
From AI Pilot to Payoff: Three AI Wins You Can Replicate
Part 3 of The Agentic Advantage. Data leaders from XMA and JPMorgan Chase (speaking in a personal capacity), plus Thomas Mazzaferro, share how to turn AI pilots into measurable results, followed by a Maia context engine demo.

Many business leaders have more AI projects than they can staff, and haven't seen ROI from the ones they've run. In part three of The Agentic Advantage, Joe Herbert, Principal Solution Architect at Matillion, talks with three data leaders about what it takes to turn an AI pilot into measurable business value.
Michael Kipling, Head of Data Intelligence at XMA, shares how his team is using Maia and Snowflake to replace a product information management data flow, with an estimated 80% cost saving and three fewer software platforms. Suma Nair, Executive Director at JPMorgan Chase, speaking in a personal capacity, explains how to choose, measure and responsibly scale AI use cases in a regulated environment. Thomas Mazzaferro, 4x Top 100 CDO, explains why context is the constraint on AI ROI.
You'll hear:
- How XMA is replacing a product information management data flow with Maia and Snowflake, including a semi-structured product dataset that would otherwise have needed a table 18,000 columns wide.
- Why to start with a small, well-defined, lower-risk pilot, keep a human in the loop, and measure impact to earn the budget to expand.
- Why context, through semantic layers and ontologies, now decides AI ROI, and why to optimize end-to-end processes rather than single pipelines.
- How agentic data ops can shorten recovery time when overnight pipelines fail.
The session closes with a live demo of Maia's context engine for a finance and retail operations team: an ontology built from more than 150 Snowflake schemas, crawlers that pull in sources, and Jira tickets hydrated with business context.
Joe Herbert: Hello, and welcome, everyone, and thank you for joining the third and final webinar in our Agentic Advantage series. I'm Joe Herbert, principal solution architect at Matillion, and I'll be your host today.
If you have any questions during the webinar, please submit them through the Q&A button. We'll answer as many as we can at the end, and we'll share the recording afterwards.
For those of you who are new to Matillion or Maia, it's our agentic data automation platform. We help enterprises do autonomous data engineering, across financial services, which we're speaking to today, as well as healthcare, life sciences, and many other verticals, both here in the UK and in the US. We do this through push-down ELT into your cloud data warehouse of choice, whether that's Snowflake, Databricks, Redshift, or BigQuery, using spec-driven development and autonomous data engineering best practices.
Our team of data agents helps create contextually aware data pipelines that preserve the quality, trust, and accuracy of your human data engineering teams. They augment those teams as you modernize your data engineering work across the enterprise.
We're joined by one of our customers, who's going to talk about how they've adopted our agentic data engineering team, Maia, and the success they've had. Stay tuned at the end, too, for a demonstration of our context engine, which helps Maia speak the language of your business.
Today's focus, though, is on three data leaders and the AI pilots they've run, or are thinking about running, in their organizations. First, we're joined by Suma, Michael, and Thomas. Suma, we'll come to you first. Could you introduce yourself, tell us who you work for, and what excites you today?
Suma Nair: Thank you, Joe. Hi, everyone. I'm Suma Nair, executive director at JPMorgan Chase, and I'm head of strategy, data, and AI for the business of the data center. I can't go into the details of my internal work, but I'm really excited to talk about how we responsibly use AI for use cases that simplify complex processes, and how we use it at scale to improve efficiency. I'll be speaking in a personal capacity today.
Joe Herbert: Fantastic. Thank you so much, Suma, for joining us. Michael, over to you.
Michael Kipling: I'm Michael Kipling, head of data intelligence at XMA. I've been with the business for about thirteen years. I started as an analyst and worked my way up. We've been building up our data team for the last four years with Maia and Snowflake as our partners, getting more mature and moving from a purely analytics support function to operating core functions of the business, through data integrations with Snowflake and Maia at the back end.
Joe Herbert: It's great to hear that. We'll hear more later about how Maia gives you and XMA the freedom to do more. And last, but by no means least, Thomas, would you mind introducing yourself?
