Sponsored by: Precisely | Architecting Agentic-Ready Data Pipelines with Databricks and Precisely

Summary

Precisely brings 60 years of data integrity heritage—from mainframe data movement optimization through enterprise data quality—to the agentic AI era, where 95 percent of executives surveyed report that AI initiatives are falling short and nearly three-quarters cite data as the root cause.
Using a multi-agent architecture on Databricks, Precisely demonstrates how free-form text from customer support tickets can be transformed into structured, agentic-ready data through a pipeline of specialized agents including NER classification, address fixing, and spatial enrichment using native Unity Catalog UDFs.
This video shows a live demo of a telecom customer support use case where a multi-agent system resolves a ticket end to end, illustrating how Precisely's data integration, quality, governance, location intelligence, and enrichment capabilities embed directly into Databricks pipelines.

Chapters

FAQs

What is 'agentic-ready data' and why is it difficult to achieve?

Agentic-ready data is data that is accessible, contextualized, current, reliable, governed, and automated enough for AI agents to act on with confidence and accuracy. This video explains that most enterprise data—particularly high-value customer data buried in support tickets, call center transcripts, and scanned documents—fails these criteria because it is unstructured, inconsistent, or lacking enrichment context.

What agents does Precisely deploy in the telecom use case demo?

The multi-agent pipeline demonstrated in this video includes a NER (named entity recognition) classification agent that extracts structured attributes from free-form support ticket text, an address fixer agent that validates and standardizes location data, and an enrichment agent that adds spatial coverage and building attributes using Precisely's location intelligence capabilities.

How do Precisely's capabilities integrate with Databricks pipelines?

Precisely's data integration, quality, governance, location intelligence, and enrichment capabilities are embedded into Databricks pipelines as native UDFs (user-defined functions) that run directly in Unity Catalog. This approach allows data engineers to call Precisely's validated data operations within standard Databricks workflows without requiring external API calls or data movement.

Why has enterprise AI investment fallen short according to this video?

This video cites a finding that 95 percent of executives surveyed say their AI initiatives are falling short of expectations, with nearly three-quarters identifying data as the core problem. The argument is that AI models are advancing faster than enterprise data foundations—models are capable, but the underlying data lacks the integrity, completeness, and structure required for agents to act accurately and reliably.

