Accelerating Industry Solutions Leveraging the Power of Agents on Databricks

Summary

Accenture's Framework for Agents is 100% Databricks-native — using Lakebase for agent session state, Databricks Apps for user interfaces, Unity Catalog for governed data access, and AI Gateway for model routing — and turns AI development from a bespoke project into a repeatable, modular platform that teams can update independently as the technology evolves.
The framework structures data platform modernization into seven modular steps from discovery through data model design, source-to-target mapping, PySpark code generation, validation, and deployment, all accelerated by agentic AI to increase developer productivity.
Accenture's Smart Coverage Assistant, the first industry solution built on the framework, targets healthcare insurance eligibility and cuts average patient hold times from 15–20 minutes down to 5 minutes while achieving 93–95% grounded accuracy on complex coverage questions.

Accelerating Industry Solutions Leveraging the Power of Agents on Databricks

Accenture's Framework for Agents turns AI development from a bespoke project into a repeatable product motion. It gives teams a reusable foundational layer that standardizes ingestion, retrieval, evaluation, guardrails, and cost governance, allowing developers to compose vertical solutions instead of rebuilding plumbing.
The framework is 100% Databricks native: Lakebase for agent state, Databricks Apps for user surfaces, Unity Catalog for governed data access, and AI Gateway for model routing and policy enforcement. Every deployment is tuned against business outcomes.
Healthcare is the first industry solution built on this framework, targeting a major pain point: patients spending 15 to 20 minutes on hold to understand coverage. Accenture's Healthcare agent cuts this wait to just 5 minutes, answering complex eligibility questions in real time with 93 to 95% grounded accuracy. In this video, Accenture shares a concrete blueprint for turning industry agents into a repeatable platform.
Talk By: Prakash Trivedi, Managing Director, Accenture Data & AI, Accenture ;

Chapters

FAQs

What is the Accenture Framework for Agents on Databricks?

The Accenture Framework for Agents is a reusable foundational layer that standardizes how enterprise AI solutions are built on the Databricks Data and AI platform, covering ingestion, retrieval, evaluation, guardrails, and cost governance so developers can compose vertical solutions instead of rebuilding infrastructure from scratch on each project. It is 100% Databricks-native, using Lakebase for agent session state and metadata context, Databricks Apps for user surfaces, Unity Catalog for governed data access, and AI Gateway for model routing and policy enforcement. The framework is designed to be modular so that individual components can be updated independently as Databricks capabilities evolve.

What is the Smart Coverage Assistant and what problem does it solve?

The Smart Coverage Assistant is a healthcare industry agent built on Accenture's framework that answers complex insurance eligibility and coverage questions for patients in real time, eliminating the need for patients to wait 15 to 20 minutes on hold for a human agent to look up their coverage across multiple contradictory data sources. According to this video, the agent achieves 93 to 95% grounded accuracy on coverage questions by using built-in RAG infrastructure with no external retrieval dependencies and responsible AI governance monitoring for all inputs and outputs. The result cuts average wait time to 5 minutes while maintaining the accuracy required for healthcare use cases.

How does Lakebase support agentic AI workflows on Databricks?

Lakebase is used in Accenture's framework as the transactional data layer for managing agent session state and metadata context during autonomous task execution — keeping the agent's working state active and consistent across the steps of a long-running workflow. This video explains that Lakebase's managed Postgres capabilities make it a natural fit for the stateful, real-time read-and-write patterns that production agents require, which differ from the analytical read-heavy patterns that the main lakehouse is optimized for. Together, Lakebase and the broader Databricks Data and AI platform provide the full data infrastructure stack needed for agentic applications.

Why does the framework need to be modular?

The framework is modular because the Databricks Data and AI platform evolves rapidly, and any tightly coupled architecture would require rebuilding large portions of the solution every time a new capability ships. This video explains that by separating the seven data platform steps — discovery, data modeling, source-to-target mapping, code generation, validation, deployment, and governance — teams can update individual steps independently without touching the rest of the solution. This modularity is described as a requirement for any enterprise-grade agentic framework that needs to stay current with the pace of platform change.

