AI for BI without the BS

Summary

• Self-service reliability: Omni on Databricks enables users to answer their own data questions with AI help, grounded in semantic layers and business context.
• Semantic layer foundation: The semantic layer provides shared definitions and business logic that gives LLMs the context they need to deliver accurate, reliable answers.
• Production-grade AI: Omni focuses on building reliable, accurate AI products for internal users and customers rather than chasing generic AI capabilities.

AI for BI without the BS

Talk by James Davidson, Co-Founder & President, Omni
Stuck on a treadmill of endless report building requests? Wondering how you can ship reliable AI products to internal users and even customers? Omni is a BI and embedded analytics platform on Databricks that lets users answer their own data questions – sometimes with a little AI help. No magic, no miracles – just smart tooling that cuts through the noise and leverages well-known concepts (semantic layer, anyone?) to improve accuracy and delight users. This talk is your blueprint for getting reliable AI use cases into production and reaching the promised land of contagious self-service.
Intelligent analytics for everyone: https://www.databricks.com/product/business-intelligence
BI meets AI: https://www.databricks.com/resources/ebook/business-intelligence-meets-ai
Explore the only unified platform for agent systems: https://www.databricks.com/product/artificial-intelligence
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements

FAQs

Why does Omni emphasize 'without the BS' for AI and BI?

While LLMs have become incredibly smart, they lack business context needed to answer questions accurately. Omni focuses on grounding AI in semantic layers and business rules rather than relying on general AI capabilities, delivering reliable answers to real business questions.

How does a semantic layer improve AI accuracy in BI?

A semantic layer defines shared business definitions and logic that gives AI context about your data. For example, "users" could mean signups, active product users, or purchasers depending on business needs. A strong semantic layer helps AI select the right definition and deliver accurate insights.

What is a semantic layer?

A semantic layer is a business logic abstraction between raw data and analytics tools. It defines business metrics, entities, and relationships consistently, enabling both humans and AI to understand data correctly and drill into metrics reliably.

What is a semantic layer in business intelligence?

In BI contexts, a semantic layer provides standardized definitions of key metrics, dimensions, and entities. This ensures all dashboards, reports, and AI systems use consistent business logic, preventing confusion about metric calculations and improving data trustworthiness.

How does semantic layer governance improve AI and BI?

A governed semantic layer ensures metrics remain consistent as business rules evolve. AI can reference authoritative metric definitions, and business users see consistent metric values across all tools—critical for trustworthy analytics.