Full transcript

[00:09] Thank you so much. So, apologies for the delay everyone. We've got a great session for you today. Dan Tatro, Senior Director Global Solutions Architecture here from Precisely, joined by my good friend and colleague Mayank Srivastava.
[00:21] Um today how we architect data pipelines are actually ready for agentic AI using Precisely and Databricks. Um everybody is racing to adopt AI.
[00:34] Um models are only as good as the data um that we feed them. And without trust, confidence, um completeness of that data, the exercise is futile uh and a bit of noise. So, by the end of this, you'll
[00:47] have a clearer picture of what this story looks like. So, a little bit about Precisely, 12,000 uh customers worldwide across 100 countries, 83% of the Fortune 500. Uh
[01:01] we've got probably close to 3,000 employees, 2 billion going into R&D. Um but I really want to give you a sense of the history behind these numbers, of where we came from. Um we started as uh
[01:14] Syncsort, so basically gold standard in the uh about 60 years ago in the mainframe space, uh specializing in data movement optimization um when the mainframe was um mainstream
[01:27] enterprise compute. Uh then through Pitney Bowes Software, um we became the pillar of enterprise data quality. Um address validation, entity resolution, customer um
[01:40] data management um at global scale. Uh so, we have watched companies navigate through these eras, mainframe era, data warehouse, big data, data lake, lakehouse, all of that great stuff. Um
[01:53] and the ones who were successful in those areas were those who got data right. So, now here we are in the AI era, the game is still the same. Um the foundation is still data integrity, um
[02:07] and that's been our North Star for decades. Okay, let's just acknowledge how fast this environment, AI, all of it is moving.
[02:20] Um token maxing in February has now become token efficiency. Uh AI went from search to summarizing meeting notes to autonomous agents, and now we've got
[02:32] compute going to space. So, the models alone, baffling. GPT Opus Flash Flashlight Flashlight, I think there's a space in there. Anthropic just dropped Fable, uh
[02:46] a nerfed version of Mythos. We have the term nerfed and flashlight in the same paragraph, and we wonder why we're failing to realize um the investment on AI. So, um all of this being so, 95% uh
[03:02] of the executives that we have um uh pulled say that these initiatives are falling short, and nearly three quarters say that data is the problem. So,
[03:14] AI is outpacing uh the data foundation, and that's probably where the gap of of return is is lying. McKinsey says there's 4 billion bucks on the table. I think that's probably just a drop in the
[03:26] bucket. Um you can go to the next slide, man. Okay. Um way back in 2025, every company declared itself a data
[03:39] company. Most didn't follow through or they forgot, because just a few months later every company's a technology company, every company is an AI company. And the
[03:51] conversations that I have is is it truly an advantage unless you're Nvidia or SpaceX, you're probably one of the 99% that's got to find your actual true competitive advantages to find out where
[04:04] you participate. So I have these conversations. It's wonderful. We talk about community, we talk about tribe, we talk about identity. The brands that you affiliate with is truly, you know, kind of the competitive differentiation. You
[04:16] think like Starbucks holding the actual cup, the logo we go every morning. Harley-Davidson, you want to have that patch on your jacket, your leather jacket. And even the Patagonia sweatshirt. Like those are competitive
[04:28] moats. The networks are fun one to think about. Every time somebody signs on to LinkedIn or or Airbnb, you know, that is that's a competitive advantage. Those are millions and millions of people that you
[04:41] would have to actually build up to compete against those folks. Is it hard necessarily to create the technology behind those offerings now? Probably not. Fable just came out, Mythos just came
[04:53] out. We could just tell it to build that, you know, those that infrastructure, but it's really hard to do. Um Just go back. Can you go back one more slide? Um
[05:05] Where was I? Ah, yes. Data, of course. All of those companies that said that they were data companies just, you know, just a few months ago. All of the companies that are here, you know, data
[05:18] remains to be a competitive moat for so many. In so many industries, we've built data sets for decades. We've built built data sets for centuries, actually. And
[05:31] at least a part of our DNA. Like that's that's kind of what makes these companies uh really sing. Um Myung's going to show a demonstration um in just a bit about a customer who made
[05:44] um customer experience their moat. They're really good at it. They wanted to kind of capitalize. They paired trusted data with AI to defend that mode and actually widen it. So, that's going to be that's going to be my uncle in just a
[05:56] couple of slides. So, um when we take a look at the enterprise AI architecture, we've kind of see these things in the very next slide. We've seen this, you know, kind of market texture from lots and lots of vendors.
[06:11] Um and we probably can from guess where a lot Well, maybe not in this room, but where a lot of the investment is. A lot of the shiny things. We've got models, we've got agents, we've got applications, we've got harness, we've
