Can AI Agents Be Trusted With Enterprise Data? | Big Data LDN Special

Two different teams in an enterprise built two AI agents, but both asked the AI the same question: What are last quarter’s customer stats? Both agents answer promptly and quite assertively. However, the figures on the same inquiry are different. The agents’ answers don’t match. Nothing has broken because neither team is aware of the other’s AI agent. 

What happened here is that each agent drew the answer from different contextual data. The business holds two variations of truth delivered with equal confidence. 

This is the kind of oblivious failure Angelita Frozza Sanches, Head of Data at Scout24, is working to prevent at the leading German real estate platform where AI agents have already been deployed in production. 

In this episode of the Don't Panic It's Just Data podcast, recorded live at Big Data London 2026 in Kensington Olympia, host Mike Ferguson, Managing Director of Intelligent Business Strategies and Conference Chair of Big Data London, sits down with Sanches and Suresh Srinivas, CEO and Co-Founder of Collate and OpenMetadata. 

They talk about who now consumes enterprise data, how much of the context layer should be centralised, and why enterprises making progress are the ones that start small and tie every agent to a business outcome.

According to Sanchez, the concerns from 20 years ago to now remain the same, and it comes back to governance. “How people consume, how they interact with data,” remains the same. What’s new is that “AI has changed the velocity and the impact of all those aspects.” 

“Now we can take advantage of AI to help us build governance,” Sanchez says. “How can we prepare everything not just for humans but agents–another persona in the game.”

Also Watch: How to Increase Agent Accuracy & Reduce Cost with Context

AI Agents To Be Primary Consumers

When asked about the comparison between before AI and after AI, Srinivas says that, as someone who has witnessed a transition from monolithic big data systems to distributed horizontal scaling in the age of Hadoop, which brought in commodity hardware. “I've also seen the shift from there to the cloud. Now, we are shifting to AI.”

“While those were technical changes, what I'm seeing right now is a fundamental shift,” he adds. “Agents are going to be the primary consumers now. We must transition from the people-ready foundations to AI-ready foundations.” 

Governance, documentation and access now have to work for agents as well as analysts.

Ferguson says that many enterprises have not grasped how much they will have to invest in context. It tends to arrive through new generations of tools they already own, and the realisation follows quickly that they need a shared understanding of data meaning and, harder still, of the relationships across data. 

“I'm not even convinced that they know about all the relationships themselves across the data. And so the realisation that there's a kind of missing piece to their stack hits them pretty quickly.”

Some firms are building this themselves while their existing vendors ship the same capability, leaving a build-or-adopt question many have not answered. 

The Chair of Big Data LDN explains that technology vendors provide distributed setups, leading to isolated silos of context. Enterprises then face the challenge of unifying these disparate systems. 

“Should they use a federated layer to connect everything, or adopt a master- model to synchronise from a central source?” He voices, “Currently, the master- mindset dominates because context management is not yet mature enough for true federation.”

What’s the End-User Business Strategy For Context?

Scout24’s business strategy is to focus on centralised policies, governance and tooling, with context contributed by subject matter experts. Sanches adds that the model has been revised in practice.

The team’s design principle comes from how people behave under load. "The more burden you put on the people that you are depending on, the less they will be willing to contribute, because their primary activity or function is somewhere else,” she says.

Scout24 runs a central data governance team, a centralised context platform and centralised policies. Automations create, maintain, alert on, and tag context, and call on humans only when judgment is needed.

The change had an effect on people's behaviour; in the case of the core domain, when they used the agents, the confidence scores for the answers were clearly improved. Sanches states that the contributors then realised the connection between their input and the output.

She firmly believes that confidence scores should be treated as an obligation just as much as a feature. "People have a right to transparency," she says, "and users need that information in order to decide whether to rely on an AI's outcome."

Srinivas proposed a principle to guide this approach: "Don't tackle a problem that you don't actually have." He cited examples of ontology projects that take six months to define all the elements before any AI work begins. 

