Big Data LDN: Why MDM is Back on the C-Suite Agenda

A massive organisation in the pharmaceutical industry deploys about six to seven master data management (MDM) platforms in addition to a few customised solutions. After facing challenges with scattered systems and their tools, the pharma company has decided to consolidate them. 

The problem is that every drug, clinical trial, or product is named differently across the world. On top of that, if you had to prompt an AI agent to decipher the proportion of the business that comes from every product, it would have to read through several systems, with no agreement between them, and no way of knowing which one is right.

This is the challenge that LakeFusion, a master data management vendor, was tasked with solving for a major pharmaceutical firm. 

“They’re using every MDM under the sun, as well as some custom-built solutions, looking to consolidate all of those on Databricks,” said Cooper Kenny, Founding Sales and Alliances Manager at LakeFusion in a recent episode of the Don’t Panic! It’s Just Data podcast.

Built natively on Databricks, LakeFusion provided the pharmaceutical companies with a way to unify the view of how much business they’re actually doing. This includes spotting areas of weak links in the supply chain or areas where the business can be expanded for sales and create focus areas for marketing teams as well. 

“Ultimately, the goal is to help them consolidate, extrapolate, and assign relationships and context to their data to allow their data and their business teams to really run with all of these new AI,” Kenny adds. 

For CDOs, CIOs and data platform leaders, it is also why a category once handed to analysts as a punishment is back at the centre of AI planning.

In this episode of the Don’t Panic, It’s Just Data podcast, recorded live at Big Data London, Podcast Host, Producer and B2B Tech Journalist, Shubhangi Dua sits down with three guests: Sanjeev Mohan, Principal at SanjMo; Cooper Kenny, Founding Sales and Alliances Manager at LakeFusion; and Panos Athanasiou, Chief Field Technologist at Databricks. They explain why agents are unable to make up for poor master data, what it means to govern context rather than just data, and the reason why MDM has to cover the tools and data products that agents depend on.

Also Watch: CEOs Consolidating Spend For Modern Tech Stack | LakeFusion @BDL

How Has MDM Changed in the Age of Agentic AI?

Dua opens the conversation by throwing a market figure at the guests. According to Master Data Management Market Report 2026 published by The Business Research Company, the MDM market is worth $20 billion and is expected to reach $42 billion by 2030. The question in the industry now remains whether AI undermined the MDM industry or had, in fact, renewed it. 

Mohan, also a former Gartner analyst, believes that the discipline of MDM before the arrival of agents was still troublesome. He shares, “When I was at Gartner, if you didn’t like somebody, you assigned them the MDM topic to research. Because MDM projects inherently would overrun their budget, overrun the time constraints, and they wouldn’t deliver the expected results.”

However, his view has now changed. He explains that when a person looks at a flawed record can often tell when something is wrong but an agent can’t.

Simply put, “humans can somehow rationalise and figure out something is wrong. Agents cannot,” the former Gartner Research VP tells Dua. “And if most of the consumers would be agents, then the question is, how do we make sure that they are reasoning and making decisions on the right data? MDM is no longer something that was nice to have. It’s a must-have.”

Dua cites another stat from Gartner: bad data already costs enterprises about $12.9 million each year. This stat is also projected to increase exponentially when agents take action on it. For instance, in the recent case of Hugging Face, OpenAI’s rogue AI agent violated the governance rules in order to carry out its assigned task.

To this, Athanasiou from Databricks responds by alluding to the oldest rule in machine learning. “In machine learning, we used to say garbage in equals garbage out. If you feed bad data to your model or if you train with bad data, your model will be bad. The same goes for agents.”

Mohan referred to a cost that is seldom included in an AI business case. Poorly justified data increases the amount the agent has to read, and this in turn impacts spending.

“If you give a lot of context, even to a child, and ask a question, they get confused and hallucinate,” he says. “That’s what agents will do. The thing that is most dangerous is your token cost goes sky-high, and that can kill your entire AI initiative.” 

What’s the Role of Context in MDM?

The question put forward in the conversation is: how can enterprises administer something they can't see at all? The topic always came back to the context. In the past, enterprises controlled data since it was tangible and fed it to deterministic applications. Mohan explains that it always produced the same output. Probabilistic models alter that situation.

“We used to govern data, which is still tangible, and the consumers of data were these deterministic applications. If you did a good job at governing your data, then the output would always be the same,” he says. “Now we’ve got probabilistic models. And we’re not just governing data; we are governing context.

Context, Mohan explained, is generated at the moment a query runs, which makes it hard to inspect.

The Principal adds that context is a sort of ephemeral. It gets generated when queries are run. “Panos and Cooper and I can ask the same question and have different context. How do you govern something that you can’t even see?”

Individuals using the product at BDK were asked vendors to provide a definition of the term, context. "Each exhibitor now includes context," he pointed out, "and the end users are telling me, 'Will the correct context please come forward and tell us what to do?'"

Athanasiou also applied this issue to the agent lifecycle. In software engineering, there’s a well-defined process: develop in a notebook, push through CI/CD, and pass both the test and production gates. Agents tend to undermine that model.

The Databricks Chief Technologist then asks how one could control the agent since it can now do whatever it likes.

“We are going from deterministic software engineering, which had a very specific software development lifecycle methodology, to probabilistic systems. The agent can basically spatter out whatever it wants. How do you control that? Also, how do you control the tools that the agent is using? How do you know what’s the good tool, actually?” 

As a solution, he talks about Unity Catalog, a product that Databricks developed to manage not just tabular data but also unstructured data, dashboards, notebooks, and the tools that the agents use. He then passed the problem on to the MDM vendors.

Athanasiou asked whether MDM actually had to operate on all these different layers, since it couldn't be merely a data product but also a tool. Adding additional queries, he asked what the good version of such a tool would be for use by which agent?

Agents cannot tell when a record is wrong, so the quality of master data now sets the ceiling on what an agent can do. The panel agreed that the old model of MDM, a separate system that copies data out, needs a small army of specialists and takes most of a year, does not fit that need. Mohan and Kenny emphasised LLM-driven automation, while Athanasiou held that human operators and an organisational culture of data ownership remain part of the answer.

Watch the episode for a deeper understanding of MDM and how enterprises need to approach MDM in the age of agentic AI evolution. 

Key Takeaways

• MDM is forecast to grow from $20 billion to $42 billion by 2030.
• Humans spot wrong records. Agents don’t.
• Governance now has to cover context, not just data.
• Context is generated at query time, so it is hard to inspect.
• LLMs can match structured records to audio, PDFs, and social posts.
• Zero-copy architecture removes the need for CDC pipelines in MDM.
• Ontology needs semantics and connections, not only graphs.
• Human operators still decide what counts as a golden record.
• Bloated context drives token costs up and invites hallucination.
• Agentic data stewards can work through human-review backlogs.
• Start with a thin slice of value deliverable in weeks.

Chapters

• 00:00 Introduction to Big Data London and Participants
• 05:02 Master Data Management: Evolution and Trends
• 09:54 AI Native MDM: The Future of Data Management
• 15:07 The Role of Context in Data Governance
• 19:55 Challenges of Bad Data and Its Impact
• 24:36 Cultural and Technological Shifts in Data Management
• 25:00 The Importance of Master Data in AI Initiatives
• 25:24 The Future of Data Management Solutions
• 26:43 Lake Fusion's Unique Position in the Market
• 29:30 The Role of AI in Understanding Business Dynamics
• 32:06 The Evolution of AI Agents in Workflows
• 35:37 Key Takeaways for C-Suite Executives

For more information, visit lakefusion.ai.