Sponsored by: Omni | AI analytics at Condé Nast: What it took to get to production

Summary

Condé Nast manages analytics for 14 global brands and markets and partnered with Omni and Databricks to encode business logic, metric definitions, and synonyms into a governed semantic layer that AI can reliably reason over.
After a six-month implementation, the team automated a five-step manual campaign pacing workflow using an AI orchestration framework built on Omni and Databricks, enabling AI-generated campaign takeaways and performance driver recommendations to be delivered at scale.
This video argues that AI supplements rather than replaces dashboards—teams still need good data foundations and semantic context before AI can reliably ask questions, synthesize insights, and surface actionable recommendations from enterprise analytics data.

Chapters

FAQs

What is a semantic layer and why is it required for AI analytics?

A semantic layer encodes the business logic, KPI definitions, synonyms, and decision-making rules that give raw data business meaning—telling an AI model what 'performance' means for a specific team, how to interpret a given metric, and which filters apply in a given context. This video explains that without a governed semantic layer, AI-generated analytics answers are unreliable because the model lacks the context to interpret numbers correctly.

What manual workflow did Condé Nast automate with Omni and Databricks?

Condé Nast automated a five-step manual campaign pacing workflow that previously required analysts to manually pull data, interpret performance signals, and generate recommendations for campaign adjustments. After encoding business logic into the Omni semantic layer on Databricks, AI now generates pacing takeaways and surfaces performance drivers automatically across all 14 global brands.

Does AI replace dashboards in modern analytics?

This video argues that AI supplements dashboards rather than replacing them—dashboards remain essential for monitoring known metrics, while AI adds the ability to ask questions, synthesize patterns across datasets, and surface recommendations that analysts would not have thought to look for. The prerequisite for both to work together is a strong data foundation and a governed semantic context layer.

How is AI changing the role of data teams according to this video?

This video describes a 'StackOverflow effect' where tasks that previously required significant technical expertise—like writing complex queries, building predictive models, or performing reverse geolookups—are becoming trivial API calls. Data teams are shifting from technical gatekeepers to context builders whose primary value is maintaining the semantic layer and data foundations that make AI-generated analytics trustworthy.

