Xebia: AI agents fail without the right data foundation

Niels Zeilemaker, global chief technology officer at consultancy Xebia, told AI News that most failed deployments of AI agents in companies come down not to weak models but to the lack of a prepared data foundation. In his words, agentic AI scales only as far as the data the agents rely on is solid: if an organization has not thought this through in advance, the project risks stalling long before the agents start speeding up the processes they were deployed for.

Why data is the foundation, not a detail

Unlike conventional software, which follows a predefined script, an autonomous AI agent decides for itself which sources to consult, which data to read and which action to take next. If corporate information is scattered across disconnected systems, unlabeled and unavailable through clear interfaces, the agent will either fail to find the information it needs or receive it in distorted form and take the wrong action. Zeilemaker stresses that an agent's performance is directly limited by the quality and accessibility of the data it can reach, and no amount of raw model power compensates for that. This is what fundamentally separates agentic systems from classic generative AI pilots, where a model error rarely has real consequences outside the chat window.

The same logic already played out in the wave of generative AI projects in recent years: companies launched chatbots on internal data en masse, then found that the answers were either inaccurate or incomplete because the model had no access to current versions of documents or could not tell an outdated instruction from a valid one. For agents the stakes are higher, because an agent does not just produce text. It takes actions: it sends emails, updates CRM records, initiates payments. In an agentic scenario, an error caused by bad data turns not into an inaccurate answer but into a real, potentially costly action.

What lies behind Xebia's position

Xebia is an international IT consultancy that helps corporations with digital transformation, data infrastructure and putting AI into production processes. Zeilemaker's message is aimed squarely at those responsible for speeding up an organization's business processes with AI agents: in his words, they should start with the foundation, making data fit for consumption by agents, and only then move on to scaling the rollout itself. The sequence sounds obvious, but in practice it is the one most often broken: companies buy licenses for agent platforms before they have put in order the data sources those agents are supposed to use.

Key facts:

• Niels Zeilemaker is Xebia's global CTO
• The comment was published by AI News (artificialintelligence-news.com)
• Central thesis: agentic AI scales on the strength of data, not on model power alone
• Xebia specializes in data consulting and digital transformation for corporate clients

What this changes for teams deploying AI agents

The practical takeaway from Xebia's position is simple: before scaling a fleet of agents, audit your data sources. Find out where they are stored, how current they are, who manages them and whether an agent can safely be given access to them through an API. Companies that skip this step and go straight to pilots with flashy demo scenarios most often find that the agent works in a sandbox but breaks on real, incomplete or contradictory corporate data. Investment in data preparation, including source cataloging, quality control and access-rights separation, pays off at the scaling stage, not at the first prototype.

For technical leaders this means reordering budget priorities: part of the money that used to go exclusively to model subscriptions and agent frameworks should be redirected to data engineering, meaning pipelines that give agents access to current, verified data in real time. Without that step, even the most advanced model will make decisions blind, and the agents themselves risk becoming an expensive showcase of capabilities rather than a working tool that actually speeds up the company's processes.