An enterprise can build a capable AI agent and still receive unreliable work because the system beneath it does not explain what corporate data means. Niels Zeilemaker, global CTO at Xebia, argues that fragmented platforms, weak catalogues and ambiguous fields become agent failures even when the model behaves exactly as designed.
Humans compensate for missing documentation by asking a colleague. An agent has no informal corridor conversation. It relies on the description, ownership and relationships encoded in the data environment—and confidently inherits every mistake.
A machine-ready catalogue replaces human back channels
A conventional catalogue may list a dataset without explaining whether a field is current, sensitive or suitable for a particular decision. People often learn those details through institutional memory. Autonomous software needs them expressed directly.
Definitions should identify authoritative sources, data owners, refresh schedules, lineage and permitted uses. Relationships must be explicit so an agent does not join two similar identifiers that refer to different customers, products or accounting periods.
Access controls also need to survive orchestration. An agent that can query several systems should not combine individually permitted facts into a sensitive inference the user was never authorised to see. Governance must consider the result of a workflow, not only each database call.
Xebia describes an Agentic Data Foundation that extends a data platform to host agents and support customer-facing and internal processes. The value lies in placing execution close to governed context rather than attaching a chatbot to disconnected repositories.
Data quality becomes observable when agents expose recurring ambiguity. Failed tasks can reveal undocumented fields or inconsistent records, giving organisations a practical backlog for improving the foundation.
Agents can accelerate migration when context travels with the work
Legacy modernisation is a promising use because migrations contain repeated mapping, transformation and validation steps. Xebia says accumulated platform context can help agents compress work that previously required long consulting engagements.
Speed does not remove the need for acceptance criteria. Teams should compare record counts, business rules, performance and exception handling before replacing a production process. An automated migration can reproduce a misunderstood rule at enormous scale.
The same principle applies to AI-assisted software engineering. Xebia’s ACE framework embeds models across the development lifecycle while maintaining established governance and quality. The attraction is not uncontrolled “vibe coding,” but faster planning, implementation and review within a process capable of rejecting weak code.
Model-based pull-request review can add another perspective, especially as code volume rises. It should complement tests, security analysis and accountable human approval rather than create a false guarantee from a long automated comment.
Production readiness depends on ownership and measurable outcomes
Organisations should begin with a bounded workflow whose data owner, business owner and technical owner are known. The agent needs a clear task, a reliable source and an escalation route when records conflict.
Evaluation should include correctness, exception rate, human correction and the time or cost saved. A prototype that produces an impressive answer but cannot be audited, monitored or maintained is not a production asset.
Knowledge sharing is part of the operating model. Teams need a way to publish what worked, what failed and which catalogue improvements made an agent dependable. Otherwise every department repeats the same experiments behind different interfaces.
The foundation is not a one-time cleanup. New systems, definitions and regulations continuously change the meaning of data. Owners must review metadata and permissions as carefully as model versions.
Agent projects often begin with a demonstration of reasoning. Xebia’s argument brings attention back to the less glamorous work that determines whether reasoning can act safely. Reliable autonomy grows from clear data, controlled software delivery and people who remain responsible for both. Without those conditions, a faster agent merely reaches the wrong conclusion sooner.