Full transcript

[00:09] So, we're going to talk to you about uh AI for data and specifically Kai is going to come and talk a little bit about how they put it into production. But first, uh we're going to talk about how AI is changing data analytics um and
[00:21] how you need to build for it. Uh so, quick background. I'm Colin. I'm the CEO at Omni. Uh we have about 800 customers, all sorts of industries, shapes and sizes. Um from tech natives
[00:33] to old school companies that have been around for 50 100 years. Um we've been in the space for a very long time even though the company is young. We're about four and a half years old, but um a lot of us came from Looker, Fiverr, DBT. Um
[00:48] have been in the space for a very long time. We've raised about $220 million. Um, and the sort of topic sentence for this, and again, sorry about the tech, is the data world is changing faster
[01:00] than we think. And I probably don't need to pitch this all to you, but I'm just going to show you a few slides that help put it in perspective. So, when this comes up, does everyone know what this is a picture of? Does anyone know?
[01:13] I heard someone say Stack Overflow. This is a a chart of Stack Overflow questions since its founding in, I guess, 2009. until today. And I think that people
[01:25] understand that AI is changing things, but I'm not sure that people fully appreciated sort of the order of magnitude with which it is occurring. Um, and sort of how drastically it's changing things. And the question is sort of why.
[01:37] And uh, this is an XKCD cartoon from I don't know six or seven years ago that I think is almost the perfect example of how things have been different. and you
[01:49] can read it through, but the idea is that it's really easy to do a reverse geolookup, but if you have a picture of a bird, you're going to need a research team and five years to actually go figure out that it's a bird. And very
[02:02] literally in the past, I don't know, nine months, that has turned from a research team of five to go figure out that you've got a picture of a bird to effectively an API call to one of the foundation models. And it just means
[02:16] that these things that were sort of functionally either very difficult or that required an enormous amount of technical resource are trivial at this point. And when it comes down to it, I
[02:30] think the same thing is happening and we are seeing the same thing happen for data analytics. Um, and I'm going to show you a bunch of different examples of sort of real life examples of where this has actually changed things. So the first one and this is probably the simplest. Uh, you can just ask for
[02:44] things So, if I can actually get these to play, we will see. Uh, on the left hand side, we've got a video of Oh, man. So, it's it's going to be a GIF of someone just asking for what they want out of a BI tool. On the right, it's
[02:57] sort of like the general navigation experience of knowing the eight clicks that you need to do to actually go find a data set that you need. And I know that that sounds incredibly, you know, oversimplified, but this idea that you
[03:09] can actually just show up to the tool and ask for what you want means that you can remove all of the context and understanding that you need to actually be effective with data. There's a lot of people in organizations that have questions and the barrier for them
[03:21] answering the question isn't the data set's available or the data is well known. It's that the person literally doesn't know where to go to actually ask the question. And sorry, I have no idea what's going on
[03:33] here. We'll keep going. Um, the second one is that AI can sort of fundamentally answer things very differently than uh, you know, a human can. And the reason I
[03:46] like this picture, this is a result from an analytics query through the natural language experience. And you can't really see it now, but over the left is a scroll bar. And this is actually a result set that is, I don't know, 12
[03:59] screens wide. Classically, that is uh a result set or an answer that is pretty hard for a human to deal with. If I wanted to ask about these three deals, um the idea of just getting that data
[04:12] back in a table that is 12 screens wide is actually pretty ineffective for me. In the alternative, if I go up to a natural language experience, the idea that you can just show up to a BI tool
[04:24] and ask and it can actually synthesize what telly everything about an account actually means. This is orders of magnitude simpler for a human to go understand. Um, and the idea is that you
[04:37] know the AI can synthesize here are the data sets that I have available. When you say give me everything, I can actually go figure out what everything actually means and the user doesn't need to understand this sort of thing. So again, this is like a real life example.
[04:49] It can read in that giant result set and actually synthesize out the things that are helpful for each person. And this is ultimately what people are trying to do with these tools is make a decision to do synthesis on top of these result sets. And AI takes this from like a
[05:02] two-step process of find what you need and then synthesize it all the way to what the user is actually trying to do, which is actually synthesizing it. Um, the second one is that it can it can
[05:14] go do the next thing for us. So again, like this is what the user is actually trying to do in a BI context. They're not trying to usually pull data. They're trying to say like how are we doing and what can we actually do to improve? And
[05:28] this is where these AIs can now take the next step on behalf of the user and say great like this is the data that I have. Here are the wins and the losses. And again, like this is a real result set
[05:40] where it's pulling and it's synthesizing and it can actually give things like tactical improvement opportunities and it just means that we can actually productize out the use cases for our end users. One of my favorite examples that
[05:53] we're not going to talk about today is we had a support leader at a customer actually go through every single member of their team and have the AI pull out transcripts of a conversation and
[06:06] specific types of failures that a support rep could make with quotes that they could then action them on how to improve down the line. And it really starts creating this sort of blurriness
[06:18] between doing data analytics, so line charts essentially, and doing business operations management. And these are the things that can actually sort of be productized. So again, like execution
