Our Take
Xebia is selling the obvious: agents need good data. The real claim—that data catalogues designed for human fallback don't work for agents—is valid but unsupported by customer wins or benchmarks.
Why it matters
Enterprise teams building agent pilots are discovering this the hard way. Zeilemaker's framing (data catalogue as prerequisite, not afterthought) may prevent costly restarts, but practitioners need specifics on what 'AI-ready' cataloguing looks like in practice.
Do this week
Data engineering lead: audit your current data catalogue against Xebia's Agentic Data Foundation criteria before approving your first agent deployment, so you can avoid misinterpretation failures in production.
The data catalogue problem agents expose
Niels Zeilemaker, global CTO at Xebia, argues that AI agents expose a fundamental weakness in how enterprises catalogue data. When humans encounter incomplete or unclear documentation, they have a workaround: they ask a colleague. Agents do not.
"If the description is wrong, the agents will not perform," Zeilemaker stated. The failure modes he identified include agents unable to locate the correct data, misinterpreting field values, and joining datasets that should remain separate. The root cause is not agent incompetence but a data foundation designed for human flexibility, not machine precision.
Xebia responded with Agentic Data Foundation (ADF), a framework that extends data platforms to host agents and deploy them in both customer-facing and internal workflows. The company packages this alongside Xebia ACE, an AI-native software engineering framework claimed to accelerate delivery by up to 40% and cut legacy transformation costs by up to 70% (company-reported). Zeilemaker positioned ADF as a way to compress 12- to 24-month migration timelines into fixed-price, milestone-bound engagements by co-developing solutions with customers.
Data readiness is the admission fee, not the prize
Zeilemaker is correct that agents are more brittle than humans when data quality declines. What he does not address is whether this is a hard constraint or a solved problem with existing tools.
The claim that data catalogues must be redesigned specifically for agent consumption is plausible but vague. No independent benchmark compares agent performance on human-optimized catalogues versus agent-optimized ones. Xebia's framing—that 12–24 month migrations compress into fixed-price engagements—reads more like a sales narrative than a technical breakthrough. Proof would require published timelines, customer references, or independent audits comparing ADF migrations to traditional approaches.
The security angle Zeilemaker raised is more interesting: as code generation scales, pull request reviews become a potential attack surface. He cited Anthropic's recent work on LLM-assisted code review as a sign the industry is waking up to this. That observation is worth watching but not yet a solution.
What to do before you deploy an agent
If your organisation is building an agent pilot, do not assume your current data catalogue is sufficient. Test your agent against a real query using your existing documentation. If the agent fails to locate fields, conflates unrelated datasets, or returns confabulated answers, the problem is your catalogue, not the agent.
Second, audit the metadata and lineage for each data source your agent will consume. Agents do not have the option to phone a colleague. Third, if you are considering Xebia ADF or a competitor, ask for a pilot migration timeline and a reference customer who can attest to the 40–70% claimed acceleration. Until then, treat those figures as vendor estimates, not proof.