Arguments, observations, and inconvenient truths about how organizations build, deploy, and govern enterprise systems.
Projects do not fail because governance was wrong. Projects fail because the solution was never known. The governance industry exists largely to manage the recovery cycle from this failure. Better governance is a painkiller. Formal validation is the cure.
Read articleAgentrification — the rush to deploy autonomous agents on unvalidated logic — is creating a new class of invisible, untraceable debt. When it matures, the remediation cost will make the last decade of cloud migrations look like a rounding error.
AI agents do not fail when given flawed logic. They improvise. They confidently fill in missing steps, invent exception paths, and produce plausible-sounding workflows that are structurally unsound. The output looks like gospel. It was generated from garbage.
Three waves of enterprise AI — thin wrappers, prompt engineering, uncaged agents — have failed for the same reason. The fourth wave is not an evolution of the first three. It is a rejection of the premise underlying all of them.
The enterprise AI problem is not a lack of intelligence. It is a lack of discipline. Language models are brilliant at language and architecturally unsuited for logic proofs. These are not the same problem and they do not have the same solution.
We test code. We inspect buildings before occupancy. We certify aircraft through thousands of hours of testing. We have never validated business logic before deploying enterprise systems. The 70% failure rate is the consequence — and it has not moved in twenty years.
Every organization operates simultaneously in three realities: the documented process, the perceived process, and the actual process. Discovering the gap between them is necessary. It is not sufficient. Diagnosis without proof is still a guess.
What is dying is not SaaS. What is dying is the fiction that CRM, ERP, and HCM are fundamentally different things requiring fundamentally different platforms. That fiction was never about how organizations operate. It was about how software was sold.
The backlog is not a queue of work waiting for execution capacity. It is a graveyard of abandoned ambitions. AI does not complete that list. It gives organizations permission to burn it and write a new one. What replaces it is the work that only humans can do.
In 1750 BCE, Hammurabi required builders to prove their structures were sound before anyone lived in them. The consequence fell on the builder personally. Enterprise software has no equivalent standard. The $87B annual failure rate is what that absence costs.
Construction proves load-bearing integrity before the first beam. Aerospace formally verifies flight control logic before certification. Banking proves risk exposure before issuing an instrument. Only enterprise software ships on assumption — and has normalized a 70% failure rate as a consequence.
Every migration that was supposed to be seamless was not. The reason is the same every time. The logic was never yours — it was absorbed by the platform. If you owned it, migration would be translation. The fact that it is not is the proof.
Every platform imposes its own constructs on organizational logic. What goes in is already distorted. Each migration starts from the previous system's distortion, not the original intent. The compounding cost is the real technical debt of enterprise software.
Flowcharts. BPM platforms. Low-code tools. RPA. Generative AI. Every generation of enterprise software tools gave you a better drawing. None of them gave you a proof. The gap remained. The cost compounded. The agents are about to execute it at scale.
Enterprise software fails for four distinct classes of reasons. The industry addresses two of them. The other two — structural failures and governance corruption — are invisible to every tool currently deployed. The $87B failure rate is the consequence.
UAT runs specific scenarios and checks specific outputs. Structural failures are properties of the logic itself — independent of any input. Testing samples behavior. Formal verification proves properties. The difference is expensive.
The kernel proves structural logic is sound before deployment. It does not prevent infrastructure failures or guarantee every business outcome. Knowing the boundary precisely is what makes the value proposition honest — and the company credible.