Sponsored by: Prophecy | From Weeks to Minutes: Democratizing Marketing Analytics

Summary

Traditional marketing analytics requires weeks of handoffs across multiple specialized teams to translate business intent into production pipelines—a process Prophecy and Acxiom have replaced with a single AI agent workflow.
Prophecy's specialized AI agent learns schema, relationships, and data distributions from enterprise marketing data, understands business intent in natural language, and automatically invokes propensity, affinity, and segmentation models to generate governed Spark and SQL pipelines on the Databricks Data and AI platform.
By embedding institutional knowledge into reusable AI skills and using Unity Catalog for governance, the solution reduces a multi-team, multi-week process to a conversational workflow that a single marketer can execute without writing SQL or managing handoffs.

Chapters

FAQs

What is Acxiom and what marketing data problem does this video address?

Acxiom is a marketing agency and data company—part of Omnicom—that specializes in building marketing data foundations using identity resolution and data enrichment products. This video addresses the challenge of translating business intent (such as identifying customers at risk of churn) into production-grade analytics pipelines without requiring SQL expertise or lengthy multi-team handoffs.

How does Prophecy's AI agent generate marketing data pipelines?

The Prophecy AI agent learns the schema, relationships, and data distributions of enterprise marketing data, then interprets natural-language business intent to automatically invoke analytical models including propensity scoring, affinity analysis, and segmentation. It compiles deterministic Spark and SQL workflows that run on the Databricks Data and AI platform under Unity Catalog governance.

What is ID resolution and why does it matter for marketing analytics?

ID resolution is the process of connecting disparate customer identifiers—from CRM systems, web activity, purchase records, and other sources—into a unified customer profile. This video explains that Acxiom's proprietary identity and enrichment products perform this matching at scale, giving the AI agent a reliable, de-duplicated customer view to work from.

How does this architecture ensure auditability and reproducibility?

The solution embeds institutional knowledge into reusable AI skills and executes all pipelines on the Databricks Data and AI platform under Unity Catalog governance, preserving full data lineage and auditability. Because the agent generates deterministic Spark and SQL code rather than opaque model outputs, every pipeline can be reviewed, versioned, and reproduced.