[06:23] got harness as a service. We've got all of these wonderful things. Um and and really there's some fantastic things. I mean, um models today can take a single prompt, work through the night,
[06:35] self-debate, course correct, deliver complete systems by morning. It's incredible. I I actually can't even think of a prompt that I would give my AI to have it work for longer than a few minutes or
[06:47] an hour. So, you know, maybe there's something there. Um Prompt iterate was a huge leap, but it's no longer a ceiling, it's the floor. So, um
[07:02] I think, you know, really what happens is, you know, we we've you know, we have the conversations uh to kind of redirect uh our customers to understand um that the trusted data layer, you know, that's really um you know, your business context, your
[07:15] rules, your relationships, the history of an organization, that's the DNA of you. And it's what makes your data yours. So, it's it's a thing that your competitors can't replicate. Um
[07:27] and, you know, those companies that want to win the agentic era will not have the best founda- not choose the best models, they'll have the best data underneath it. And that's that's the real story today.
[07:40] Um You can go to next slide, Mike. So, this is a fun one. This was a surprise from my marketing team who gives me a new slide and expects me to speak to it on a moment's notice. So,
[07:52] I'm going to say all of these things are wonderful. But, I want to challenge the room. Really? So, I wrote this this morning because I didn't really want to present to this thing. And I said,
[08:04] have all these models are so powerful. They're creating things overnight, you know, but the stakes have changed. It's not just prompt and iterate. It's not just, you know, it's not just meeting
[08:16] notes. It's really I listened to a podcast on the way you know, probably the fifth flight trying to get here and I came across the term task imagination and and the gentleman's name is Nate B. Jones. I
[08:28] listened to him on TikTok. And the idea is that the next unlock isn't better prompting. It's fundamentally imagining what problem you're solving and what a solution could even look like.
[08:42] So, most organizations are using AI to do existing things faster. The ones that are pulling ahead are doing things that they couldn't have conceived to do before.
[08:54] So, yes, all of these things matter, but the bigger question is are you up-leveling your understanding of an AI? Um because the gap between how most companies are using it and what is possible is is pretty staggering.
[09:08] Okay. Little obligatory slide on how Precisely powers data integrity. Organizations have these pieces. They have data sources, they have pipelines, they have AI tooling, but what they're
[09:20] missing is the connective tissue and that's what Precisely provides. So, all of these are the key ingredients of building a trusted data layer. Uh they're not nice to have. These are the actual requirements to to kind of um
[09:33] realize that return on investment. We have things like interoperable APIs, you plug in your stack, you're not going to rip and replace. Uh we have common meta metadata foundation that goes across all of these capabilities from the catalog to quality
[09:48] uh to enrichment so that you're kind of breaking down those silos. Um this is going to run where your data lives, cloud on prem hybrid because the data's not Well, the data's moving.
[10:01] Where you store the data isn't moving. Uh so, neither should your integrity layer. Uh and without it, you don't have any AI architecture. So, you have a very expensive experiment. So, building a trusted data layer, I'm going to get to my aunt. Yep, very next slide.
[10:15] So, this slide is the blueprint. Um the foundation on the left and guardrails on the right. And we're going to touch on guardrails here real quick. Um but today, Mayank's story is going to be about the foundation. Guardrails, we
[10:29] got governance, we've got policy, we've got access controls, all that additional information that knows how a business operates and how to leverage the data. That's on the left-hand side is part of that DNA. It really matters.
[10:42] It's not so much fun to demo. So, we're going to we're going to focus on the foundation. So, before on the foundation side, you're going to be bringing the data together. We're going to be leveraging data quality tools to make sure it's clean and complete. We're going to enrich it to make sure it's
[10:55] contextual. And then we're going to watch it. Um we're going to observe it over time to make sure that it stays fit for use. So, um Mayank's going to walk you through uh that foundational element. Uh of course,
[11:07] on Databricks not slides, well some slides. Um >> Some slides. >> Yeah, but real agents doing real work. Thank you, Mayank. >> Awesome. Thank you, Dan, for laying out the
[11:19] importance of uh the data because um we are living in this world of AI and ML and we only focus on models, we focus on uh how much tokens we we saving, how quick the models are running, but in uh
[11:33] the information and the importance of data which you covered then thank you so much that lays a good foundation. So, I'm going to cover some fun part which is included the the actual demo and then few slides.
[11:47] Before that I want to talk about the agents. I was working on this slide and my dad called me and he asked me like all we know about AI is I can click a picture, I can
