Inside Nav-AI: Accenture's No-Code Agent Factory, Powered by Databricks Apps and Lakebase

Summary

Accenture's Nav-AI is a low-code/no-code agent factory rebuilt on Databricks that lets business users build and deploy AI agents through a visual interface, using Databricks Apps for one-click multi-cloud deployment, Model Serving for the agent runtime, and Lakebase for persistent state and self-improvement.
Nav-AI uses a requirement-to-dataset-to-model-matching pipeline to construct bespoke state-of-the-art models for specific problem sets, demonstrated through a global grid AI digital twin that forecasts electricity demand growth driven by AI consumption.
Unity Catalog governance spans the entire Nav-AI platform, while Databricks Genie is used both to test model health—querying metrics such as MAPE comparisons and AIC weights—and to expand and build on foundational models within the application.

Inside Nav-AI: Accenture's No-Code Agent Factory, Powered by Databricks Apps and Lakebase

Accenture's Nav-AI lets business users build and deploy AI agents through a visual interface — no deep technical expertise required. Every customer deployment used to be a bespoke engineering project: weeks of custom work per cloud, fragile ingestion, and no foundation for agents to improve over time.
In this video, Accenture shows how Nav-AI was rebuilt on Databricks to become a true product — using Databricks Apps for one-click multi-cloud deployment, the medallion architecture for reliable customer data ingestion, Model Serving for the agent runtime, and Lakebase for persistent state and self-improvement. Expect live demos from real deployments and a concrete blueprint for turning agents from a science project into a repeatable platform.
Talk By: Amir Bagherpour, Global MD- Databricks Business Group CTO, Accenture ;

Chapters

FAQs

What is Accenture's Nav-AI platform?

Nav-AI is Accenture's low-code/no-code agent factory that lets business users build and deploy AI agents through a visual interface without deep technical expertise. It was rebuilt on Databricks to become a repeatable product, using Databricks Apps for deployment, the medallion architecture for data ingestion, Model Serving for the agent runtime, and Lakebase for persistent state and self-improvement.

How does Nav-AI's model-matching pipeline work?

Nav-AI converts a business requirement into bespoke datasets and then applies model matching to construct the ideal model for that dataset and context. The result is a purpose-built, state-of-the-art model for the specific problem—demonstrated in this video by a global grid AI digital twin built to forecast electricity demand driven by AI consumption growth.

How does Accenture use Genie within the Nav-AI platform?

Genie is used in Nav-AI both to test and monitor the health of deployed models—enabling natural language queries against metrics like MAPE comparisons and AIC weights—and to expand and build on foundational models already within the application, creating a conversational layer over model performance monitoring.

What are the cross-industry applications of Nav-AI on Databricks?

While the demo focuses on electricity demand forecasting for the global grid, Nav-AI's architecture is designed for cross-industry deployment. The Databricks Apps one-click deployment model means the same agent platform can be deployed across multiple clouds without bespoke engineering per customer or industry vertical.

