It usually does not, and the reason is more uncomfortable than "the model was not good enough."

Across 623+ case studies and 65+ practitioner interviews, the pattern that repeats is not a model problem:

Most production AI failures are meaning failures. The system answered a question nobody needed answered, optimised a metric nobody cared about, or ran at an autonomy level the organisation was not equipped to govern.

"If this gives the right answer to the wrong question, how would you know?"

Not rhetorical. It needs a specific operational answer before architecture work begins, and the most dangerous response is a confident, fast one from a team that has never considered it.

Vendor-agnostic, first principles, no BS. I want the abstractions that hold, and the real lessons from what shipped, including the failures. Inversion first: ask how this fails before asking how it succeeds. And intellectual honesty about the gaps, including my own: the published work carries a corrections ledger, because a confident answer that has not been stress-tested is the most dangerous artifact in the room.

_{What each one has had to fix, in public: Agentic CoE. ARIA's is the gap, and it is next.}

From production post-mortems, not theory. Each exists because something went wrong in a way the existing vocabulary could not name.

_{Five Modes. Autonomy transfers down the ladder. Responsibility does not, so the further you delegate, the more of your job becomes deciding what "correct" means in advance. You will not be in the room when it is decided.}

A compliance assistant that checks its own answers before you see them.

Every assistant cites sources. A source that supports the claim, a real document that says something else, and a document that never existed all look identical in the answer.

ARIA verifies every citation in the request path (with ordinary rules, not a second model asked for an opinion), then measures whether that verification actually works:

corruption of the evidence | ARIA's checks | a deliberately blind check
---------------------------------------------------------------------
                       10% |    +0.118     |          +0.000
                       50% |    +0.471     |          +0.000
                      100% |    +1.000     |          +0.000

The flat column is the one that matters. Without it, the first is just a number.

247 tests · no API key · ~2s. An independent review found four real defects, all fixed, and what it missed is published too.

Agentic CoE is the operating model, governance and quality gates that move enterprise AI past pilot purgatory, the Agent Card standard, a ten-gate Pre-Flight Checklist, the BXT Scorecard. Most programs stall not because the technology fails, but because nothing around it is built to fund, govern or scale what works.

The wiki is where the operating model is actually written down: 27 pages covering the factory and who staffs it, the agent patterns and risk tiers, the three drifts, the proof gates and the deployment ladder, what to do in the first 90 days, and what is not verified on any of it.

If you are working on the hard part, not whether AI can do something, but whether an organisation can trust it, govern it, operationalise it and prove value from it at scale: I would like to hear from you.

_{Interests · AI Transformation · Data & AI Strategy · AI Adoption · Intelligent Automation · AI Governance · Practice Leadership}

_{Technical rigor matters. So does building systems that serve people, not just metrics.}