DATA & ANALYTICS
Josh Fecteau serves as Chief Data & AI Officer at Teradata, an autonomous knowledge platform. What’ s more, the CIO title was recently added to his remit, focused on accelerating the modernisation of Teradata’ s internal data ecosystem and creating an integrated foundation for enterprise-wide AI and technology execution. In this combined role, he guides the Technology Services organisation and integrates it with the Data and AI function, which Josh describes as“ a unique opportunity in this era of AI disruption”.
Q. WHAT MISCONCEPTIONS DO ENTERPRISE LEADERS HAVE ABOUT WHY AI AGENTS FAIL WHEN MOVING FROM PILOT TESTS TO PRODUCTION AT SCALE?
» The single biggest misconception is that LLMs alone should be able to solve enterprise problems. Personal productivity tools don’ t magically translate to enterprise use cases. The early instincts were to blame the models, but the models are only as good as the context they are given. For an agent to reason reliably, data needs to carry meaning, not just values: metadata, lineage, entity relationships, business rules embedded at the data layer.
Pilots work because someone prepared the data. Production fails because the real data is not like that, and company data and context are always evolving. Teradata’ s research found that 84 % of UK executives report that 20 % or less of their data is sufficiently described and contextualised.
The problem is in the data foundation. And it is hard, because data is always changing and agents do not understand the consequences of those changes. That gap between a decision and the reasoning behind it is what I call temporal truths, and I do not think the industry is talking about it enough yet.
Q. WHAT IS A TEMPORAL TRUTH, AND WHY DOES IGNORING THIS CONTEXTUAL REASONING CAUSE AGENTS TO BREAK DOWN OVER TIME?
» A temporal truth is a decision that was right when it was made, but since it was recorded without any background information behind it, nothing can inform the system when the reasons it was made no longer apply.
There are two parts to this. The first is completeness. Even on day one, the record was a fragment. The full context that made a decision the right one – the market conditions, the relationship dynamics, the conversation that happened 10 minutes before – almost none of that makes it into any system.
The second is temporality. Because that context was never captured, nothing can watch it change. The record sits there looking accurate while everything that made it true quietly shifts.
When that happens, agents execute decisions that are no longer valid, with no error message, no signal that anything is wrong. The agent just keeps going, executing confidently on a truth that stopped being true some time ago. That is what makes this failure so costly at scale.
aimagazine. com 127