[11:59] make modification in that picture, I can write emails for my for my vendors, but what are these agents? And my dad owns a retail shop and I asked how many staff you have and
[12:11] he mentioned three. So, what about if we take those three people and convert it into an a machine that will do the job for those three people. He didn't understand. Then we break it into the
[12:23] pieces. He has three people. I was like, okay, you have one person who handles the sale, another person handle the inventory. Another person is handling the accounts. So, what if we have a machine that has a heart
[12:37] which will always make sure that the machine is up and running, that has a brain that will have skills of doing accounting, of doing the sales and doing the inventory. Then a way of communication
[12:50] which is your your human agents can talk. Similarly, my agents can also talk up from from an AI prospective or from a data bricks interaction the genie prospective.
[13:02] So, if we combine what a human provide you and replicate into a machine, that is what an agent is. I'm going to talk about three agents here. First one is the data classification
[13:14] that will take the natural language and classify the data. Another one is the address fixer. We are the location guys here and one of the reason I came late because location guys doesn't find able to find
[13:26] the building. so sorry about that, but the address fixer going to take the prompt and fix the addresses based on precisely MCPs. And then the last one is the enrichment layer, which will take
[13:39] the customer data and precisely data, combine it together, use the intelligence of AI, and um find a optimal solution of your problem. Data classification. We've been training
[13:53] the open source NER models. We have our data quality layer, so we combine the best of the NERs. Then we have decoupled it with precisely data quality, and on
[14:05] top we did some kind of reconstruction, and we come up with this ML agent, which will take a natural language and classify it into
[14:18] names, phone numbers, address, location, and noise. So if you see how it works is you give it a natural language, it will classify the data, and it will also add a score, because what if the agent is
[14:30] not sure about it, and it has a lower score, you can add the boundaries of what is acceptable and what is not. The next one is the address fixer agent, which is doing a particular job of fixing an address. Precisely has out of
[14:44] the box APIs, and we leverage those precisely out of the box API APIs and created an MCP layer on top of it. And what is MCP? It's model context protocol. That means
[14:58] you have an AI LLM, and then you're saying, "How about I give you a context of how my API works?" And then I also give you some rules how it works. So model, context protocol. We added them.
[15:12] We have address verify address API, we have geocode API, data quality API, noise filter APIs, we combine them together, put it in an MCP, and then added on top of the LLM layer,
[15:26] and that's how we created an address fixer agent. Now, this will just take an address, it apply the Precisely intelligence, and stop. But, an another step which my team has
[15:39] taken, they we have created a reasoning chain, we have created a multi-agent within this address fixer agent, in which my smart LLM will take an address, run it through
[15:52] Precisely, see that it doesn't look good, I can do still do better. It will go back. It will have a recurring passes that will will continue to run
[16:04] against the MCPs and my Precisely APIs until it get a satisfactory answer. So, anything which you can get out of a regular addressing API, my agent has
[16:16] smartness to take that address and run it multiple times with multiple configuration until we get a acceptable solution. It's similar to my dad's sales
[16:28] agent who will sell who will try to sell and convince the customer until the the until the customer is ready to buy. So, similar to that, we have this address fixer agent.
[16:40] Now, next agent which I'm going to talk about is an enrichment layer. So, you have an address, you have a name, you have phone number, you validated all of those, but what next? Until we add any value to
[16:54] your data, I think the work which we have been doing is not worth because we are here to add value to your particular data. So, for any address, if you I'm
[17:07] I'm just going to talk about address because I'm an addressing guy, I told you, sorry about that. But, for an address, we can append approximately 10,000 attributes. This This one on your screen on the right-hand side, are few
[17:19] examples based on the property. So, on this property, which we have seen in the previous slide, it's a residential apartment. It's It's of three floors, two bedroom, one living room. All that information we
[17:31] provide alongside 10,000 and other attributes. The right side of the The left side of the screen, you see we have some coverage map going on. I will You will get an understanding once you move forward into the demo, but just based on
[17:47] what's happening here, we have coverage maps. So, we all use phones. We all have different cell providers, and each has a coverage map.
[17:59] Somebody provide 5G with a certain speed at one location. Somebody provide 4G at the same location with certain speed. So, this data, obviously not owned by Precisely, but it comes from the