[06:30] insights. The the next part that is really important here though is that we can actually see what our users are doing. And again, this doesn't sound like something that is uh like
[06:43] transformationally interesting, but for folks that have been in the analytics space for a long time, we are often given tasks by our end users like, hey, go build me this dashboard or, you know, I'm trying to understand XYZ. I need to
[06:57] understand every type of data about that sort of object. That's still only a portion of what our user base is actually trying to do. like most of the usage of a data environment
[07:09] is sort of unwatched by the builders. And the interesting thing with AI and sort of natural language is we actually get a transcribed log of what our users
[07:22] are trying to do in the data environment. And these are very literally real life examples from our this is actually from our sales team. And what it means is that we can
[07:34] actually now reverse engineer the types of data and the questions that our organization needs and we can start sort of reorienting the way that we deliver things out to think about very directly
[07:47] the types of questions that our customers are trying to answer. And you can sort of see the difference between a sales team and a support team. But the idea is like these these questions can be really different and it means that we can sort of action the improvement that
[07:59] we need to take in the environment to go answer questions for the business. Um now the next thing that we have seen is that for all of the magic that I just explained we are not seeing natural
[08:12] language eat everything about doing data analytics. So this is again like a real chart. It's a it's a month old now of usage in our platform and you can see that AI is growing faster than everything. That is not surprising. But
[08:25] what we also see is that it's driving more consumption of dashboards, more consumption of spreadsheets. The other things that exist in sort of a classical BI environment don't go away. We still need people that are delivering pixel
[08:37] perfect dashboards and spreadsheets on a daily basis. The AI is just sort of supplementing what people can do. And so when we think about sort of the landscape that is required to do data
[08:50] effectively, uh, conversational AI has become sort of the centerpiece of all of this. Like I will say personally that probably 90% of my data usage has moved from technical things to now just asking questions. But we think that all of
[09:04] these different products should exist in one platform to be effective. Like you still need to go share data externally. Technical users still need to write SQL. You're still going to want point-and-click interfaces to sort of refine things. All of these ultimately
[09:18] have a place in the sort of BI landscape. And I'm going to sort of finish here before I bring Max on talking about sort of how you get this done. And what we have seen is that you really need two
[09:30] things to do this well. First, you need to do all the things that we've been doing for the last 25 years. Like you need to get your data in good shape. You want to name it. Like you want to have good pipelines. You need reliability. There's lots of vendors downstairs that
[09:42] can help with all of those different things. But the sort of new thing that we need is we actually need to sort of more tightly explain language in the context of data. And this is probably one of my favorite examples internally. Like we learned this the hard way by
[09:56] just doing it. We we are a Salesforce shop and we had a lot of people come in and say like how much business have we closed this quarter? For people that don't use Salesforce, you have opportunities. You close them one or you close them lost. If you close them one,
[10:08] they're a customer. If you close them lost, they're not a customer. The AI would sort of happily return a mix of closed one and close loss deals when we say how much business have we closed. So, uh, for people that are not in SAS,
[10:21] this is, uh, not the correct result set. It's going to like overstate everything. It's it's essentially not meeting the expectation of the user. And it's this mismatch between language. I said I wanted closed and there's a column called is closed and intent. The user
[10:34] when they said closed means one. They don't mean closed. And again like uh a thoughtful AI and um learning from history, you will get some of this sort of absorbed over time into a semantic
[10:47] layer or however you're doing data pipelines. But ultimately we see every single business has a lot of this nuance in their everyday existence. And they're things the data team often understands and maybe an end user doesn't understand. If you're an e-commerce
[11:00] shop, you might have, you know, returns and cancellations alongside completed transactions. you need to be able to explain this nuance at query time and effectively make your environment thoughtful in terms of answering these
[11:12] questions. Um so yeah this is again like the example of closed and one um the other thing that we can do with these feedback loops is I sort of mentioned this earlier is we can build around them. So this was another question that
[11:24] got asked show me my team's performance relative to other teams. We didn't have any HR data in our analytics environment, but because we had that feedback loop to understand what users were doing, we could say great, a lot of
[11:37] people are trying to do org chart centric things. Let's figure out a data pipeline to serve this up. And now we can sort of in an automated way answer these new types of questions because we can see what our users are trying to do.
[11:49] And so ultimately, it's that AI is this combination of a data problem and a language problem. Um so kind of to to wrap it all up, we got to describe how things work.
[12:02] Uh overfitting is actually a good thing. You know, like that closed example. Um if if we have statistitians in the audience or data scientists or whatever you call yourself. Um like overfitting
[12:14] was generally viewed as a bad thing and we still might learn in AI that like too much context becomes a problem over time and it probably does. But the nice thing is that when we know what our users are trying to answer, we can make sure that AI can actually return those responses
[12:27] successfully. And then we want to just constantly build on top of these things. So it is very much just extending the things that we do today with these new amazing tools. But because of the level of insight we have, we can really design