Full transcript

Uh, thanks thanks everybody for coming. Uh, super excited to chat a little bit about AI for BI and give you a little bit of a self-admitted uh, a little bit of a skeptic uh, uh, and sort of my view. I'm one of the founders at Omni. Uh, we've got a booth just over that way. Look for the pink if you want to continue the conversation. Let's see. Oh, here we go. Uh, so if you're like me, you've probably had a bunch of conversations at the conference already about the potential for LLMs and generative AI to transform data.
We are super excited about the use case and I think in in general, I've been working for data for a long time. I've been working at uh at BI companies. I've been a data practitioner. Uh I think everyone's always hoped for the potential to actually enable self-service and never have been able to realize that you know you can get at most sort of a quarter of the organization according to like Gartner stats using data and maybe text to data actually has the ability to lower the frict lower the friction sort of lower the bar and actually bring everybody in to actually use data.
There's a ton of prem promise for uh you know generative AI. We're really excited about it. But while LLMs have gotten incredibly smart and the sort of uh their abilities like honestly, you know, continually change on a week- toeek basis, they don't have context about your business. If you ask a question like how many users do we have, it effectively implies a bunch of context. It has a bunch of different reasonable interpretations.
It could mean how many people have signed up for the service. It could mean how many people have used the product and in what time frame? How many people have made a purchase? What does a purchase mean? Like these things are all nuanced questions. Does a purchase include fraudulent things? Does it include things have been returned? Does it include test accounts? Like they're they're difficult definitions to make and certainly can't be just inferred uh without without any context by an LLM.
See if I Oh, this has got a little lag. Sorry. you know, our our fundamental belief is that AI alone cannot uh realize the potential of uh of you know, enabling self-service and solving all of the hard parts about BI, but we think that there's ways to do it and we've got a bunch of customers and a bunch of use cases we can talk about um just in a little bit too. A few a few principles that we sort of have developed. You know, one the data quality matters.
Uh and this is something that may maybe sounds obvious to a bunch of data folks, but uh you know, if you're making decisions based on data, it needs to be accurate. You need to have consistency. You need to have a mechanism to trust it. You need to have humans to be able to actually understand it. You also need uh ability to provide guard rails for LM. LM are inherently non-deterministic systems. You have to start from the premise that they're not going to be 100% accurate.
It's not you. Natural language is inherently imprecise. How many users do we have does not contain enough context to actually tell you what a definitive answer is. There's no right right answer. There's lots of kind of contextual answers that you can give. All right. So, accepting that LLMs aren't going to be perfect, they're going to be nondeterministic, you need to design for a human in the loop system where the system can actually become, you know, better with usage, you can refine it over time.
And then finally, you we think the mechanism to do this is effectively a shared vocabulary to enable humans and AI to use consistent definitions. And that is the wrong direction. I must be holding it backwards. Uh, and we think the way to do this is a semantic layer. So, a semantic layer at our at our core. And this is something that I think the industry is starting to do. We're excited to be launch partners with data bricks and their new metrics view where you effectively can encode some semantics into your data bricks instance.
But semantic layer just to sort of really quickly define it. It's a place where data teams can define data relationships and business logic. It's an abstraction layer that enables folks in in an organization to use a single definition of complicated metrics like gross margin and use it everywhere consistently too and then have change management uh on top of it. So if you update that definition of gross margin, you do it in one place.
You don't have to go into a workbook across uh you know every single workbook to go and update these things. Just to show you what what like a kind of articulation of it uh is uh you can think of a semantic view or a semantic layer as an interface to data. It creates a basically a shared vocabulary with basically business context and semantics and data relationships being defined explicitly to and it gives the LLMs effectively a a shared space to operate within too.
You can strain the input space and output space and dramatically increase the accuracy of it. It also makes the answers that an LLM generates interpretable to the non-technical user. So we've got lots of executives, we've got lots and lots of uh you know operations folks, customer success folks, sales folks using data through chat and they're able to interpret the data because it's not a single number which is completely uninterpretable and is you know highly potent.
It can be can be inaccurate. It also is not uh it's not a tech it's not a SQL statement too. That executive can't read the SQL statement and understand the gross margins correct. that salesperson doesn't know what your definition of your pipeline is or the workflow that you maybe use to define you know arro say for example the semantic layer effectively creates that shared vocabulary the way you build a semantic layer we think it is effectively iterative you want to start start with the sort of core primitives of uh of the schema for a g a given data warehouse and then enumerate transformational logic in analysis you want to build up uh you know that semantic layer um iteratively so as you're answering questions as you're building up you know analysis as you're building up dashboards, you're effectively creating reusable pieces of logic.