[18:11] customer. So, how we bring in the customer data in a safe ecosystem, use an agent on top of it is the magical thing which I'm going to show you next. And Databricks helped us a lot on this
[18:23] piece. So, customer data, Precisely data, joined together, and using multiple agents on top of this is what I'm going to show you next in my
[18:35] demo. So, in here, as Dan has shared a little spice that we've been working We're going to show you something where
[18:47] a human interacting with maybe a customer agent who is also a human, but not a highly technical human, or maybe talking to a bot. So, in that case, how
[19:00] a human's problem will be solved entirely by AI, and a smart AI is not just, "Okay, I cannot do that." or maybe a simple reference to documentation. No, we we have crossed that level. What we
[19:13] going to do is we going to do the agentic multi multi-agentic view running the agents in multiple view and then solve the problem of of our agent. And before that, we talked about the agents. Now, another thing is uh the data which
[19:28] Dan has talked about. Um data is very very important. So, what we have done is we have created a uh a source of truth table. So, think of this this this is a source of truth table for a telecommunication uh
[19:42] company and we have a different data within this company. If you see, we have the service number, the service look the service coverage, whatnot. And then this is the view of the sample you can see.
[19:54] In this we have the service type if this provider is providing 5G LTE, whatnot. And then approximately I have added uh 700 attributes in here. Uh all are neat and um
[20:06] uh very valuable in uh information that we have appended to this. Uh and another example of the uh data that I have added is the property type. If it is condominium, single family, multi-family. So, then I took this data
[20:20] and put it into the genie. Now, your customer data when aligned to the genie, the genie can answer anything. So, in here I'm saying, "How many 5G customer are living in the multi-story building?" Here is the answer. It's giving you a uh
[20:33] uh pinpoint answer. It'll also show you that on the map. Then you can say, "Okay, give me a breakdown of the construction type." And what it's going to do is it's going to go look into your data and then on that it's going to use
[20:46] precisely attributes and on top it will give you the best 5G coverage contains this this this uh groups of property type, uh construction type, uh the poor coverage has this construction type, and
[20:58] the legacy 3G has this construction type. Uh and then it will also show you on the map or on the on the table. So, the beauty of this is um This This has become your data engineer
[21:10] and I have not even asked a single thing. What I did did is ask a simple question, give it to Genie, and Genie will take the take care of rest of the actual things. Now, this is an example where one
[21:23] a person has contacted a telecommunication support, and I don't know if you can see it, but if you see at the bottom, a customer has been saying, "My name is this, my phone number is this, and I
[21:35] live at this location. And I'm totally done wasting my time with you guys. You promised me a speed of 700 megs, but I'm only getting 100 megs. So, an agent will look at this prompt, and
[21:49] obviously it's a kind of a negative one, but if you see the three agents which I have talked about, they're going to work together to solve Alex's problem, and going to call precisely MCP, going to call the data
[22:01] classification to classify the data, and then boom, they come back with the response that, "Okay, I got an account summary of Alex, whose phone number is this. And then, uh based on the enrichment
[22:13] spatial analytics, if you remember the last slide where we have 5G 4G coverage, based on that, we have found that there may be something going on. But let's see,
[22:25] we have some standardized address over there. We have the building ID associated to this. This is all coming from Precisely. We have the county data, we have the coordinates, and then uh it
[22:37] the agent will give you why exactly you're getting 100 megabytes instead of a 700 megabytes. So, your office, ma'am, is in the LTE space, your living room is in the 5G space. So, if you move your
[22:50] modem to your living room, the problem will be solved. Now, Alex will say, "I am not ready to do the do that, right?" Back and forth, but we now get the source of the problem, like why it
[23:03] happens. And then the agent will also uh give a recommendation that either you can do an equipment swap. Maybe that equipment's uh will will handle the edge cases. Or like um we will uh give you
[23:17] some credits. Uh whatever but whatnot. But if you see, it will come back with an definite response of what's going on. Then we have this mapping tool which is attached to the agents in the data bricks where
[23:29] you can click and go to the source of the problem on a on a on a map view where exactly I I can see Alex's property and I can see on the map we
[23:41] have 5G network and 4G 4G network covering 95 and 19% respectively. I can you can see the stories. You can you can see the property attributes, how when it is built, the condition. Exterior wall,
[23:53] it is very important for the networks and that is what we have been showing and this is all AI who gathered out of my 700 fields, relevant fields which will be important for this agent to talk