[12:40] an environment where every subsequent sort of experience is better than the last one. And you actually build this intelligence and share it out. So this is what it looks like technically. Um, but with that, I'm going to turn it over
[12:53] to Max to talk about what it looks like in reality. Sorry for the slides.
[13:07] See what happens. Cool. Hey everyone. Uh, I'm Max. I um I will tell you how we monetize our analytics stack with Omni and Data Bricks. Um, and also leveraged AI. And I will try to not comment about the slides after this. Um, so I'll give
[13:20] you like where we came from, how we evaluated the tools that we needed, um, what we built and what we're doing with AI right now. So when we started as a group, it was a
[13:33] team of one, which was just me. Um, we oversee all of our performance and measurement for Kane NAS advertising. Simply put, when someone sponsors the Met Gala, my team is the one telling you
[13:45] why they sponsored it, how they sponsored it, and what happened now that they did sponsor it. We tell how Kane NAS drives outcomes for everyone. Um, so we have a team within my larger team
[13:57] called performance and insights. They established like our global reporting guidelines, effectively our semantic layer, right? um for all of our users. We drive actionable campaign insights, tell people how things performed,
[14:09] click-through rate, good, video completion rate, great. And then the biggest thing that we do is we now democratize data and support across our entire revenue and advertising organization. Um
[14:21] so previously we were on a um different BI tool um one that left us looking for a lot more. So if you're familiar with this UI, that is Looker Studio. Um, when
[14:33] we started, we again we just were a team of one. So, we had to prove out why this actually worked. What were we going to do to do self-service analytics for all of marketing and sales across the entire organization? Um, that was tough, but
[14:47] this did actually give us a ton of opportunity because we knew that self-service analytics was actually something people needed and wanted. Um, so we had a heavy reliance on all of our data teams. We would just get bombarded with questions all of the time. uh
[15:00] everyone would ask us what's this mean why does this look this way how does this work operationally so heavy reliance on us as the data experts um and that left us looking for a lot more
[15:13] when we did evaluate our criteria for BI tool we needed something that was innovative right so my larger team is called global enterprise innovation so we kind of like had to innovate being under that team umbrella um we also
[15:25] needed something that had a lot of value for us um and we needed to be modern flexible UI and then as Kai NAS since we are home to like Vogue and Vanity Fair and GQ, we needed it to look really
[15:38] pretty. We needed to look great and this allowed us to have customizable embedded analytics through um a third party that helped us out. Most importantly though, we needed things to be fast. We needed things to be speed um speedy and then we
[15:51] also needed to ensure that we had things that service technical and non-technical users. We had some people that were very tactical and then some people that just needed point-and-click, which is what Omni offered out of the box.
[16:07] So, in six months, um, we actually started this project in October. Um, which if you're in like sales and marketing, Q4 is like a little bit busier than other times. Um, so we had to make sure that we had something
[16:19] really fast and really quick to people to show people. So we have sources across social display which is just a display ad on a website branded content um our sales systems and our first party audiences. This is all ingested and
[16:33] housed in data bricks. So um we have a pipeline that uses DBT and data bricks asset bundles that then connects to Omni um which is then embedded into a a uh
[16:45] octa app called the performance studio because as Kane NAS we have to brand all the things that we do especially internally um so now that we have the performance studio we launched and we have three main
[16:59] pillars of our of our dashboard suite so it's the insights hubs which gives people access to things like how did this brand perform and I'll get to this later in the AI thing. Brand at Conde Nast means something very different than
[17:12] brands to most people. Um brands are at magazines and we need to know how categories performed and be able to also give people insights into what markets are performing. So we have 14 different markets and 14 different brands that all
[17:25] need to be serviced holistically and consistently across the entire organization. Um, campaign monitoring is very revenue operations focused, campaign management focused to ensure that things are actually tracking and pacing. And I'll get to like again, this
[17:39] might seem simple and table stakes, but beforehand, this was not something that we all had. Uh, and then resources, just benchmarks, definitions, etc. And again, we're still making sure that this moves
[17:51] along because now that we have 15 data sources, 14 different performance dashboards, people are spending 14,000 minutes. And now last we checked was about 30,000 minutes a week for a 100 people is a ton. And then we've
[18:04] automated five different processes with this whole process. Um and again like dashboards are great and you hear a lot about dashboards kind of going away but as Colin showed they're not. Um but
[18:16] where we were before versus where we are now is we had multiple reporting endpoints. People would go to Facebook or to Google directly to get data. you had inconsistent reporting inputs that would go to a client. Um, so when my
[18:29] team joined, we made sure all this stuff was centralized, standardized, normalized and we did it with engineering best practices in mind and also ensuring that we involved our end users from the start to ensure that they
[18:42] got the questions to the answers that they needed. Um, this allowed us to automate a ton of work and a ton of processes in the process. U, this is one example. Omni has a really cool spreadsheet functionality. Uh before
[18:54] people would have to go to like three different CSVs, download them, put them in Google Sheets, write a macro and make sure it all kind of talks in one place and now we just have one filter that allows people to get all the information