Thomas Mazzaferro: Sure. I'm Thomas Mazzaferro. I spent over twenty years in banking and financial services. Now I have the pleasure of partnering with a number of the top institutions in the world, helping them drive transformation with AI and evolve their businesses for the AI era. Great to meet you and the team.
Joe Herbert: Fantastic. Thank you all for taking the time. I'm really excited to get into the AI wins our audience can replicate or draw inspiration from in today's session. First, we're going to hear from one of our customers. Michael, I'm going to bring up some slides to walk through. Could you give us an overview of what XMA does, what AI initiatives you're showing today, and what business problem made this a priority for you and the team?
Michael Kipling: XMA is an IT reseller. We've existed for over forty years, and we sell all sorts of IT solutions and the services that go alongside them. So embracing new technology is very much on brand for us.
We were closely linked to our parent company, West Coast. Essentially, their IT did our IT. About four or five years ago, in 2022, we split out from them and needed to set up our own IT function, including the data function. We had one analyst who looked like he might have the aptitude to cross over into data engineering. Me and Anthony set up the team and built it out. We got help from a consultancy, SDG, to get us up and running: selecting technology, getting upskilled, and running an apprenticeship program to grow the team. We brought on apprentices, including Malachi, who's now been qualified for some time and is helping us build out more of these data solutions.
As for the problems we've had: to do what we did, we had to spin up an ERP system from scratch. That meant we had to learn that system, and we had to provide all the reporting and analytics off it, quickly. Matillion ETL, as it was at the time, and Snowflake were our partners, and that's how we built it out. Even though those products were more of a traditional ETL approach at the time, being cloud-based meant we could get up and running quickly, with quite a green team.
What we've discovered over the last year or so of the ERP being live is that some of the solutions we put in place aren't the most efficient, in terms of both performance and cost. That's the problem that's been put to us. We have Maia and Snowflake, which are very efficient and very powerful, so can we start using them to fix some of the problems with our integrations?
One of the things we're looking at right now, and we're quite deep into development, is our product information management data flow. At the moment, that requires a business application to manage the data. It needs its own ETL and its own integration to push the data out to the end platforms. We're in the process of replacing all of that with pipelines designed and supported by Maia, into Snowflake, using the medallion architecture you'd expect, with gold data products that present the data in exactly the right way for all those business applications.
Joe Herbert: So there was a level of modernization here. You had a solution that already worked in the old way of doing things, but the new challenge needed a new approach and a new tool set. How did you decide AI was the right approach for this problem, and what did you do tactically to get it moving?
Michael Kipling: As I said, we're quite a green team, so we knew we needed to modernize not just the technology but our approach to data. If AI has all the repositories of information about best practices and the latest approaches to handling and architecting data, then we have that as our support. We already knew Maia could provide that kind of support and troubleshooting when you're in the realm of development. It's always there to assist you.
We'd never built out business processes and that sort of architecture before. Our end product had always been analytics, so the whole thing was new to us, and we knew we needed that extra support. The business also brought in a data architect on consultancy for the really tactical stuff, like how we were going to approach this. So when we looked at Maia, it was almost obvious: we already had that in place. It's what's going to help us get to production quickly, and it has.
One really specific head-scratcher was the semi-structured data we were getting for product. The range of products we sell is quite wide, so you can imagine the permutations of data are huge. The specifications for a Wi-Fi router are massively different from those for a computer monitor, and so on. How were we going to structure that? Our understanding was that it would come from tables of data, and you're talking about potentially a table that's 18,000 columns wide, which is preposterous. But Maia helped break that down into a really elegant solution for dealing with that semi-structured data. When we came up with it, our data architect called it "cheeky."
Joe Herbert: Brilliant. Eighteen thousand columns. That metadata is so rich with tribal knowledge and context that the XMA data people will understand, in terms of what each of those columns means, how it moves through the medallion architecture, and how it ends up in the downstream reporting and in the data products your team supports. What I heard is that the productivity gain in getting to that end state outweighed the trade-off against other competing priorities.