Not all of that is going to be shared, but but there needs to be basically a workflow to go and take that and sort of standardize it directly. So, you know, the way way we think that this sort of human in the loop uh workflow works, I'll just fast forward. Users can ask questions in natural language and get results. Those results are interpretable because we're showing them the fields and the context for which they're getting that data.
Users can actually iterate on those those queries. They can use that as a starting place, but very very rarely, I'm sure all the folks that sort of are data practitioners have been doing it for a while, very rarely is where you start necessarily the exact answer. It probably actually spawns four or five more questions like I'm looking for, you know, a trend for this KPI over time. Why is it changing? What are the next, you know, what are the next degrees of freedom I want to go and ask questions of to be able to go better understand it?
And AI can even help with that, summarizing it and basically recommending next steps, recommending things that can kind of remove friction from that process. Excuse me. Uh we we think that you know AI has the potential to uh remove friction in a variety of contexts. We've seen a bunch of success where folks are building out customerf facing analytics. So using their data to their customers to show usage or or demonstrate value for folks too.
Those are places where uh inherently because you're building for your customers, they don't have an internal team to go and rely on to go ask questions of. They are forced to effectively product manage the data assets. Forced to product manage the and curate the experience so that the user can interpret what date to use or what metrics mean um how it ties back into you know the workflows. Those are great use cases for LM because you've effectively built a data model in in doing that too and you need a data model to go in um and adopt it.
We think that it's great for business users to ask questions sort of for the whole premise of self-service sort of predicated on how do you make it as easy as possible for people to get trusted data. Uh and then finally and I'm sure this is going to appeal to this this audience in general. We think that it's going to help data teams too. It's really good LLM are really good at unstructured data and folding in workflows to be able to go and derive insight and value from structured data is super interesting.
So we'll talk a little bit about that. Um you talk first about customer experience. We think it it can dramatically increase sort of the degrees of freedom you can give your customers and ask questions. So no longer are you just generating a PDF and sending it to them uh you know on an ad hoc basis. You can basically open up the sort of the full realm of possibility of uh you know the customer's data too and can be done in a way where it's super intu interpretable to an end user.
Uh we also think that uh it it means you're lowering the friction so that everyone everyone can consume it. You don't have to be numerate. You don't have to write SQL. So you don't have to even know how how to use a spreadsheet to necessarily be able to go and use natural language. Uh we think that these things need to be uh integrated into a variety of experiences. So there's rich APIs and you know an MCP server to integrated with your existing chat interfaces so that when you're looking for organizational data, you're also able to intersperse sort of structured data too with sort of context broadly.
And then you know finally too it's incredibly important to uh enable sort of a feedback loop. They can ask subsequent questions. They can start from a dashboard. They can go into a chat interface. They can use that chat interface in an iterative way so that it can actually answer questions too. Really basically matching the sort of uh classic analytical workflow to talk to talk about business users. Business users uh this we think this is has the potential you know there's probably a billion people or so that can use Excel.
Everyone speaks natural language. We think this has the potential to dramatically increase sort of the uh sort of aperture of what data can be applied to. We we can also uh leverage some of data bricks new functionality. So in Omni Omni the name is hey we want to bring data to everyone in the organization. That's sort of the the genesis for the company. We want to have an interface for everybody. So for the most technical folks rich IDEs, data modeling, SQL.
For the less technical folks it's point-and-click interfaces, it's Excel. It's it's natural language. you can leverage the sort of even the more technical things the things like you know the AI classification you know via SQL inside Omni and create standardized reusable metrics so that you can actually do things um in an iterative fashion you can do things where they're parameterized um and we think that this is actually going to start to uh enable folks in their everyday workflow because we use direct onetoone parody with Excel uh we have an infinite corpus of content to train on too LLMs are incredibly good at writing things like Excel functions so you We basically we can give people accelerators and basically leverage sort of powerful things that sort of the underlying infrastructure gives you also give you accelerators in the in the front-end interface.
And then you know uh finally too we've talked about this a little bit already but like enabling that sort of natural language in a in a semantic context means that end users can ask questions they can interpret results. they can be, you know, uh uh guaranteed that they're referencing uh revenue or customer requests and the common definition of it too and have it be integrated into the full analytical experience. We've seen customers create a huge amount of value for this.