[19:06] that they need um through spreadsheets. This gives everyone like a client ready deliverable download, export, send out to a client that saves about 10 hours a week for about 50 people. So um we get
[19:20] asked for reporting like every single day and this allows them to do one click download send it out really simple application but again incredibly helpful. Uh another one that we've done is identified a super manual workflow.
[19:33] So every day we have to make sure that we are pacing to deliver for our clients. Um we define pacing revenue at risk and all the delivery metrics differently across teams. So we identified a manual workflow where
[19:46] different teams had different spreadsheets with different numbers and had to communicate all of those to one CBO. Um so we worked with everyone to define a shared business logic. Make sure we're all talking the same language. Again, that's obviously
[19:59] incredibly important now. Um, and we put it all in one spreadsheet that one dashboard that that automates everything across all those teams and it does it four times a day instead of one person pulling one thing at 11:30 a.m. and
[20:12] ensuring everyone has that data once. So this is now again four times repeated. And then we also added AI on top. So my team is uniquely positioned as incredibly data literate and then also
[20:24] really close to client deliverables. So we know how people are talking about data in the marketplace. what are the KPIs for a deal and how should we be talking about our revenue and campaign performance overall. That means we can then are we are then well qualified to
[20:37] add in context to our users. Um that allows for really smart and actionable AI to be placed within the dashboard directly. So the biggest thing that we have is servicing marketing and sales
[20:49] people is that we are you know marketing sales people are not necessarily the most data savvy which is fine. That's not their job. Um, but we want to be able to place data points and and data
[21:01] coverage within AI and have them be human readable next to like the charts that they're looking at. So, this again allows a really seamless experience, allows them to trust the data, allows them to trust themselves when they're
[21:13] actually talking to a client. Cool. So, how are we going further with AI in just a dashboard world? So, one thing we've done and what Colin shared earlier and I think um you heard in the
[21:25] keynote today about context, right? That's the biggest thing that we know with AI will make it most impactful. So we started shared documentation and again Colin said earlier like shared feedback and shared documentation that
[21:37] we need is essential right you need to continue to iterate and make sure these tools are producing the results that you expect. So we started with defining all of our metrics with our data models just exporting those. Um, and then since we
[21:50] are the ones that defined what a metric was, what a definition was as the uh subject matter experts, we added in our context. One thing that was really fun for us is we add in our own context that again synonyms, couple words that we
[22:03] just would provide more information. We would also then add that into chatgbt and be like, is this good? And that was actually a really fun process to know like how those things could iterate. Um, so once that was all done, we put it all
[22:15] in added to Omni as a context in the model layer and then also added in on definitions and metrics along with synonyms. So when people do ask questions about like revenue or click-through rate or engagements, you
[22:28] know what they're trying to decipher because we added in that context for you. Oh, is that right? That's right. Cool. Um, sweet. So
[22:41] whenever someone goes to get reporting right so this happens every single week we you go in you filter for a dashboard and then we provide the context for you so you have like an actual campaign insight that we would write ourselves.
[22:54] So our my team is small we have to service the entire global organization. So doing this for every single campaign is just not feasible from a person. So we basically encoded our brains into context that is living in a in Omni and
[23:08] that allows people to get the insights that they need to prove performance. So, it's referencing our benchmarks, our definitions, all the pre-sale assets that come in of what the campaign objective was and then joins it together with the data we have in Omni in order
[23:21] to spit out like a result that someone can take internally, communicate that to their salesperson and everyone feels confident in the results and feels confident going to a client and delivering this information and saying how condrove performance.
[23:40] Cool. Um, again for more AI, we are now focused on actually getting something in the dashboards that is an AI overlay. So when someone goes into a dashboard, they can ask a follow-up question self-service. This involves a ton of monitoring, a ton of logging. The one
[23:53] thing we've done is we've scraped all of our Slack channels and all of our emails to ensure we know the most common questions. And then we've encoded business logic that allows us to get consistent results. So I'll take an email that I got of you know what brands
[24:06] are best for Bisfy tech. Bisfy tech means business, finance and technology and it's a category that we all service. So encoding that context into the a into the data model itself and saying bisfy
[24:18] tech is either business or finance or technology and a brand is actually not like you know Oreos. It's actually Vogue. So those that type of information is key for it and it learns over time
[24:30] and gets smarter with context. Um, so we're currently logging these questions and asking it and scaling these things. Um, once we get an answer back, we will all grade it as a human on a one to five
[24:42] scale and ensure that it actually gives the answers that we would expect because we're the ones providing those answers normally. Cool.
[24:55] Awesome. So overall, again, this is like five years um ago that I started at Kai Nest. We were a super reactive industry. Um, and we were making sure that people got answers, but we were only doing that by fielding it. And now we put everything to the forefront of um,
[25:09] people's eyeballs because now instead of us asking questions, we can point them to a dashboard that has self-service self-service analytics and gets insights that they need instantaneously. Um, cool. And that's our journey.

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.