Now that you've adopted a more agentic way of building your data products internally, what's the impact on your business? What do you expect, and what would you tell someone else starting out about how to approach an agentic way of working?
Michael Kipling: Within my team, the advantages were immediate. We can spin up solutions much quicker. Maia takes on the repetitive tasks, like the huge views we'd otherwise have to build out and code line by line, and the repetitive pipeline development. So our route to a solution is much faster.
The real end result, once we finish, is this. We're currently in UAT, by the way, so we're very close to getting this out into the real world. We could have three fewer software platforms. That simplification is fantastic. We're all focused on one set of products, we know how it works, and we can all troubleshoot the same platforms with the same knowledge and skills. It also brings reduced cost: three fewer platforms and the professional services that go with them. We're estimating about an 80% cost saving versus what we spend today.
Joe Herbert: Wow.
Michael Kipling: We're also seeing that some of the integrations and the ETL in the existing platform are quite niche. With Maia and Snowflake, given their scale and the amount of data they can deal with, and how modern they are, performance is much improved and we can handle much more throughput. Sometimes a big product launch can grind some of those integrations to a halt. We've tested that with the initial product load, which was 780,000 products. We know now that there's nothing we can throw at this that's going to break it. It's going to stand up and perform.
Joe Herbert: Fantastic. You also mentioned to me last time we spoke that it doesn't stop there. There's the monitoring and observability angle, and the ability to make change requests in an agentic way. How does that affect the speed of changes to these data engineering pipelines, and what does it mean for freeing up your team's time?
Michael Kipling: We have so many pipelines that are quite repetitive. Think of the pipelines for ingesting supplier and cost feeds: you have to repeat those for every single supplier. If we want to improve the efficiency of a pipeline, we can do it in one, experiment with it, see if it works, and then tell Maia to deploy that across every pipeline. We know it will be consistent, and we can deploy at speed.
On monitoring, as I've moved into a leadership role, my data skills have gotten a bit rusty. Now, with Maia, I can go in and diagnose what the problems are, and we can organize the support to get things up and running much faster. If there's ever been any downtime, the business rarely notices, because we're back up and running so quickly.
Joe Herbert: Fantastic. Michael, thank you for sharing that story. I'm sure there are plenty of questions in the chat that we'll come back to shortly. Now I'm excited to be joined by Suma as well. Suma, how can large, highly regulated financial institutions identify where AI can help solve business problems, and what does it mean to adopt it responsibly? Perhaps you could speak to that in a personal capacity.
Suma Nair: Absolutely. Thank you, Joe, and that was a really good description, Michael. I enjoyed listening to it.
Large financial institutions have lots of complex processes, with multiple teams across business units working through different processes in a highly regulated environment. That puts privacy, legal obligations, and security controls front and center. Bringing AI in will improve efficiency and save time, but you can't do it all at once. You can't boil the ocean.
So we have to find the highly frictional processes that are well defined, pull them out of the ocean, and define them properly. Then we reimagine the process to find pockets where AI can make things faster, better, and easier, and eventually free up time for the people doing that work today. And we always have to keep in mind how we'll measure the improvement. Is it processing time? Accuracy? Issue resolution time? Figure out what the process is, then figure out how to bring AI into it in a very responsible way.
If you want to move to the next slide: some of what Michael talked about is also what I have here. In these large institutions, we identify the good enterprise sources to use in the workflow we've chosen. We use those initial data sources and bring them into a trusted data layer. As Michael described, that can be a medallion architecture, where you land raw data in a bronze environment, and then in silver and gold you add clear definitions and the right context, along with thresholds and guardrails that reduce hallucinations and other inconsistencies.
That becomes really important in a highly regulated environment, because you have to maintain accuracy while also following all the regulations and controls that go hand in hand with it. Context becomes extremely important. This is where an ontology and a semantic layer come in, and how exactly you'll use them: are you using a RAG-based model? All of that matters when you set it up, and it depends on your use case.
Joe Herbert: You mentioned use cases, and there's almost a paralysis of choice in which ones to go after and how to value and assess them. How can organizations measure whether an AI use case is delivering the business impact it claims? How would you think about ranking and prioritizing these different pilots?