Full transcript

[00:09] Good afternoon everyone. Welcome to the session on talking about democratizing the marketing analytics. Um we're going to introduce ourselves uh in just a moment. But before we start the session um this is just a disclaimer from data
[00:22] bricks that whatever presentation whatever we say whatever we do here is basically um our own views and data bricks doesn't take any liability for that. Um also just a reminder about
[00:35] filling the survey. Please fill the survey after the the presentation is over. that will be greatly appreciated by Axiom Prophecy as well as uh data bricks as well. So here is our session
[00:48] uh basically it's about democratizing marketing analytics and the aim here is to kind of show you what the traditional architecture what the traditional way of harnessing analytics from your marketing
[01:01] data versus where the future is going powered by Axiom powered by data bricks and powered by our partner uh prophecy as well. So with that um we'll start with some introductions. Baser you want
[01:13] to introduce yourself? >> Yes. Hi everybody and thank you for joining us in this session. I'm Basar. I lead uh AI product management at prophecy and my background is uh basically building distributed data and
[01:25] machine learning platforms for over 17 years and last couple of years I've been building agentic AI platforms and uh prophecy is essentially based on agentic data prep analysis platform and uh
[01:37] trusted by enterprises like JP Morgan, HSBC, Microsoft and of course Axium which is why we are here today. Um yeah >> thank you. Thank you Baskar. And my name is Ankur Jane. I am chief cloud and data
[01:50] modernization officer at Axiom. U I joined uh Axiom a couple of years ago and um I don't know how many of you are familiar with Axiom but Axiom is a
[02:02] marketing agency part of a bigger larger group called Omnicom. And think of Axiom as um you know an agency which helps you put together marketing data foundation
[02:14] um bringing in disparate data sources and connecting the dots you know through our proprietary identity and data enrichment products you know so that's a little bit about Axiom um and you know let me start the
[02:28] presentation with this uh question you know which is okay imagine a marketer who is trying to get to find out okay who these audiences are or who these customers are who are basically at risk
[02:41] um at churn you know like who will lose this brand in next 30 days 60 days 90 days and what should the brand do to retain these customer it's a very logical question but yet very powerful
[02:53] question um just a show of hands how many of you are actually uh from data science background because it this question does involve a little bit of data science I see a few hands how many of you are data
[03:06] engineers in the room. All right, we have some few data engineers. Uh anybody directly related to marketing? Anybody um from the business side? Okay, we have a few, you know, people who represent
[03:18] marketing as well. Well, that's that's good good to know, you know, because I think this question, as you all can imagine, requires a little bit of everything, you know, right from it spinning up the platform to data engineers bringing the data and then
[03:31] data scientists performing the analysis and then kind of providing those insights to marketing. So, I'm guessing this picture probably resonates with all of you. Um to answer this question you need an environment like this where you
[03:45] bring in you know multiple different data sets whether those are structured, semistructured, unstructured data sets and depending on the frequency of how you u acquire that data. You bring that
[03:58] data in a streaming fashion or in a batch fashion. That batch fashion could be nightly batches or weekly batches or hourly batches, you know, depending on the frequency. And that's when the fun
[04:10] begins, right? Like you see this data processing block in the middle. That's where you have relational databases or platforms, you know, highly parallel platforms like data bricks in the middle. You bring in some tools to kind
[04:23] of connect the dots. And then um ID resolution. How many of you are familiar with a concept called ID resolution? Um I know XM folks are here, but ID resolution is basically customer
[04:36] identity. you know like how do you resolve the customer identity across various different systems how do you recognize them it's the same person who comes from various different systems right and that's also um where agencies
[04:49] like Axiom come into picture right how they work with the brands to bring that first party data resolve the identity and provide the results back and then you put some governance on top uh which is what we are calling observability to
[05:02] see okay what the lineage looks like you bring in specialized tools like Calibbra or Monte Carlo or a few others to kind of see okay how how the system is actually performing and then towards the
[05:14] end is when you start to actually derive the insights you know your reporting your BI uh comes into picture nowadays you bring in specialized AI tools to kind of answer some of those questions
[05:26] and then finally once you have the insights finally once you are able to build your audiences or finally once you get the questions that you are looking to answer Then you activate that across various different channels. Right? So I
[05:40] mean it's a very complex picture. I've simplified this. I'm sure you all data engineers, data scientists know what all it takes to kind of put this together. But these are the three main things that we want to actually hone in, right? This
[05:54] picture is pretty sophisticated picture. But there are three barriers. You know, think of it from a business perspective. Think of it from a marketer perspective. It takes time to build this architecture to build this solution right to bring in
[06:06] this data to put observability to resolve the quality of the data to resolve the identity to derive the insights it's a very timeconuming exercise plus it takes you know a village of different kind of skills to
