Dustin Snoap
to you Imagine, for a moment, that AI actually worked. Not demo-worked. Not impressive-in-a-controlled-environment-worked. Actually worked, reliably, at scal
Imagine, for a moment, that AI actually worked. Not demo-worked. Not impressive-in-a-controlled-environment-worked. Actually worked, reliably, at scale, in the real world.
The industry would still need 18 months to evaluate vendors.
This is not a criticism. It is how large organizations work, and for good reason. There is a name for what happens when they don't. It is called a data breach, a regulatory fine, or a congressional hearing, depending on the industry and the ambition of the hallucination.
AI systems are probabilistic. They hallucinate: generating plausible-sounding information that is factually incorrect, with a confidence that bears no relation to accuracy. MIT research found that AI models use 34 percent more confident language when wrong than when correct. The model is most convincing precisely when it is most wrong, which is a design characteristic with consequences.
The legal profession has been stress-testing this in real time. In December 2025, fifty-one judicial opinions involved lawyers submitting AI-hallucinated citations in court filings. January brought thirty-six. February, thirty-three before the month finished. One firm billing $759 million annually was sanctioned for hallucinated briefs, promised updated AI policies, and then did it again. Generative AI lawsuits grew 978 percent from 2021 to 2025. As of early 2026, fewer than half a dozen AI liability insurance products exist worldwide.
Inside enterprises, the result is what the industry calls pilot purgatory. IDC found that for every 33 AI proofs of concept started, only four reach production. MIT's analysis of 300 deployments found 95 percent delivered no measurable business impact, and large enterprises abandoned an average of 2.3 initiatives in 2025, each carrying a sunk cost of $7.2 million.
The problem is not the model. It is the org chart the model has to pass through to become useful.
Security reviews. Compliance assessments. Legal sign-off. Integration with systems not built to communicate with anything after 2014. The pilot that impressed the board in March, awaiting approval from a stakeholder on sabbatical since April.
Only 14 percent of organizations deploy AI agents with full security approval. The rest run shadow AI. Shadow AI incidents add an average of $670,000 to breach costs.
The electricity analogy is instructive. Electricity was commercially available in the 1880s. Most American factories didn't electrify until the 1920s. Forty years. Not because factory owners were slow. Because replacing infrastructure takes time even when the outcome isn't in doubt.
AI is not forty years away from meaningful enterprise deployment. It is also not eighteen months away, which is what most press releases imply.
The future is in procurement. It is sitting in a queue. It has been assigned an owner. There is a meeting scheduled for Thursday, and the right people have been invited this time.