Suma Nair: Absolutely. Let me give you a couple of examples. A basic one: think of a financial analyst working on many different requests that come from senior management, or from regulatory requirements they have to fill out. The work they do today is very manual. They go to all these different sources, put the information together, do the analysis, identify the insights, and create a report, a graph, or even a presentation to hand back.
Think of that as an end-to-end process. What data sources do they use? How do they curate that information, and how do they finally present it? If you lay that out, it might be as simple as creating an agent that looks through the analyst's emails or Jira tickets, however the request comes in, and gathers all of that. Then it figures out which sources to go to, whether external or the company's internal sources, and compares the results back against what the request was.
You could do this as a deterministic workflow. The agentic part, the way AI enhances it, is the reasoning capacity. It can work with both structured and unstructured data, bring it together, gather the insights, and create the presentation and graphs for you. It pulls it all together, puts it into an email, and keeps it ready for the analyst to review end to end. The analyst does the accuracy checks and makes sure all the controls were followed, that the AI isn't going where it isn't allowed to, and that it's scrubbing out secure information such as API keys.
Imagine the time the analyst saves. Their time goes into more strategic work: innovation and what the next step is, rather than repeating the previous steps. It becomes a really powerful companion.
Joe Herbert: What you're describing feels a lot like context engineering. You're hydrating a task or use case that's sitting on a backlog, on a Jira ticket somewhere or just in someone's mind, and fleshing it out into something more substantial, with more hydrated information. That makes it a more trusted, higher-quality, accurate data product, and it frees up the human brain to do higher-value work, like detailed financial analysis, and get to meaningful impact. Could you share some other example pilots you'd think about, Suma?
Suma Nair: Yes. That was an individual's journey. But think of a large enterprise application where, as we talked about, you're curating data from multiple sources, bringing it together in a medallion model, and eventually making it available for AI analysis. It could be a chatbot, or automatically generated dashboards, or something in Snowflake or other tools, as Michael mentioned.
The whole process of loading the data, running the pipelines, and creating the data in the required data product structures is a long pipeline process with multiple steps. It usually runs overnight so the data is available when the analysts are back in the office in the morning. But what if something fails at night? Now you're calling a production support person who has to come in and troubleshoot. What happened? Was a system down, or was a threshold not met? There could be many reasons, and the troubleshooting takes time. The analyst is in the morning, and the system hasn't run.
In this case, too, you can put together deterministic approaches: the pipeline fails, create a ticket, and rerun it to see if it works. But you can also add an agentic layer to that step-by-step process, to troubleshoot and figure out what caused the failure. Was it a system failure, or a larger outage? Was it an individual job that failed because of a memory error or a storage problem? Or did it need a quick outage to be taken care of? The system can do the entire troubleshooting process in the background before even calling the production support person. When that person comes in, they can see where the errors happened and which part needs to be reset, and then restart the process. That's really easy to put in place.
It could also do the accuracy checks before the analyst is in. Usually the analyst comes in and checks a few datasets to make sure things match up before going back to senior leaders. Now the system itself does the accuracy checks and checks the thresholds. If it finds an issue, it goes back, does further checks, and brings back its analysis. So when the analyst comes in, they already know what the discrepancies were, why they happened, and what steps they can take. There are so many possibilities, and it's left to our imagination to figure them out.
Joe Herbert: That's fascinating. What you're describing at the end is the context engineering we were talking about earlier, in Michael's example from a really wide range of data sources. But it's also about mean time to recovery when pipelines fail, which you want to reduce, with agentic data ops in the middle of that process to make it faster.
Another source of that context is the pipelines themselves: the metadata they contain. That can be used to turn something that breaks in the middle of the night into a well-fleshed-out, hydrated fix, ready for the engineer to pick up in the morning.
For teams that are just at the start of that agentic data ops journey, and who may be thinking this is a wonderful future state, with pipelines almost self-healing, but surely impossible, what would your advice be, Suma?
