, right? Uh the data intelligence platform if you're aware of that. Uh but really I I want you to um focus on what is called as a system of records that that basically I'm trying
[17:57] to flash my um laser but it's not going there. Okay. Um so what I really wanted to focus on is called as the agentic layer which is the system of can anybody take a guess execution.
[18:10] Agentic layer is your system of execution which is acting on top of your data, right? Um and the architecture changes. Um this whole agentic space that you see which is the agentic data intelligence platform is basically
[18:23] changing the role of AI in the enterprise, right? Uh it is able to act on everything that you see in those blue boxes. Um and and once software starts acting on your data, what you really
[18:35] need is governance, right? uh and it becomes more and more fluent on your business context the semantics uh and you know where the next slide is going uh but um I would like to pedal back
[18:48] before this this was kind of finalized and set in stone um we did we did iterate and I would I would like to show some of the facts that went uh behind these how what agents have used um what
[19:02] agents have applied uh and what what the systems and the agents basically touched pretty much. Okay. Now, there is some exceptions to to this architecture which we can discuss um probably as a Q&A, but
[19:15] really what I wanted to show you is the journey. Um and here we are. I want to welcome you all uh with this slide to a whole new world of loop engineering. How many of you have heard this term about
[19:28] loop engineering before? Fantastic. I do see a couple of hands up. That means um this actually predates and it's built on top of what is called as prompt engineering. And if you peril back in history, people were like, "Oh yeah, now
[19:41] with jai the world is going to be transformed. Everybody's going to be prompt engineers, right? Uh and and you started prompting your agents and so on and so forth." But the reality is and this was the test. This is how we set it
[19:54] up. This is how we loop through it. And you can see this end toend agenting architecture which is kind of multi- aent. So what you see here is um a single sequential set of tasks which is also called as lang chain set of events
[20:08] in a lang flow right which are chained together. You have an input prompt which does and executes a task and gives you an output prompt. But most of the AI
[20:20] pilots fail and why do they fail? They fail because the agent does not understand the business structure. They doesn't understand the underneath
[20:33] question. And when we were iterating through this exercise where you have a chain of events, you take a a module in your revenue cycle management and you try to put or throw agents on top of it.
[20:45] Let's automate it. Right? What happened? What happened? What did we notice? We notice that the models were able to produce very fluent answer. Absolutely, I can do that. Right? Thanks for
[20:58] bringing this. You probably are familiar with these adjectives that I'm talking about. Uh and then very very quickly we realize the answers may still be wrong. Also called as hallucinations, right? Um
[21:13] so the agent may treat an account, a customer, a client as a single entity as as as a company with the same thing with the same with each system is using the
[21:26] words differently though because now with the lakehouse when they were able to bring systems a definition of a client, a definition of a patient, a definition of um a customer in every module is different. when you consolidated them into a lakehouse
[21:39] architecture, the definition actually started to come together, right? So then the agents pretty much were able to uh almost act on this but they really did
[21:53] not know what those prompts uh actually are reasonable, right? So take an account, iterate through prior authorization and they start hallucinating, right? So um some of these business
[22:06] rules as they say um are needs to be validated. You need a human in the loop and so on and so forth. Also called as RLHF reinforcement learning with human
[22:19] feedback. So that was the last tail end of that third box that you see here um in the in the setup and testing. Now I want to fast forward to another term that Grant um used in his talk just now
[22:33] called as hill climbing. Anybody who has not heard the word hill climbing okay so hill climbing climbing is basically a set of exercise when you're trying to put together an agentic
[22:46] workflow an agentic task when you're trying to stitch them as chains it is one of those critical and crucial exercises and I'm going to show you now. So I'm going to move off the slide and show you how in the presentation mode um
[22:59] this all came together uh and the rise of agent ops started to become a reality right um and I don't see that slide on on
[23:11] okay u let me see how what I can do
[23:23] awesome this is So,
[23:42] All right, let's look at the entire agent ops. And when you have an agent in your system which is acting on your data, acting on your software, acting as a headless user, uh what goes behind, right? So you have two new personas who
[23:55] are interacting on top of your user base. One is called the AI prompt engineer. So those of you who you said yes I I know prompt engineering. Welcome to loop engineering. Um there is an AI
[24:08] prompt engineer in your system right now. I was trying to act on your data. There is a tool engineer who's building tools or improvising those tools in your system as well making it better and better. Now they go through what you see
[24:21] on the top and the bottom is the hill climbing exercise and soon enough you will notice that um there is a loop that sounds to form and we have able to what I call as a six-step process how this
[24:34] agent ops should actually operate and while ensemble was able to squash these features into into single target you have a candidate branch how it is served
[24:46] all of them is tied together and stitched together with full auditability. ility eval trace and benchmarking on the managed ML flow on data bricks. Um so once you have this entire loop on the left hand side of the
[24:59] agents ops that we kind of went through it is ready for serving it is ready for RHF. Now let's let's zoom out a little bit. Now let's look at the right side of the agent ops because while the engineers
[25:12] and prompt engineers and tool engineers are iterating through those loops um what is actually happening what are they doing what is their dayto-day right the day-to-day pretty much changes so what is the SDLC cycle pretty much changes to
[25:26] is the prompt engineer is is running those models locally so you can see on the top um and then they are taking a baseline eval then they spin up a subset of those eval files and then they do a
[25:38] partial run um of those similarly capturing feedback and then continuously committing and pushing that code to a remote version of those of that model essentially. So this entire cycle of
[25:52] agent ops needs to come together somewhere and where do you document these steps? How do you monitor this? How do you audit and report all of these capabilities? So Ensemble was able to
[26:04] iterate through their agentic system very very quickly with partial eval to full eval um building those pipelines on the data bricks with MLflow. In fact
[26:16] some of those inputs actually led to MLflow 3.1 update in the open source community. Um so thanks for um the users and and u folks who provide um feedback to to us. Now let's come back to um the
[26:32] actual hill climbing exercise. So now I'm going to go back to the presentation mode. All right. Here we go.
