The load-bearing detail in Red Hat's account is not the agent, it is the guidance file, and specifically who writes it. Each major data domain has a named owner accountable for quality and documentation, and the guidance is maintained by those owners rather than drafted by a central team [3]. The semantic layer that makes plain-language querying safe is therefore a set of documents kept current by people whose job is the domain, not the platform.
Red Hat is unusually plain about sequencing: the data platform was necessary and, on its own, not enough [6]. What is telling is the list of problems the team says it hit after the foundation was in place. Trust, because an agent will produce a wrong number with no hesitation, and a wrong figure in a presentation or one that triggers an automated action burns credibility, which is why the agent has to show sources and reasoning [7]. Then usability, business context, and ownership. Model capability is not among them.
That matters because the word doing the most work in the announcement is "trusted" [1], and it rests on process rather than on any published measurement: the post reports no accuracy rate, no adoption figure, and no query volume [1]. The process itself is specified, though. The data product owner writes the guidance and runs user acceptance testing on common questions, the AI team builds an agent that explains its reasoning, and the user makes the call [10]. That is a division of warranty, and it is the kind of thing that is written down after someone has been burned.
Deduplication is the part nobody demos. Consolidating redundant sources left the agent fewer places to look and less conflicting information to reconcile, and settling on source-aligned and aggregate data products lets humans and agents land deterministically on the same source of truth [4]. Auditability depends on that ordering. An answer cannot cite its source if the organisation has three of them and no ruling on which one counts, and Red Hat describes starting from data locked in source systems, legacy warehouses, databases and spreadsheets [5].
Open sourcing the foundational AI templates [11] gives other teams the scaffolding and none of the substance. Nobody can ship your definition of annual contract value, or the ruling on what "active customer" means for each product line. The templates are the week of work. The guidance files, the owners, and the retired duplicate warehouses are the year of work, and they are the reason this particular agent is allowed near a number that someone will act on.