Suma Nair: Start small. We used to say crawl, walk, run. Find a simple process first. You always have to build trust before you can expand any kind of pilot. Choose a very fixed process that you can describe from start to finish, and figure out which pieces of it you can re-engineer to bring AI in, based on how your ecosystem is set up. Each new use case depends on the existing process, the tools available, and your end goal.
Definitely keep controls top of mind. Human in the loop is the way to start, until you gain trust in the system and feel you can allow it to make decisions on your behalf. Keep a log so you can always come back and check it, and always keep people in control at key points along the process. Then measure the impact, because that's how you tell your story and how you get budget to expand your pilot.
If you keep those three things in mind, start small, work your way up, and measure and tell your story with the right metrics, there's a large success story that anyone can have.
Joe Herbert: Suma, thank you so much for sharing that. I know we've already seen some questions come in for you, which we'll come to shortly. My last guest today is Thomas, who has spent twenty years in financial services and banking and now partners with companies to accelerate their ability to build and deploy solutions. Can you give some examples of where you've seen success, and solutions that work and drive real ROI, across companies and industries?
Thomas Mazzaferro: For sure. Very happy to. I heard a few themes over the last twenty minutes or so, and I'll weave them in as I go.
The first thing I'll add to what Suma said is that people are learning. Technology is advancing so rapidly that how we use it is changing almost monthly. Continuing to understand how it can be applied to different use cases is a really important skill to keep building.
You heard from Suma already, which was fabulous. The reason I started there is that for the last six months you've been hearing in the market about context. This is context: semantic layers, ontologies. I always ask why. Why is that such a hot topic today? Semantic layers and ontologies have been around for almost twenty years. It's not a new topic. So why is it now top of mind?
The reason is that we now have technology advanced enough to reason, but it can only reason about what you inform it with and how you tell it to use that information. That's why business context and ontology matter so much now. This technology has become advanced enough to work at a pace and speed faster than humans, but without the right business context and intelligence, it doesn't know how to unlock that value.
Which brings us to Maia, and what you've enabled here: the ability to bring that context into your pipelines and unlock their throughput, so a human is no longer the constraint on how fast you can build and accelerate your business. You can do it at a much faster pace and unlock the corresponding value.
I've worked with many clients across many industries, and the first biggest problem I hear about is context. They don't understand what the context of their business is, and they say, "Give me the easy button. How do I actually get context and make it real for my company?" That's piece one. In the past, you had to do what was mentioned earlier, build a medallion architecture, and that takes quarters and years. It's not an overnight thing. So how you accelerate that piece of the puzzle is really important.
The second piece is, once you do that, how do you unlock it? When you talk to business leaders today, they say, "We have so many AI projects that we can't get to them all, and we haven't seen ROI from them." Why do you have those two constraints? The first constraint is people. The reflex is to staff up and hire people to unblock the demand. That's the wrong approach in the AI era. You're hiring humans to manage the robots. That doesn't make sense. You should be enabling humans to unlock the robots, and the AI, to go faster. That's what Maia does. It lets you unlock capacity and throughput using agents, rather than being human-constrained.
On ROI, why do people struggle to get it? In my view, it's three things. One, they can't get access to their data. Two, they can't build solutions they can trust, because the output is probabilistic in nature rather than deterministic. How do you make sure the output is deterministic? And three, what are you trying to achieve? Insights and analytics are great, but giving you ideas and visibility into what's happening doesn't change the business. How do you start enabling changes to processes and client experiences, optimizing process flows, unlocking throughput, and getting the visibility to make a change or evolve your business? That's where you start unlocking true business ROI and value. Those are a lot of the constraints and blockers we're tackling today.
You heard from Michael and Suma already. What Michael mentioned about the transformation he's doing is great. It's helping teams go faster and accelerate. I also work with clients across other industries. One of them said, "We want to connect all of our marketing systems. We have thirty of them: Google Analytics, Sprout, Salesforce, and so on. We have so many disparate systems that we don't understand the marketing lifecycle in a seamless view." What's happening with the campaigns? With customer interactions? With conversion? When we launch a product, how do you view it all the way through conversion and financial management, and see that come together?
