In the majority of failed deployments, the model did exactly what it was asked to do. The problem lies in what it was asked to do, in business processes that were never really documented.
The analysis published by InformationWeek takes the opposite of the dominant discourse on the failures of AI agentics in companies. Justin Bolles, from the resulting firm, sums up the field report: in most of the cases observed, the technology did exactly what it was asked to do; The trouble comes from what was asked of him.
The recurring cause is the same: actual business processes do not resemble their documentation. They are full of informal circumventions, blurred responsibilities, never-recorded exceptions and steps that only work because an experienced employee knows what to do. A human absorbs these grey areas without thinking about it; An agent, no.
The resulting recommendations reverse the usual order of priorities: to start with simple and clearly defined processes, to add deterministic validation points, to ensure complete traceability of actions, and above all to redraw workflows around what agents really know how to do — before any deployment, not after.
Read the full article on the original source
Go to the source →