[26:59] Okay. Something happened.
[27:29] Sorry about that. >> Go back to where you were. No, no, where you were before.
[27:41] >> Awesome. All right. So, um let's quickly give you a walk through of so we went through the hillcoming exercise. So when you're trying to put together um an agentic AI uh system onto suching a system of processes and datas you go through this pretty much six processes
[27:54] right uh so make sure you get your one manual run reliable once you do that you turn that run into a skill once you have done that you convert that into a state file and then you have what
[28:07] is called as the hard gate make sure your baseline evals and the partial eval are always uh matching the business needs and requirements. So that's the hard gate. That's where you pretty much go on no-go your newly built uh tool or
[28:23] the or the prompt engineers finally tuned model for the entire agentic stack to work with your business processes and then welcome to what is called as write it in a loop and then schedule it. Right? So this is now the new thing
[28:36] which is called as the loop engineering. Then obviously you need to have what is called as a failure mode um and and the observability that comes with the stack of it. So um after all of this what
[28:48] really um we want to show a demo but I think we are uh gone time because of the logistics but I can show you offline how the entire um um revenue cycle model with specific modules like AR follow-up
[29:01] utilization management um cash viewer and and claims denial management all of those stats and analytics were performed using AI um and on on a continuous basis and being available into dashboards
[29:15] um on the data bricks app stack. So that's the demo that I have. Um hopefully we'll have some time to the end and I can I can show those. But what you really saw is the agents the domain model the prevention not rework ROI um
[29:29] is all displayed and uh it is part of the agent data platform. Uh and again this has not come together um just just like just like that. It has taken years of consolidation into the lakehouse
[29:42] stack and once that is there they were able to really merge the old TP and OLAP uh without writing any data pipelines and surface all the analytics out into the data bricks apps. Um so uh with that
[29:55] there is some future work uh from on top of the agentic system that we talked about and what we what we really wanted to um showcase is how this stack is going to evolve and you can see
[30:07] similarities between what you have seen yesterday with genie ontology right so genie is is the heart of your agentic work layer and what you're actually seeing is what we call it as the ontology binding right so you what you
[30:20] see at the center where your data platform your agentic data platform the governance is now getting glued with this agentic ontology. Um and what what is actually happening is these models
[30:34] are are defined on what entities exist, what is the definition of a customer, what rules govern when you act on these data sets and what actions are actually
[30:46] valid, right? Um and finally these these language models propose an answer with a plan which is what you see here as page ranks. Um and probably have heard that.