We got one customer that has uh you know a reasonably non-technical sales team and the sales team was effectively prospecting into uh a bunch of uh you know startups that were you know seeking to go public. I'm sure they'd love to sell the data bricks themselves. All right, they they wanted to go ask a bunch of questions of a bunch of structured IPO data and sort of this is kind of you know cut think of this as enrichment data for their CRM and they were able to enable their their sales folks to actually ask questions their data and reduce the sort of the number of tickets coming into the data teams by a huge order of magnitude too and dramatically increase sort of the feedback loop and enables people to ask more questions of their data and get more value.
So sort of last last we can talk a little bit about data teams. Super excited about removing some of the manual effort from uh the data engineering workflow. You know at our core Omni is all about building semantic models. We want to make it super easy to have that workflow go from lightweight ad hoc insight new metric new piece of metadata and have it progressively sit wherever it makes most sense. It can live in our semantic layer can live in data bricks semantic layer.
It can be materialized inside your data warehouse too. We've got a tight integration with data engineering workflows. We'll even go so far as to do things like generate a git pull request in your DBT project with a rich birectional integration to enable you to take the that logic and put it wherever it makes most sense. If you want to go for cost performance or standardization reasons, materialize it. That makes a ton of sense.
We want to facilitate it and we'll go so far as to to do it for you. Something that I think uh data people are typically very focused on and is incredibly important is the governance associated with their data too. The sort of the last but like not least uh benefit of having text to data ultimately pass through a semantic model is you can actually apply rowle column level permissions because you're marrying the stochastic system that the generative AI does which is great for asking new questions.
It's great for summarizing data. It's not so good when you want to have strict adherence to you know rowle security where you don't want to hallucinate joins. don't want to hallucinate new uh new potential definitions for gross margin and have it be misinterpreted by an end user. In order to basically have that sort of tight tie to a deterministic system, we basically generate semantic queries. It means that every single chat uh chat generated piece of data is enforced at a row and column level, which is great.
We want to give you very granular controls over this and think it's incredibly important. So, so to reiterate, you know, we think there's a couple of really important principles that we we've got for AI highest level, you first have to accept that generative AI is not going to be 100% accurate. I think probably everyone who's played with it will will accept this and see uh you know, it's hallucination prone. It has huge potential though for removing friction from the analytical workflows.
The way you you can incorporate it though is by basically giving it context and structure. You want to constrain the degrees of freedom that the AI has and effectively design for a human in the loop so that the human can interpret it, the human can iterate on it to build a mental model for what the AI is actually helping it with or helping the humans with and uh and sort of fold it in to create leverage for the uh the organization.
Just to close at a very high level, we got a a quick 2 by two I sort of talked to with uh you know customers thinking about implementing you know generative AI for uh for data projects. You know on on one axis you've got the difficulty of a question on the other axis you've got level of effort. So it either can be you know easy or hard or you know difficult uh and or not not difficult low low you know low low difficulty.
The way way I sort of frame it is uh data teams should be strategic thought partners when the questions are hard when they're not pre-modled when they're when they're things that where you're you don't know exactly the space that you're going into. When the questions are easy and loweffort it uh excuse me uh uh it's a great spot to start with generative AI. So things like lookups, things like what uh what did the account team on this this deal?
What was the performance for this specific campaign? what was you know uh uh what does our pipeline look like? These are the simple types of questions that people can actually uh you know almost immediately out of the box get on top of a semantic model. The questions that are slightly more uh slightly higher effort are also still possible. They effectively are possible though with data modeling and you want to do that in sort of an iterative workflow where you say like things like uh as an example a customer showed me uh a use case where they were they were doing a cohort analysis and they wanted to see the performance of a specific cohort.
That's a sophisticated analysis to typically go do, but they had effectively pre-modled it inside inside of the Omni experience and they were able to ask questions of it in an incredibly intuitive way using natural language because that context was already pre-built in the data model. So, I sort of say you want to start start with the easy stuff. You want to work your way up in in complexity too by leveraging the context that you can build uh in in into the the data model.
Um, so that's that's the you know that's the sort of core focus you know firmly believe in the value of generative AI. You need to basically encode the context of your business into the semantic model and we think we've got the best way to go and do that. So ju just uh you know if you're interested in chatting I'll I'll stick around for a little bit too. We've also got a booth right over there. Look for the pink uh come come talk to us.
Thank you.

Learn more about the Databricks Data and AI platform.