[06:19] come together to actually answer these questions so you need people with data science background with SQL background with some statistical analysis type of background and then also u coming back
[06:31] to the democratization ation which was basically the title of this whole presentation. Um you know marketers are basically u at the mercy of these u different kind of skilled uh resources which may reside in different teams
[06:43] different organizations and they basically have to go maybe open a service help desk ticket or a trouble ticket and then somebody will pick it up then basically turn this into a program
[06:56] which will then basically u engage these specialized resources to provide those answers. So it takes time, it takes heavy skill sets and it also basically um is not basically as fast as one would
[07:09] imagine. And that's the whole thing that we want to talk about in this presentation that okay, how can we in this age of AI, how can we uh make this process faster? How can we democratize this analytics in the hands of business
[07:22] people themselves? So that's the whole idea that these traditional ways of sharing data and serving intelligence won't work anymore. Um what marketers need is basically a
[07:34] self-s served data foundation which is powered by AI, right? Because AI is disrupting everything. That's the future of marketing as well. Agentic AI and that's what we want to show you through this presentation of how this would
[07:47] work. So imagine a scenario where instead of um opening those trouble tickets and reaching out to those different teams to kind of get you the answers, what if you could just ask what if you could just
[08:00] say okay type this question to say okay who are those customers who are at the risk um in basically losing this brand in next 30 60 90 days. So that's a very simple question but then it gives you it
[08:14] builds your contextual audiences and we'll talk about more about what we mean by context in just a second but these audiences which are built are built in that context in which the question was asked and that gives you the ability to
[08:27] activate these audiences across various different channels right so how do we make this possible this is basically made possible by data bricks and prophecy and this is what Baskar is going to show you how this all kind of
[08:39] comes to life so bas let me hand turn it over to you. >> Thank you. All right. Thank you so much uh Ankur for setting the stage here. Uh now let's quickly look at what prophecy is like.
[08:51] We're a self-s served data foundation you know for business users. Now obviously prophecy as a core is an agentic prep and analysis platform. It's a fancy way of saying that agents do all
[09:03] the heavy lifting of data prep and analysis which you wouldn't want to do which is the most complicated part of that business and you're basically doing decisions and strategies and so on. Now
[09:15] prophecy is uh works on your data which is sitting on your data bricks environment or any other cloud data warehouse as well as in-memory system which is your execution environment in case you don't have a cloud data
[09:27] warehouse. But why we're here today is to talk about verticalized AI products on top of this platform. How you as a marketer, as a uh structure in structured finance or as an auditor or
[09:40] in fact any other business user can very quickly do data prep, data analysis and basically look at it from a business point of view. So Angur earlier talked about
[09:53] uh the three biggest challenges time skills and the enablement and this really why we are here to show that prophecy is trying to eliminate this entire handoffs
[10:08] by leveraging agents which do all the heavy lifting. Now what you're seeing in the screen is an AI native process of how you would actually go about as a
[10:20] marketer. Now, you know, in case many of you are not aware, one of the first things that you need to do as a marketer is to build an audience. And building that audience requires you to gather data, real-time data across from
[10:32] different dozens of different systems, being able to make sure that it's clean and it's resolved in terms of identification and then of course continuously getting signals from different systems. So what we do in
[10:45] terms of prophecy is that we enable we offload the heavy part of generating that initial audience to the agents and I'll explain a little bit on how we you know what's the secret sauce behind it.
[10:57] But this process what it does is is just like you're writing code using AI and then refining it and until you go into production the same way you're generating audiences and once you go back and refine depending on what your
[11:10] business strategy is what your criteria is for that particular campaign that you want to launch and finally once you're happy with that you activate the campaign you start measuring the data that's coming in how many people click the campaign right what was the
[11:23] conversion rate and so on and based on that you get iterative feedback so this process as you can see here right requires data requires knowledge and skills and of course it requires handoffs across different groups and this is what takes several days or weeks
[11:37] when you're working in a business in an enterprise marketer context where you can't just make a decision uh because it obviously you know comes at a cost so
[11:49] the next part is actually going into the demo but I think this is where I'd like to >> yeah no I mean Bhaskar what you just described here on general generate, refine and deploy, right? That was basically something that caught Axiom's
[12:02] eye and that's how we partnered with Prophecy to test it a little bit further, right? Because um this is basically what marketers are asking for and that's the business we are in. So we we partnered with Prophecy and we'll
[12:14] show you uh what the results look like. But before we get into the demo, um I just want to um emphasize that before you even put AI to anything, right? the quality of the data uh the the