[30:58] Um and all of these ontology with knowledge graph all the different policies the source of truth of data um that they that is there and and is grounded uh valid and safe for agents to
[31:10] act upon. Um so we talked about how basically uh these work. So I'm going to just quickly mention um how this pieces will actually come together and what the future work is um going to be and you
[31:23] have heard about onto ranks yesterday um and and what you're able to do with this whole process of building the gene ontology in between the agentic layer and the data layer um is you're able to reduce these duplicate records. So
[31:36] imagine every business person asking the same question. How do you reduce that? You reduce the prompt size with the ontology on top of it because they exactly know what it how many halves it will take for an agentic system to get
[31:49] to to build the prompt in the most recent in in the most efficient manner as possible. But why is this important? Why is ontology important? Because your context length is limited. And what is
[32:01] context? Context is tokens and hence token maxing, right? Um so this is how it all comes together. And what you're looking for really is reducing all the manual reconciliation. Um reducing
[32:13] invalid actions with this knowledge extraction. Um you're pretty much looking at zero ops uh if you were there uh yesterday. So that's a critical part of a future work. Um and then also
[32:26] reducing what all the audits and decisioning that happens what on this tab what is called as what you would like to know the prompt itself right uh and overall once you're able to tie together you reduce your operational risk significantly uh while bringing
[32:40] Genie and Genie ontology on top of your agentic data platform so that's really where we are headed as next step uh with ensemble awesome so u I want to make sure that I hit home with this u this is
[32:53] not something that you need to stitch together. It automatically autolearns at the enterprise level which basically means the enterprise stack right uh which has long been separated with your
[33:05] system of records at somewhere um and your system of intelligence somewhere is now finally getting connected uh without stitching together too many services and so on and so forth. So agentic layer now
[33:17] suddenly becomes your system of execution in the entire stack. All right. So that's where the multimodels and and and other things come to life. Now having that said that I would love
[33:29] to go back to Grant to um provide some closing thoughts um on how the company is able to move fast. >> Yeah, I think a lot of great points that were talked about. Uh the way we organize our teams at Ensemble is in pod
[33:43] structures and we really want to be able to do is kind of recreate that feel of a startup. So we want to have a complete team. Our product, AI, engineering, those entire teams are encompassed, focused on a single task. And with that,
[33:56] we want to be able to give them every tool they need to be able to solve that problem. What you heard a lot of that was just described is a big kind of blocker for that had been getting them the data they needed when they needed it
[34:08] and the freshness they needed it. with these tools. Now, what we've been able to do, we can now be able to let those T teams be more self-service, be able to go through those hill climbing exercises that you heard described uh through
[34:21] faster iterations. So, what maybe took, you know, two weeks to build a data pipeline, create a new transactional store, maybe that happens in minutes or hours. And now I can go through five or six of those iterations and take what would have taken two weeks, I mean, two
[34:34] months down into a couple of days. And what that allows I think everybody that's going to need to be able to do is is that level of experimentation, right? The that's the hill climbing exercise that that you heard described, that's kind of where the secret sauce to all of
[34:47] agentic lives in my mind is the ability for you to go, hey, what are the are we getting closer to that that gold set? Are we getting closer to that quality mark? Uh are we getting closer to that overturn rate? Are we getting closer to
[34:59] that clinical diagnosis? Whatever that is that we can hill climb for that, then we can iterate. And what we'll tell you is that through our experimentation through that what we thought were the the data the the signal that we needed
[35:11] in AI to be able to solve those problems very rarely ended up being the exact signal we needed. So through experimentation we found that a lot. So what that's really been a gamecher for us is that is that catalog data platform
[35:25] is now being able to be real time available to our agent layer and it's going to be an absolutely already is and going to continue to be an absolute gamecher. Some of the thought that you talked we talked about as future works.
[35:37] We're super excited about that ontology in healthcare. If if you know ontology is gold and making sure that you can bring that clinical ontology together, that administrative ontology is really the secret sauce to being able to get
[35:50] these agents to start to to deliver the promise that they have. And so kind of the the last on it is um my teams hear me talking about every time compliance and security is always
[36:03] treated like a first class citizen in everything we do. We look at that first. So we've got to make sure that we're HIPACO compliant. We're aligned with our high trust. We're we're in compliance for our AI agents have the right
[36:15] traceability and observability and auditability of the results that they had. So that is that has been another part of this that we can kind of bring all of that lineage and have that being monitored through our unity catalog as well now that this is one single
[36:28] platform and not having to measure that in individual transactional data stores as well has been has been huge for us. So those are the kind of key components uh from that perspective that have been really exciting to us uh at Ensemble. So
[36:43] kind of my last closing uh statement on what we've been able to do and I kind of hit it before. We're not we're no longer building a data platform. That's what we started. That's the journey we started three years ago. We are now building an
[36:56] AIdriven operating system for all healthcare revenue cycle. And so this is now one seamless platform that we're able to bring all together. And I the the new theme that everybody understands today is that your AI is only as good as
[37:10] your data. And those things being connected are what have to happen from from that perspective. So we're really excited about what we've done. We're even more excited about the future holds for us and what we're able to do uh and
[37:22] the services we're able to build to our clients and healthcare in general. So with that, I'll say thank you and I don't know We got a couple minutes for Q&A or >> Yeah. So, uh, thank you very much. U
[37:34] Yep, we do have some time for questions, so we're happy to take questions. Yes.

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.