[24:05] to the human and satisfy that human. Then uh Alex is not ready to do it to to to accept my proposal. So in that case, what we have done is we the agent will show you, "How about the fiber? I have a
[24:17] nearest fiber in 100 ft. Uh what about instead of the 5G box, I give you a fiber line. I have two options. I'll give you the the nearest one and that will solve your problem."
[24:30] Right? So this all happening at the agentic level, multiple agents working together to solve Alex's problem. Uh whereas today, it's been a cycle of technicians going back and forth. This
[24:43] is another one where the agent will try to find, "Okay, maybe there was a hail storm last night." So we check the hail storm. Nothing was in there. So we know the answer that it was because of the overlaying overlaying
[24:56] spatial analytics and that's the reason Alex was not able to get the speed and um the service which she asked for. Um now, you have seen a social truth
[25:08] table, and then you have seen uh a demo for for my agents. So, this is how we build this source of truth table, where we took uh data sources, and it can be a cloud
[25:20] storage, it can be flat files, it can be shape files, anything. That is where everything starts, and then we will classify the data. As I mentioned, we are using the advanced NERs. So, we will
[25:33] um take an input address, and then bifurcate into names, phone number, email, organization, addresses, and from there we will call an appropriate Precisely service.
[25:47] Uh names will go to the name parsing, email will go to the email verification. And addresses, obviously, will go to the addresses uh APIs, which is address geocode and address verification.
[26:00] From there, uh we will have a quality check. And if the quality is not appropriate, then I will run my agent at the batch, and I'll ask my agent that this quality can still
[26:13] be improved. Go ahead, run the multi multi agent, and run it until you get up an appropriate response from that from an address. I can give an example. 123
[26:26] Main Street, care of Myank Asturia, New York, New York, 1001. Now, this address may not get resolved in the first call of AP geocode API or verify API.
[26:39] This agent will make sure that it will understand that maybe care of Myank Asturia is something which is making a problem here. So, how about I take it out, and then run it again? So, that is
[26:51] the rule of address fixer agent. Um and it will also give you a confidence tier and based on that, you can accept it or reject it. And finally, the last layer which you have seen that we have 700
[27:04] attributes which I attached in an example, that is where the advanced uh enrichment agent comes into the picture. Uh now, thanks to AI, if you are specify your verticals, specify your use case,
[27:17] uh using precisely catalog of data, you take uh the catalog the AI will give you suggestions on which attributes to use. And then based on that, you can also create flags. And um moving forward, you
[27:31] those flags will be helpful to the agents also. So, that is how we created the source of truth. Now, how like why why a customer why why my users will accept this kind of
[27:43] architecture? Is because the bottom line. If you see, Databricks help us to provide the governance because AI as a cloud API, we all use this for our
[27:56] personal use, but when it comes to running it in production, we all have our doubts. So, Databricks come into the picture, give you a surety that your data is not leaving. You use the agent
[28:09] one agent, multiple agent, all those will be highly governed and highly secure. Not no data will be trained no AI will be trained based on your data. So, that satisfaction, that level of security is
[28:24] been provided by platforms like Databricks. And then, we have audit logs, we have retry and recovery logics, observ- observability, which means you have a batch file going on and based
[28:38] on that, how many times there has been change in the data, how many updates are there, all these will be automatically generated within Databricks notebook. So, that's for the batch, and here is
[28:51] the architecture of how Alex's problem been solved using the Databricks Genie and multiple agents, where Alex
[29:03] point number one interacted and directly reached to a multi-orchestration agent, which will then call the NER model, and Alex's her
[29:15] phone number, her address, her request, or her uh >> Right. >> Yeah.
[29:30] Alex's name, Alex's phone number, verified Alex's phone number, and then taking the address, running those addresses, and then understanding that the problem is at the spatial level, where we have this 5G
[29:43] service in our in our docs or in our data, but why we not able to provide it? Then we move next point, which is the enrichment one. We looked into the enriched data, and
[29:55] then that agent the enriched agent will classify that the problem is based on the spatial computation. Enrichment will done on the fly, and it would categorize the location and that particular
[30:08] location into 5G and 4G. And then finally, we have seen the root cause has been explained by the agent. The the next steps has been explained by the agent, and those thing will be taken
[30:21] care of take care by the orchestration agent, which will be running these three agents entirely in the Databricks, and running the show with the governance, audit logs, and and observability. Thank
[30:33] 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.