[12:27] foundation of the data also makes sense right because the results will be as good as the data which is underneath. So just to take a moment here you know like so we part we actually piloted this solution with one of our automobile client and think about the automobile
[12:42] scenario and this particular client was actually thinking in terms of um reaching out to customers who were coming up on lease in next 30 60 90 days so that they can target them you know for um another luxury segment. Now
[12:56] imagine you know the kind of data that marketers in automobile industry deal with. You know not only do they have customer information which is coming in from the CRM system but they're also getting telemetics data which is coming
[13:08] in from the cars itself. They're also looking at financial data like how many leases are coming up uh what kind of car loan information that they have. uh if they are running campaigns how those campaigns are performings, how many
[13:21] performing, how many people have actually opened or clicked or um not even opened those emails. So lot of different data. Now how do you connect going back to that identity resolution topic that I was talking about earlier
[13:34] that's where basically the identity comes in how do you stitch all this together so that this becomes an actionable data foundation and that's basically um the point I wanted to emphasize that you need to first sanitize this data clean this data
[13:47] resolve the identity enrich it with more attributes so that you know more about your customers and that's the foundation you know which is necessary before you even actually put AI on it so u with that Basar I'll give it back over to you
[14:00] >> so that you can describe what the uh demo looks like. >> Thank you. So I have uh separated the demo out into segments so that it's easier to digest. Uh the first thing is once you have this data, right? This is
[14:13] live data. This is not a set of data that's lying around where you're getting a dump and they're running analysis. You constantly getting signals from dozens of different data sources. Now once we have this always on data what's the
[14:25] first thing that you do as a business user as a marketer you need to understand what it is right and then you have a strategy right what is it that you would like to do in terms of the next few weeks let's say you want to launch a campaign um so the the first
[14:38] need of it is to understand what this data set looks like and this is where as you can see here you know you're talking to the agents and the agents is like hey I got this new database okay like I want
[14:51] to first explore what it is. I need to understand what tables are there, right? What kind of data is there? What is the quality of the data? So the agent understands everything in terms of the
[15:03] data distribution, referential integrity. Uh it looks at what kind of data is there, how many unique individuals, how many unique households, all of these information it extracts and it now creates what we call as a context
[15:17] graph. And part of that context graph is the knowledge layer. Now without this information as to how joints are you know have to be done across tables you know what how do you dduplicate data without knowing all of this you cannot
[15:30] really generate actionable outcomes. So this is the the first thing that you need to do in order to be able to actually go into the generate refine deploy sort of uh you know AI native way
[15:42] of exploring data. Now some of you might be thinking hey you know any AI today can write SQL and and that's true of course I mean the LLMs are extremely advanced in terms in
[15:55] terms of uh generating code now the challenge is that is it explainable governable code is it reusable code can you inspect it if you cannot then are you going to invest hundreds of millions
[16:07] of dollars in terms of your marketing budget and are you actually tracking the metrics right so a lot of the optim optimization has to happen based on that and and how we do this right is like when you're asking the simple question
[16:20] of what the business intent is what comes in between the input and the output is this whole concept called context graph uh I'll just pause there and see if anyone is aware of what a
[16:32] context graph is in terms of AI any raise of hands all right we have a few folks uh very few folks here but fundamentally what happens is that you know in order for in you know leveraging
[16:47] LLMs as you can see like when you're chatting with it it needs deep context and the deep context happens based on the domain it has to understand what your business workflow is it has to understand what audience means in terms
[16:59] of automotive database context it needs to understand what you in the in terms of tribal knowledge how you actually do things as a data engineer how do you join two tables between which which keys things of that sort right how do you
[17:11] eliminate certain sort of problems that are there in the data a lot of these things reside in people's heads. So the context graph is a way of uh taking all of that decisioning knowledge right
[17:23] memories memories in sense of what did you do earlier and then it's a compounding layer and this layer is what really differentiates between how effective your AI is going to be when
[17:36] you're using it in context of a business operation. So on the right as you can see here right is governed output. What it really means is that the output is
[17:48] not just hey here is the action. It tells you what that data pipeline looks like in visual way which you as a user can go and look at the aggregates the joins the ddups all of these gems that
[18:00] you can see on the top and these are backed by SQL code or Python code depending on which system that you're using. Now all of this is inspectable and this is the only way that you can actually govern your data pipelines and
