I run a scored thesis on enterprise AI, and one line in it sits above the noise at 85% confidence: the bottleneck isn't capability, it's implementation. Every few months a new report lands that makes the same point with different numbers.
Camunda's 2026 State of Agentic Orchestration and Automation report surveyed 1,150 senior IT and business leaders and found that 73% admit there's a real gap between their vision for agentic AI and where they actually are. 71% say they're "using AI agents." Only 11% have anything in production. Gartner tells the same story from a different angle: it expects more than 40% of agentic AI projects to be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. MIT's Project NANDA puts a third number on it: roughly 95% of generative AI pilots show no measurable return on the income statement.
Three different research houses, three different survey designs, converging on the same answer. This isn't a capability problem. It's an orchestration problem.
The gap isn't technical, it's procedural
I spent close to two decades before starting Hundred Solutions inside the ERPs, CRMs, and accounting platforms that most transformation slides never mention. I was the one who got the call when an integration that looked perfect in the demo started failing in ways nobody had scoped for. That job taught me something every one of these reports keeps confirming from a different angle: a pilot rarely dies because the model was wrong. It dies because nobody built a path for it to leave the lab.
The pattern is consistent enough that the research has a name for the cause. Camunda's data shows 84% of organizations cite business risk when IT lacks appropriate controls over AI agents, 80% point to a lack of transparency in how AI gets used inside business processes, and 66% cite compliance concerns directly. Add it up and you get the same story I saw for twenty years with ordinary system integrations: the pilot works in the demo, then someone asks who signs off on it touching real customer data, and that single question routes the project into a security review, a compliance review, and a budget conversation nobody scoped for at the pilot stage. Three months later the pilot is technically alive and organizationally dead.
Why the model was never the real question
A huge amount of energy in these rooms goes into debating which foundation model, which vendor, which architecture. I understand the appeal, it's a concrete decision that feels like progress. But in every failure I've traced back to its root cause, the model was rarely the constraint. The constraint was that nobody had defined, before the pilot started, what "ready for production" required and who owned that decision once the demo stopped being a demo.
That's also why I put so much weight on one specific number from the same body of research: partnerships that work alongside a customer to understand their actual workflow before building anything succeed roughly twice as often as internal teams that skip straight to building. The gap isn't the tooling either team has access to. It's whether anyone bothered to solve the real problem before automating an imagined one, which is exactly the argument I make at length elsewhere.
The question that should replace "did the demo work"
If I could change one habit in how enterprises run AI pilots, it would be this: stop measuring success by whether the demo worked, and start measuring it by whether someone has agreed, in writing, on who owns the decision the system is helping make.
That second question forces the governance conversation to happen at the start instead of the end. It also enforces something I think gets lost in a lot of automation enthusiasm: the AI does not get to make the call. A human owns the decision, every time, and the pilot's job is to make that person faster and better informed, not to quietly replace them. Pilots that skip this step aren't unfinished, they're structurally incomplete, and no amount of additional model capability fixes a missing owner.
The failure rate isn't a verdict on AI. It's a mirror held up to how casually most organizations still treat the handoff between "we built something that works" and "we built something we can run." Fix that handoff, and the technology mostly takes care of itself.