Full transcript

[00:08] I'm Prakash. Uh I'm managing director Accenture AI and data and I'm also datab bricks MVP. Uh today I will be talking about uh excellating industry solutions leveraging the power of agents on data
[00:21] bricks and while I talk about this I will use smart coverage assistant as an industry use case that was deployed using 100% databris native solutions on
[00:36] the agendics but before I do that there are many building blocks of creating such industry use cases and the two major building block for such thing is one the data platform and second the AI
[00:51] or agentic AI foundation layer. I'll start with data platform modernization. How did we first enable the data platform leveraging the agentic solutions to increase the developer
[01:04] productivity? So driving or leveraging agentic AI solution for data product development on data bricks that's the first building block and you can see why we need this
[01:19] what is this needed for around automated core life cycle around the various deliverables that needs to be accelerated and how we actually developed it using agentic AI layer
[01:32] using platform and tools and the differentiators that you see on the screen I'll go on to talk about the first the set of the seven steps that leads us to create the data platform
[01:46] >> that not only is modernized to the recency of the technology advancements that we have in this space but also ensures that these seven steps keep the solution very modular, very flexible and
[02:01] very editable. The moment there is a changes there are changes in the technology you should be able to work independently on these seven steps to modify it as the technology demands. So
[02:14] from discovery to data model conceptual logical and physical to source totarget mapping and all the way and to uh pispar code and then validation and deployment. The seven steps takes care of everything
[02:28] and this is a larger overview of what are the major eight steps that leads to data modernization implementations. So you can see that
[02:40] there is a source that you need to ensure that it is there is a seamless uh connections to that there is a datab bricks app that host the entire solution and leverage the unity uh enabled security features. Then you develop the
[02:54] agents. Uh the agents are the ones that does the majority and the bulk of your workload that it develops and create the data products and the data platform uh
[03:06] for you. to be able to enable all of these lakebased a perfect solution that not only keeps your session data active when the agents are running but also it
[03:19] gives a very right context to the agents while this is running. uh storing in it all the meta data that you need for this to give you the results that you need
[03:31] with respect to the data platform the very custom platform that you are developing for your enterprises. Then we have the model subbing. Uh that's where all of these abstraction
[03:43] comes in and all of these uh models gets hosted so that your developers are agnostic to what model and how they have been hosted and they can go and start leveraging it. You have data bricks
[03:56] unity catalog that eventually provides the context back to the lake base and the agents and then there's a dashboard to ensure that everything is rightly auditable end to end. In addition, you
[04:08] now develop the custom UI because in coming months, not even years, your developers are the ones who are not going to be very very hands-on. They just need to
[04:21] understand what it takes to build the building blocks of your data platform. So the custom UI creates that abstraction on top of that.
[04:34] these steps come together to create the data platform for you.
[04:50] So you will see how in this video the various components that I just talked about the overall architecture view the seven steps the eight different components comes together to create a developer productive tool productivity
[05:04] tool in order for a data platform to be accelerated and deployed and either to all the modern learn all the recent technology advancements that's now
[05:16] happening. U I always uh say to everyone all my clients that I talk to that if you go on a two weeks vacation you are probably outdated in data bricks. This
[05:28] is how fast it is changing. So in order for us to be able to keep pace with this framework helps to change on a very flexible path.
[05:42] on those seven modular steps that I talked about u that is coming together to look as a unified product and the solution you can see all of these steps while this is getting completed your developer
[05:56] doesn't need to be SQL expertise pispark expertise or any expertise in transformations what he needs to understand is a very conceptual working knowledge of what it takes to build the data product and while he had been
[06:09] building this he can go and review what is this product creating uh at a very detailed level while your data platform creation is happening at the back end.
[06:21] And he can go you can just see that he just type the sandbox where this is running and he is able to check what this is running in the in the back end and all of these on the left side of the