[18:12] of course all of this runs natively on top of data bricks. So not really using it on shadow it you don't have to worry about security you don't have to worry about trust. So this is really what the you know differentiation is in terms of
[18:26] what prophecy brings to the table. So let's go back to the first question right who are my high propensity trading customers. Okay. Now when you ask this question, the first thing the agent needs to do with all this context graph
[18:39] and the unique ID resolution is that it figures out what is the intent of your question. Now in this case as you can see here u the agent actually figured out that hey you you asking a question
[18:52] in terms of tradein propensity. So it understand what that propensity is and he says hey I'm going to build you a model and that model is a reusable model which can be used across you know different use cases as you'll see later.
[19:05] Now in this case the model is coming and saying I built this model and this is what I can see in terms of what this model is doing and these are certain signals. for example, it's saying that okay uh these are your high tier
[19:18] customers who are potentially ready for tradein and three to five years sweet spot right is basically giving you a set of signals you can now iteratively change this right you can come back and say I just want to know for the last one
[19:31] year how somebody's you know doing right and who are my trade in you know for the last 6 months so being able to change these signals iteratively is what we call as your post after generation you go back and forth and this is really the handoff part where instead of you having
[19:45] to go and re write SQL or write a pipeline. The agent does that for you and you inspect it. So happens within a matter of minutes. Now in this case I'm going to show you what this pipeline looks like. Right? So
[19:58] this pipeline right is what it gets generated. It shows you which tables are joined in what way. Uh how it was filtered, right? How it goes all the way
[20:10] to the final output that you're seeing on the application. Now all of this is backed by code. So this code is what you own and you can inspect it, you can change it. All right? All of this code is reusable is CTE SQL. So you can
[20:23] basically change them and you don't have to change it manually. The agent does that for you and eventually it compiles it. It it tests it out. So all the heavy lifting of compiling, testing, verifying
[20:35] it on your data, the agent does that for you. So once you have this let's look at what's the output right and finally the output is an analysis app that you can
[20:48] actually deploy. So what do we mean by deployment is that you know you want to track this over time. You want to come back and say you know I want to see the service threshold over different two
[21:00] months or I want to change the high propensity threshold over a different types of days. Right? So being able to leverage this data pipeline that was created and look at the output and and and remember that this is always on data
[21:12] which means that as new data is coming in as somebody's clicking the email or as you're getting conversion data or other audience interaction data all of this is getting updated live and these are um apps data apps which you can
[21:26] deploy and once you deploy you have the final output saying okay these are my scored customers right this is my final target campaign main you know um folks that I can target this uh application
[21:40] and you can export this to your data platform whichever you're using to run these campaigns. Now of course all we're doing all of these things right but one of the key things why we have an AI agent you know
[21:53] to close the loop is to uh really strategize on what your uh you know marketing strategy is. Now for example
[22:05] um you want to optimize on revenue let's say okay like when you're trying to target customers for upgrades the agent now with all this context is coming and telling you here are potentially a set
[22:18] of campaigns that you can actually target. Now, in this case, as you're seeing, he's coming back with a premium loyalty upgrade campaign. But the target here is that revenue lift that you're seeing, which is $13.3 million, right?
[22:30] So, it tells you that uh if we target these users, these these customers, and based on the criteria that you had set, this is what your revenue uplift is going to be versus another campaign
[22:42] which is converting highinccome non- buyers with first-time incentives. Now these are recommendations which in certain industries you would see as next best actions or recommendations but this is what where you're strategizing as a
[22:55] marketer and of course this is a back and forth uh sort of system right you can go back and say I want to change this and I want to reduce the number of people that I'm targeting and so on. So all of it is doable immediately on the
[23:07] spot with the agent and this is really how we close the loop from understanding the data from generating refining deploy as a loop and finally strategizing executing your campaigns. >> So with this I'd like to hand over to
[23:20] Ankor. >> Yeah. No thank you Basker. I mean this is basically coming towards the end of our presentation. Um I hope you saw you know how powerful conversational AI is when you pair it up with platforms like
[23:32] data bricks where you actually bring in marketing data across various different platforms. Right? So this diagram here basically just summarizes that data bricks becomes your underlying fabric where you store the marketing data and
[23:44] then you bring in you know um capabilities you know uh that things like XM provides which is basically your data and identity um enrichment capabilities um to kind of connect those dots connect
[23:57] the identifiers and then essentially bolt-on conversational AI capability from a platform like prophecy to kind of build your audiences and activate them so That's basically the the the the you
[24:09] know uh presentation here. Thank you all. Appreciate it.

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.