In contrast, his suggestion is to identify a business problem, define the ontology required by that problem, and then expand the ontology as new requirements emerge.

Metadata, Meaning & Memory Supplied 59% to 94% Accuracy

Srinivas traced how OpenMetadata moved from serving people to serving agents. The first version built a unified graph of metadata so humans could understand data, judge its quality, and see issues. Business meaning came from the people using it. Once Collate began building its own agents, that stopped being enough.

“Before context became a buzzword, about four-five years ago, one learning from our experience in building data platforms, we realised that context of data is important for people to understand the data, what to use, whether it is high quality or not, and whether there are any issues. That's why we built a metadata platform,” Suresh said.

The metadata platform provides complete context of the data for different users as they require it for different purposes. This is what led to the building of a unified graph of metadata by Collate, 

Srinivas tells Sanches and Ferguson that metadata helps people understand their data, but people have to bring their business understanding, the business context, and the processes. This includes understanding for domain experts and business users with distinct personas. 

OpenMetadata's first iteration was building context for people. When AI came in and the company began building its own AI agents, they realised that giving AI the understanding of data is not enough.

“What data exists. It also needs to understand the business context and the meaning. People interacting with AI will ask business questions. That means you have to have the understanding of the business meaning of what they're asking and map it back to data, and data becomes an implementation detail,” he explains. 

The context that Collate built for users with metadata expanded to include metadata, meaning, and memory. Memory covers the corrections, confidence scores, and up-voting and down-voting that let learning carry across agents instead of staying inside one. 

He cited a benchmark in which the best available LLMs reach 59 per cent accuracy without that context. With metadata, meaning, and memory supplied, the figure rose to 94 per cent, and for frequently asked questions, further feedback can take it to 100 per cent.

Organise first, then build

Ferguson's advice to enterprises is to start away from technology at first. He backs a strategy involving a programme office that ties each agent, data product, and machine learning model to a stated business outcome, kept visible on something like a large Kanban board of what is built, in progress, and still to do. 

The Managing Director of Intelligent Business Strategies believes that this coordination is what lets democratised development work. Teams know which piece they own and why.

From there he favours divide and conquer. Rather than a monolithic enterprise context layer, he asks which knowledge bases (the term many of his clients now use for vector stores) each agent and part of the business needs. Some clients are moving toward what he calls mass contribution to common outcomes, with humans and agents both feeding the same goals.

Srinivas added the people factor. In his experience, the customers progressing fastest have a motivated, empowered champion, and the biggest missing piece elsewhere is exactly that person. Sanches agreed that a data leader needs technical grounding and a sense of which work earns business confidence first.

Ultimately, AI agents are now becoming the main consumers of enterprise data, and they answer only as well as the context they are given. Scout24's approach is to centralise guardrails and the context, let experts contribute where possible for them to do so, and show confidence scores so trust is earned. Ferguson's counterweight is that none of it matters unless each agent is anchored to a business outcome.

"Do not run before you can walk. In other words, plan this out. Divide and conquer and figure out what you're going to build where and for what purpose,” the director said.

Sanches says, "AI multiplies what the context represents. There is a lot of FOMO involved," she said, with people building to avoid being left out of the cool crowd. Ferguson replies that you don't want to "do bad things faster."

Key Takeaways

• Agents are becoming the primary consumers of enterprise data.
• Existing data foundations were built for people, not agents.
• Scout24 centralises policy, governance and tooling, while experts contribute context.
• Cross-domain "core data" has no owner, so Scout24's central team curates it.
• Visible confidence scores made domain experts willing to contribute.
• Without context, the best models scored 59% accuracy on one benchmark.
• With metadata, meaning, and memory added, that benchmark result reached 94%.
• Don't build a full ontology before starting; define it per business problem.
• Duplicate agents with different context produce confident, conflicting answers.
• Poor context drives up cost as agents retry and guess.
• Context layers are intellectual property, which raises European sovereignty concerns.
• Tie every agent and knowledge base to a stated business outcome.