[06:33] screen you see what particular component is being built while this video is in progress. So at this particular instance you see that uh from a conceptual to logical to a
[06:46] physical data model is now being created and now this is being deployed as a SQL DDL in the production uh world. We won't see the deployment in this case because uh we just uh used an intermediate layer
[06:59] to keep this uh demo handy uh for this session. But when we do this deployment uh all of these uh datab bricks asset bundle can be equally created leveraging
[07:11] the agentics if there has been a lot of changes to dabs and our developer is not up to date with respect to it. So now you see that bronze to silver mapping
[07:23] has already been created. So your two hops is now actually in motion right. So what used to take you 3 months of a PC development could now be done probably in 3 days because you have tools like
[07:36] this. All you need is a is a reviewer who's has a conceptual and a working knowledge of how a data platform should work and in that conceptual and the working knowledge he is able to verify
[07:50] the various outcomes from each one of these agents that I talked about in the previous slide. So in this uh at this moment you're saying that how this how does the STM code generator converts the
[08:02] mapping into a pispark ETL modules and while this is being created you can see that there are steps where all of these things are reviewed uh with respect to
[08:15] what you have on the u on the mapping side done. So with the mapping in in in in place uh this uh helps uh ensure that you have
[08:28] uh yeah with this mapping completed now that this is ready for the deployment to your production stage. Right now our deployment strategy for this demo was
[08:40] only to deploy it until the silver because uh in in the scope of this 20 minutes we won't be able to show the deployment in the productions using the datab bricks asset bundle. So we've completed all the seven steps all the
[08:52] eight modules that come together to create your data platform modernization and have your first building blocks of the industry use case ready. So now that we have created this uh data platform,
[09:06] I'll go back to the industry use case and then I'll go back to the AI foundation layer that actually enabled this this uh business use case. And this this one is is for a healthcare industry where we automated the processes first
[09:20] by building the data platform then by the AI foundation layer which was leveraged to uh create the agentic solutions again 100% datab bricks native agentic solutions. The first step of
[09:33] that we have seen that we have created the data platform we converted that into an industry solutions various datab bricks services used. You can see this uh productivity gain of 45 to 55%. Why
[09:47] 55% is because we are still keeping human in the lead u to ensure that there is a right review and that the regulatory requirements of the particular industries are also added to
[09:59] because every industry has its own regulations. Even the codes get audited and healthcare industry is something that needs to have very high accuracy and we work with the business to ensure that this is an acceptable accuracy
[10:13] which is 96%. You see at the bottom uh left the datab bricks aentic AI foundation that will be the next next topic I will I I will be talking about after the demo that we
[10:26] have for smart coverage assistant. So you see that the challenges that we solved in the industry use case the challenge was that there are fragmented systems and data silos. The variability
[10:39] in human agent responses. Every time I call a customer care there are multiple responses that your customer care gives you. uh within 30 minutes of you calling you probably get two different details
[10:52] and you get frustrated which one to trust because healthcare is healthcare industry is as such where customer friction is very high. The reason is that you are needing some medical interventions. your your goal and
[11:06] objective is to focus on your health but you end up talking to a customer humaniz that gives you two different answers and you're now confused that which one is the right one and the eligibility if at all you have leveraged the North America
[11:18] healthare insurance eligibility is one of the most complex uh applications that you have to go through sometime the benefits that get sent to you you're you
[11:30] have to scratch your head to understand what is the explanation of the benefits and how does the various amount that gets calculated out of pocket deductibles how much the insurance pays and how much you have to pay this this
[11:43] uh solves that problem by automating document intelligence integrated data intelligence and then the life uh live call assistance so I'll go on and play the demo
[12:03] so we will see now that all of this The every time you needed to call to customer care to understand that what is your current eligibility running, which is your in network provider, which is your out of network provider, what is your out- of-pocket deductibles, what is
[12:16] your max and if how much of a co-ay co insurance that you have to pay when you visit a doctor. And this all you do um being on the call
