From People to Agents: Why AI Demands a Different Data Foundation

In the age of AI, a growing number of organisations are beginning to realise what their real goldmine is: their own data. Not as a static raw material for a management report, but as the dynamic context that makes a company unique. Years of transactions, decisions, customer behaviour, and accumulated expertise form an asset that no competitor can copy. After all, the generic AI model is the same for everyone and can be bought anywhere; genuine differentiation only emerges once that model is fed with the organisation's own data. That is when contextual intelligence emerges, and that is where the true promise of AI lies.

That realisation has now hardened into a consensus echoing through many boardrooms: the added value of AI is directly determined by the quality of the data foundation. But the crucial question often goes unasked: who decides how good that quality is? Is it the human who has traditionally looked at it, or the AI agent that has to work with it?

Why companies are now investing in their data foundation

"Everyone now knows that it's all about the data. Almost no one asks which data, and for whom."

Research by MIT (2025) shows that roughly 95% of companies deploying generative AI are not yet achieving a measurable return. Strikingly, the cause does not lie with the large language models themselves, but with these systems' inability to retain company-specific context and adapt to it. The conclusion is inescapable: the bottleneck is not in the AI, but in the underlying business data.

The effect is already visible in how budgets are allocated. Spurred in part by binding regulation such as the NIS2 Directive and the AI Act, companies are seizing the moment to get their data foundation and data governance in order in one move. Investment that previously flowed to isolated AI experiments and standalone tools is now shifting to the basics: cleaning, centralising, and securing data.

Virtually every organisation is currently striving for the same ideal of "AI-ready" data: clean, structured, housed in a modern architecture, and equipped with tight governance. There is nothing wrong with those qualities, but they are the product of decades in which data was prepared exclusively for analysis and reporting by humans. Now that humans are no longer the only readers, we have to confront a new reality: are we in fact preparing this data for an employee with a dashboard, or for a machine that has to interact with it?

AI-ready data is (for now) still defined for human readers

What suffices for a dashboard is not yet equipped for an AI agent. The current data and IT landscape rests, broadly speaking, on two pillars: the lightning-fast processing of operational transactions (OLTP) and the analysis of those transactions through fixed structures in data warehouses or lakehouses (OLAP). Both architectures are designed to answer predefined questions efficiently. That works superbly for business intelligence and traditional, predictive algorithms, but it rarely produces the structure an autonomous AI agent can work with directly.

An AI agent works in a fundamentally different way. Where a dashboard passively aggregates numbers, an agent retrieves information in real time, retains that context, reasons, and then takes action. For that, the agent needs meaning to be captured explicitly, in a machine-readable form. To unlock that context safely, with strictly controlled access to the right sources, a new, semantic data layer is required.

Extend, don't tear down

The good news is that the existing IT and data foundation does not have to be razed to the ground; it needs to be extended. This is an enormous advantage for organisations whose fundamentals are already in reasonable shape. The transition to genuinely AI-ready systems calls for a targeted architecture upgrade, built on three crucial pillars:

• From flat tables to a Knowledge Graph: A semantic layer is not new, but AI acts as a merciless X-ray of its quality. Where a colleague intuitively bridges differences in interpretation, an AI agent immediately grinds to a halt on contradictory data. The architecture must therefore evolve towards a Knowledge Graph. Think of it as a digital org chart: it not only stores individual data points, but also makes the relationships and meaning between those data points directly comprehensible to a machine.
• The rules of the game: metadata is 'king' again: If the Knowledge Graph is the road map, the metadata is the set of road signs. Metadata used to be a reference tool for analysts. For an AI agent, it has suddenly become the remote control. Instead of simply labelling a column 'revenue', the metadata must now literally instruct the machine: "If someone asks for monthly revenue, always use this field excluding VAT, and ignore table X." This turns central data governance into the literal rules for how AI answers.
• Execution: from passively looking on to acting autonomously and securely: Where a dashboard looks passively backwards, an agent acts in the present. This requires real-time data streams and bidirectional API integrations with source systems (such as ERP or CRM). As soon as an agent starts modifying records autonomously, an entirely new security challenge arises: Identity & Access Management (IAM) for machines. The authorisations and safety margins (guardrails) of an AI that takes action must be watertight, hard-wired into the system.

Questions for the boardroom

Tomorrow's market leaders will not distinguish themselves through the language models they buy, but through the robustness and precision with which they have made their own business logic machine-readable. For that, the conversation at C-level must move from the exploratory question, "Can we deploy AI?" to the structural question, "How should our organisation serve AI, and is our foundation genuinely AI-ready to extract as much value from it as possible?" The following strategic questions offer a starting point for that dialogue:

• Who are we building our foundation for? Do our investments in data still go mainly towards building better dashboards and reports for our employees? Or are we now also explicitly spending budget on making that same data 'readable' for an AI? And who, in fact, is ultimately responsible for that?
• The question of trust: If we gave an AI the authority today to carry out actions independently based on our current business data, where would it go wrong first? Have we really got data quality and security in good enough shape to trust a machine blindly, and what does IT need to cover that risk safely?
• Where does our real business knowledge sit? How much of the unwritten know-how, the definitions and the context that make our company unique still resides solely in employees' heads rather than in our systems? As long as that knowledge is not captured in the system, every AI remains generic. How are we going to secure that knowledge in a durable, systematic way?