People think of ontologies and say, "We'll just do an ontology for one domain." That's the wrong way of doing it. You want ontologies across the entire business process and department, with data that understands how it connects, so you can optimize the entire department or channel. I want people listening to this webinar to think beyond optimizing one pipeline. That's what we've done for the last twenty years. We're in a world that's changing. Think about optimizing an end-to-end process, or an end-to-end sub-department or department. How do you use these new solutions to unlock potential beyond a single pipeline, for the department, or for processes that cross departments, to help accelerate the overall business?
Joe Herbert: You touched on a few really interesting points there, Thomas. There's the tribal knowledge in organizations around how processes were set up and what they cover. It takes eighteen months or so for new joiners to learn and understand those processes. What I'm hearing you say is that this agentic data team gives us a capability not only to step into one part of that process with automation to make it faster, but to step back and treat both the process and the connective tissue of the organization as a source of data context. That can unblock the productivity of team members regardless of who owns the process. Is that a fair summary? And I don't want to cut you off from elaborating. How do you think about AI fitting in from a broader, holistic perspective?
Thomas Mazzaferro: I think you hit the nail on the head. That's a great synthesis. You hit on self-healing earlier. You run batch data overnight so that things are ready in the morning, but that batch data is feeding business processes. So the data and the business are ready in the morning. It's not just self-healing pipelines. It's self-healing business processes. That's where we should be going.
Ask business leaders, "How do you run your business?" Do you know what they don't say? "My business runs through operations," or "My business runs through technology," or "My business runs through Suma." They're not saying it's a person or one department. The common answer is that the business runs on business processes. Those processes are optimized to serve a client, meet a regulatory requirement, enable a marketing campaign, or deliver an outcome. That's how businesses have been built over the last hundred and fifty years.
So, to your point, it's like self-healing pipelines, but for me it's one level higher. What we're doing now is building self-healing business processes. We're building the assembly line to create parts of the process agentically, and then to triage, troubleshoot, update, and maintain those processes across the business.
Joe Herbert: You've been involved in some of these use cases in the organizations you work with. Could you paint a picture of how to approach that end-to-end view of an AI pilot with a use case in mind? Not only the assembly line creation that helps people build more, but also how you unlock and understand the business value those processes underpin.
Thomas Mazzaferro: Let me give you two answers. The first is this: if you wanted to build a new solution, or launch a new product or channel, what would you do? For the last hundred years, in any business, you'd hire an SME, an expert in that area. You'd say, "I want to hire you to build out this business, this solution, this product, this capability." They'd come in, spend a month and a half getting up to speed, and start doing their work. Nine to twelve months in, they have an MVP. You've spent a pretty penny hiring an industry expert, and you've had to give them a team to support them. In today's world, that's changed.
If you already have that knowledge but haven't pulled it together, that's what we call tribal knowledge. For the last hundred years, you've had to talk to your SMEs and either interview them or have them write business requirement documents hundreds of pages long, then hand that to a technology team to try to understand. Imagine a world where you can plug in solutions that identify and create context from that tribal knowledge. How? The tool isn't talking to that person. But it can plug into what that person has done and built: the PowerPoints they created, the Word documents, the transcripts of their meetings, their messages on Slack and Teams. It can start putting together what that person's tribal knowledge actually is. You've then brought together knowledge that you can unlock and massively accelerate as you go forward.
So how do you start bringing together structured and unstructured data? Not to make it conversational. It's the opposite. How do you create contextual knowledge that enables employees to become two, four, eight times more productive every day? That's the unlock.
Now you can take someone who has worked at JPMorgan Chase for five or ten years and built up that tribal knowledge of their area, and empower them with a true copilot in their workspace. You unlock that value as they go forward. That's a different way of thinking. You're no longer just giving it a RAG to search. That was AI two years ago. Now the question is how you enable that context to unlock value. And I'm just using that as an example. I don't know whether they're doing this or not. But that's exactly what we're doing with the clients we work with: plugging in and unlocking visibility and transparency.