[12:28] for 20 minutes while you you focus and objective and your primary concern is your own health or health of your near ones uh your dependents your your uh
[12:40] kids right what this does is that this gives you the tool at your hand to ask the questions in your plain English language or this can be enhanced to the other
[12:52] native languages that you speak to get the answers that you need. Every time you ask a question, it would be one answer. One answer that comes from all of the health plan documents that you
[13:04] have subscribed to, all of the medical interventions that you've gone through in your previous doctor visits. This takes care of everything. It scans hundreds of documents, hundreds of
[13:16] tables, brings you the unified view and you don't need to talk to a customer agent. for a customer human aent to interpret your questions and maybe give you two answers. Two times you ask the
[13:30] same customer agent for what is your eligibility when you are needing this uh medical interventions. You can see in a very simple uh phrase, very uh simple
[13:45] English. It doesn't have to be grammatically correct. As long as you put your keywords here that someone is needing help with the sleep, you get all the answers that you need and you go and
[13:57] and read through it. While we have created this just for the English text for this demo, we also have this in in various other languages. um that you can opt from. Um
[14:10] now that you've you've gone through what were the coverage or what were the interventions needed with respect to your sleep cycle. You had a very small incident of your baby u broking uh
[14:23] breaking uh the bone while he or she was at play. And uh the first thing that you have to do is to rush to the baby, call the ambulance and while you are doing this you can go to the app and just ask
[14:36] the very simple process that all right tell me what are the various uh eligibility that I am entitled to and how that will uh uh span out when I go
[14:49] to the doctor and while you're doing this you can quickly ask that which is your in network and out of network provider you don't need to call a human agent and be on the call for 20 25 minutes and all of this was done with
[15:02] 96% accuracy which is very high the requirement is of high frequency in healthcare industry. Um and this all you can get it in just 5 minutes or less while you are also
[15:16] concerned about your kids uh health. Now uh what was done how this smart coverage assistant came in? You had one platform for data and AI and governance 100% data native. I did not have to pay
[15:30] any complexity tax that I say when you include too many text tags in the process. Your built-in rag infrastructure I did not go and search for any outside of the vector databases. I I use the what was coming in as native
[15:43] data bricks elastic enterprise scale serving u I didn't have to create text tag uh maybe data bricks and AI foundry for example or sales maker or bedrock I
[15:55] could do all uh just the data bricks native and then I I could over the time evaluate observe and keep enhancing for a better and faster uh production results.
[16:08] So what enabled this? You saw the data platform. You saw the industry use case. Now we'll see that AI foundation that was created on top of it.
[16:20] So these are the four key challenges. Unstructured experimentation in healthcare industry because your data sets and the use cases are diverse. Your siloed deployment while you are doing your production deployments. Each team,
[16:32] each line of businesses are working with their brain of their own. high development and usage cost because each one of these developments incurs some cost and is running the CPUs uh the DBOs for you that otherwise is very hard to
[16:46] monitor and then lack of centralized control in a very highly regulated industry like healthcare. So how we supported this? So we created this AI foundation that I'm talking about now. We created a centralized AI
[17:00] platform 100% datab bricks native. We standardized it, governed it using Unity and then we roll rolled it out at scale for vast number of healthcare customers for
[17:14] our client. So what were the five building blocks? So we first had this AI foundation setup that would create your state u uh
[17:27] state-of-the-art architecture your repository your L&M ops your deployment monitoring observability and refinement you're ensuring that you're having the right AI and data management and
[17:39] everything is combined and integrated with responsible AI whatever comes in and whatever goes out is governed monitored and ensured that there is no incident with respect to any regulations. These are the other
[17:53] building blocks that we talked about the responsibility across the verticals application enablement prompt library and the rag. Right? So, so we saw just now that how a industry use case was
[18:07] enabled and gave you productivity from being 20 minutes on a call when you need and your focus is on medical intervention to a 5 minutes very plain simple native language supported uh AI
[18:19] agentic solutions and building that was your data platform modernization and the AI foundation that we created. Thank you. That was my time.

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.