Full transcript

[00:07] Okay, good afternoon everybody. My name is Amir Bagherpour. I am the head of our ADI Accenture Decision Intelligence practice and our Databricks business group
[00:19] CTO for Accenture. It's a pleasure to present today a demonstration of our demonstration of a global grid AI digital twin using
[00:32] NavAI, our low-code no-code agent factory that is built on Databricks. Now, why Databricks? It's obvious. Databricks is the application that allows us to construct
[00:47] advanced scalable applications in a way that we weren't able to before using a combination of apps one click away. We can take our low-code no-code application or any type of application that we're using. In this
[01:00] case, we're showing you our NavAI application and then grounding it with the Databricks features using Delta tables, lake base, Unity Catalog, and the Databricks apps function. Now, what you
[01:14] see here is, as I mentioned earlier, the global grid AI digital twin. And what we did is through our our workflow whereby we are taking the requirement and then
[01:26] converting that particular requirement into bespoke datasets and then converting those bespoke datasets into what we call model matching whereby we are constructing the ideal model for that dataset and the context to
[01:40] construct bespoke state-of-the-art model for that particular problem set. In this example, it's the global grid for electricity pertaining to AI consumption.
[01:52] Why AI consumption? Because we're all seeing it in the news. It's It's resulting in increased demand for energy and we want to know what's that demand going to look like 4 5 6 years down the
[02:05] line and how do we manage it. What you see here right off the bat is the underlying application in our model. You could see we built this in Databricks and then we're using Genie not only to test the model, but we are
[02:18] using Genie to expand and build on the foundational models that we've constructed within the application. And then we also have the dashboard feature whereby we are testing and
[02:31] monitoring the health of the models, whether it's back testing on what's called the MAPE comparisons or AIC weights and then using the native Genie functions to ask questions pertaining to
[02:44] the health of those particular models. Now, this is the foundation that this digital twin is built built on. You could see that on the front end we have
[02:56] live API uh that tracks consumption, demand regarding that consumption, what the energy transition looks like including forecasting energy demand going into the future and
[03:09] you could see we've built scenarios whereby we could look at the status quo trajectory versus what AI acceleration looks like. Uh so let's look at for instance AI acceleration and we could go down and click on AI
[03:23] acceleration and you could see that we will can now run the forecast off of AI acceleration where we hit run the models. We could even conduct back testing and it shows
[03:36] under what scenario um and what the best model is for consumption going into the future. Now, we also have as part of this digital twin the GIS layers
[03:48] um and then in that GIS layer we can um put overlays on what the AI data centers look like, grid stress, regional PSI, nuclear SMR, you name it. And then the
[04:00] last part of the digital twin is the policy and understanding what the different characteristics of policies affecting electricity consumption are going into
[04:12] the future. Now, here we're using one of the key models in our NavAI suite of tools called Schelling, in which we capture the dispositions of the stakeholders and can then simulate
[04:25] what the future looks like in terms of the direction of policy making pertaining to electricity grid. But this story is not necessarily about any particular model pertaining to the electricity grid.
[04:37] This story and this demonstration is really about the power of Databricks empowering our low-code no-code application platform called NavAI.
[04:49] Now, what NavAI is our enterprise solution for low-code no-code enablement to our C-suite customers. And you can see that when we construct
[05:02] this this this is the front end. It's a low-code no-code solution whereby we can construct agents and then have those agents consumed by our executive customers. Now, here you could see the three types of agents that
[05:15] we have built. We have a geopolitical governance agent, an energy demand forecasting agent, and then the orchestrator. All of this is built on NavAI NavAI applic- sorry,
[05:28] Databricks applications. And what we do here is we can construct the agent using the low no-code low-code interface built on um
[05:41] our open-source llama index. And then you could see here we have the function tools, boom, we hit Nave AI, MCP, update the function tools, and then now our application is running
[05:55] on top of a Databricks in a way that's scalable with the ability to not only build our agent factory using Nave AI, but to actually build grounded uh models with enterprise scalable data
[06:09] on Databricks. So, for instance, I could go into the agent, and then you could see as an example here, um
[06:23] I ask it to run a full inventory on energy forecast and models available on the Databricks platform, and you could see that Nave AI does so by producing the historicals and then retrieving the data
[06:35] and then showing the models that are available, and then we go to work by running those particular models on any set of data that's relevant to the digital twins.
[06:48] The other feature that we have that's inspired by um na- uh by by Databricks is our ability to be able to track tokenization. Obviously, um
[07:00] we we can we can track uh uh uh tokenization, but this isn't just about tracking token consumption. This is about tracking um the type of work that we're doing, that we're delivering,
[07:12] that goes beyond just what our base token consumption is. So, for instance, if I'm using my energy demand uh uh agent, I could track the cost on
[07:24] this. And you you right now doesn't show the exact uh the num- uh uh the numbers on the cost, but it can
[07:36] essentially run the cost off of data sets, function tools, visualization tools, and what the cost is in regards to the type of output that is being constructed. So, for example, if I am
[07:51] delivering an output that is, let's say, for a C-suite that is pertaining to a strategy of a company, that's a different type of token output versus a high consumption automatable task that's
[08:04] on the back end. Anyway, um this is a demonstration of our Navi AI low-code no-code agent factory built on Databricks
[08:16] um using Databricks apps lake-based native features. Thank you.

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.