To your point, what Maia does is let you unlock that from a data standpoint, for pipelines across enterprise sources. You bring it together, which lets you move data and make it accessible for those different solutions. That's the unlock: bringing together tribal knowledge and social data to really empower the business forward.
Joe Herbert: I love that. The AI pilot you'd recommend others start thinking about today is how to get the context engineered the right way across the organization, so you can then unlock even more capabilities with agentic AI. Thank you, Thomas, for sharing those thoughts.
We've had some questions in the chat. On your last point, Thomas, someone asked how to take the tribal-knowledge SMEs in organizations and have them contribute to, and seed, different context engines and knowledge graphs. I'll stick around at the end and we'll do a demonstration of how Maia does this.
Another question: "I work at a large enterprise, similar in size to JPMorgan Chase, and security is a big concern that gets in the way of us experimenting and testing. What advice would you give on how to navigate this?" I'll go to you on this one, Suma.
Suma Nair: Like we talked about earlier, security and regulations are a huge part of large financial institutions. We always have to think about security as a design requirement. That means thinking about security and controls at the beginning.
Whenever you start a pilot, think of the lowest-risk one to start with. Make sure it's already been discussed with your control partners, and that you've talked through a low-risk option before implementing your plan. That helps reduce friction and lets you focus on the solution. Then, as we said, measure your metrics and show the impact. Once you've done that, you have a better chance of going after a higher-risk option. You show your impact, you get the right controls in place and de-risk up front, and then you continue from there.
Joe Herbert: We talk about agentic by design, but, Suma, that sounds like security by design, which I'm also a big fan of. Thank you for that. It's really great advice, and thank you to the listener for the question.
Another question, perhaps for you, Michael: "We've got loads of legacy tools across our organization, and it's a headache, especially for business reporting. It takes ages to fix things, and it eats up hours. Looking at the XMA use case, could Maia help us automate that troubleshooting?" We covered a little of this when we talked about mean time to recovery, but could you elaborate on how Maia helps XMA with agentic data ops?
Michael Kipling: The troubleshooting has been a massive boon for us. We have a lot of pipelines running from a lot of systems, and not just XMA's. We have subsidiaries we work across, so we ingest all of their data as well. There's a huge amount of data from a range of systems, some on modern clouds and some on-prem, so the opportunities for things to go wrong are varied. We also have our suppliers feeding us data, and sometimes they have what they think are really good ideas, like adding an extra column to a report, which then breaks the schema.
We're on a bit of a journey with Maia at the moment. First, we use it to interrogate failures. We ask what went wrong with a pipeline that failed, and it tells us and suggests a fix. We review the fix, tell it to go and do it, and then find everywhere else the same thing could go wrong and apply the fix there.
Now we're looking at something further. If some of our suppliers are sending us information that's different from before, we can get Maia to log that and feed it back to the suppliers. Otherwise, someone from the supply chain team has to go and tell the supplier they've made a mistake with their report. We'll just get Maia to do that work for us. It's making us so much more slick and efficient.
Then we can go through the log tables and ask, "What happened in the past thirty days?" We can say, "These are all the problems we did have, and they resulted in almost zero downtime." So yes, it's massively helping us with troubleshooting.
Joe Herbert: Amazing. Michael, thank you so much. And to all my guests, thank you for sharing your insights today.
If you want to stick around, we're going to run a quick demo of Maia to show the context engine. Thomas's point earlier was that context should be departmental, so this looks at a whole finance department's context layer and how you can interact with it in an agentic way. It helps enforce your standards across building pipelines, and it helps Maia understand your business logic and your organization's naming conventions. You can also visit maia.ai right now to read about some of our case studies of how teams have adopted autonomous data engineering.
Before we get into the demonstration, thank you, Suma, Thomas, and Michael, for sharing your insights and stories today. Please do stick around to watch the demo.
Suma Nair, Thomas Mazzaferro, and Michael Kipling: Thank you. It was a pleasure.
Joe Herbert: For those sticking around, I'm going to demonstrate Maia's context engine and its ability to speak the language of your business. I've set up an example finance and retail operations team, to Thomas's point about thinking of context in terms of departments, not just a single data product.
I'm going to share my screen. You can see it's quite busy. There's a lot going on, and that's kind of the point. This finance and retail operations team has over 150 different schemas inside its Snowflake environment, and Maia has pulled all of that information together and created this ontological layer. I'm going to simplify the view so we can start from the basics of the graph. Inside the Maia platform, I can toggle what I'm viewing and hover over a few of these entities.
Let's start with a fun entity: sales order. It's something we can all relate to. This is an ontological description of a sales order in an organization that uses SAP S/4HANA. Maia has pulled this from data inside the Snowflake environment, but as I'll show you in a second, we can also upload unstructured data and connect it to other data sources. The sales order entity isn't necessarily a particular column in a database. It's an entity in itself that connects to other business objects that make sense from a natural-language perspective, like invoice and customer.
We can see that connective tissue visually in the graph, and we can also have Maia help us navigate and understand it. I'll say something like, "Show me how customer, invoice, and sales order relate, and what data sources we have to hand to help us build a data model with these." Maia semantically searches against those keywords, and on the right-hand side you can see it identify which entities I might be talking about. It then performs that search across this knowledge graph in real time and shows the end user that subset, the breakdown of that connective tissue.
Crucially, it also helps bring to life the underlying tables and schemas that form part of the knowledge Maia has built up. If I now introduce the attributes and domains, you can see a domain of order management and a domain of accounts receivable. Those are different teams that intersect with this question. I can also see that the invoice entity has multiple attributes, I think 26 of them. These map one-to-one, or one-to-few, onto columns. That helps Maia speak the language of this particular finance and retail operations team.
When I'm interviewing Maia and doing that context hydration, I can say something like, "Tell me about the invoices, and use the IDs from that to build me a data product that serves this up in a particular way." I can interact with my context engine here in the UI or, in fact, at that design and hydration step as well.
If I go up a level, I can show you how we set up this context engine. We have what we call crawlers. They crawl through your Snowflake, Redshift, Databricks, or BigQuery environment and populate the engine based on the data in your cloud data warehouse. You can also manually upload documents, browse through unstructured data sources, and have code repositories crawled as well.
And to Michael and Suma's point, you can treat the pipelines themselves as a source of context. You can crawl through the lineage of your pipelines running inside the Maia Foundation, and through the past thirty days of their history and how they ran. So when you ask, "Can you go fix my invoice and billing pipeline?", Maia has that contextually aware knowledge, populated in the knowledge graph as well.
You can see the variety of source types we can connect and pull into this knowledge. To use it, I'm going through MCP, or Mission Control, as I'm doing here. I can select the project I'm running in, and I have some branches where I do a lot of this data discovery work. I'm running in my dev, test, or prod environment, and I've got my context engine automatically selected.
I might give it a prompt like: "Go and find the Jira tickets I need to work through. Use MCP to connect into my Atlassian and find all of my tickets on the FE demo Jira board. One of them is to do with the invoice data product we're building. Go and find, in my finance and retail team's context engine, the relevant information about invoicing and what accounts receivable actually means, and flesh out a plan for how we can take the existing Jira tickets and hydrate them with this additional information from our context engine. While you're at it, check whether any of these pipelines have failed in the past twenty-four hours, and whether there are any fixes on my backlog to do as well."
That's a full variety of tasks we can now kick off for Maia. I'll put it into plan mode, and we'll see in Mission Control that it goes through and performs that semantic search of the context engine, while also connecting to my backlog's Jira tickets. It brings all of that together so we can perform the hydration that Thomas and Suma were mentioning earlier. We take the knowledge of our pipelines, our context, and our process, and have our agentic data team handle the building, testing, validating, and verification loops you'd expect in your data engineering process.
If you'd like to learn more, to see in more detail how this builds out, or to see how we can crawl other data sources to create a more populated, higher-quality, trusted, and curated context engine for our agentic data team, Maia, then please get in touch. Thanks for joining us in this third part of the Agentic Advantage series. I hope you have a wonderful day. My name's Joe, and I'm from Matillion.
Featuring
Our latest resources
Data management






