# Enterprise AI Projects fail because governance is ceded to a vendor, says CTO advisor > Source: > Published: 2026-08-11 14:38:00+00:00 al/ml # Enterprise AI Projects fail because governance is ceded to a vendor, says CTO advisor [CTO Advisor](https://thectoadvisor.com/) Keith Townsend believes that enterprise AI fails at the agentic AI reasoning plane, not at the model or data levels. In an [X post](https://x.com/CTOAdvisor/status/1987509660706861395) he cites an oft-repeated statistic and says: “Ninety-five percent of enterprise AI projects fail to return anything. Almost nobody asks the question that would actually help. Where in the stack do they fail?” He claims to have learned the answer to this the hard way: “Last year I [moved a production AI system off Google Cloud onto an on-premises NVIDIA DGX Spark](https://labs.layer2c.com/labs/vctoa-to-spark). Not to leave Google. To learn what leaving would cost. The data moved in an afternoon. The judgment took weeks. “The embeddings, the retrieval logic, the semantic relationships that decide whether the model reasons or hallucinates. None of it ported. I rebuilt all of it by hand. That’s when the lesson landed. The data moved fine. The judgment didn’t.” Townsend has a model which presents a layered view of an AI infrastructure stack. Overall, there are eight layers, grouped into four functional planes topped by a value plane; the AI Application layer. This is logically separate from the infrastructure stack layers: For Townsend the [Layer 2C](https://layer2c.com/what-is-layer-2c) is the reasoning plane of an AI infrastructure stack — the layer where the system decides *where* to run a model, *which* agent handles a task, *how* to route between models, and *what* evidence gets recorded. It is the control surface that determines whether an enterprise retains meaningful governance over its AI, or cedes it to a vendor. Townsend has [blogged](https://thectoadvisor.com/blog/2026/08/10/where-enterprise-ai-actually-fails/): “I’ve [assessed 26 vendor platforms](https://us.list-manage.com/7VyAwoONn5v?e=d345bf7ea9&c2id=77e7e74733a5b73e7ab46053b9d53533) across all eight layers of the AI infrastructure stack. When you score every platform at every layer, one layer keeps failing. It isn’t the one the headlines name. His view is that everything below the sixth layer, his layer 2C, the reasoning or judgement plane, is portable between vendors, between the public cloud and the on-premises world, but the reasoning or judgement plane is not. He blogs that “it sits above your data and below your application.” And then says: “it answers a different question than the rest of the stack. Not “can the model run.” Whether the organization can trust what it does. Policy, escalation, evidence, and decision authority all live here. Most platforms have no explicit home for any of them.” It’s the judgement upon which the AI system is dependent. Typically a vendor will provide it but not explicitly and not in a standard way. It is a lock-in, owning the logic that decides what your data means. A customer’s AI system will make decisions and Townsend asks: ‘Who owns the authority to make these decisions?” Is it a person, a policy engine or the system vendor’s platform? He has devised another model to explain this, the [Decision Authority Placement Model (DAPM)](https://thectoadvisor.com/blog/2025/12/18/the-decision-authority-placement-model-dapm-dap-eem/). In this and across enterprise IT domains, automation is adopted for its execution benefits, while decision authority placement remains invisible until failure forces it into view. Enterprises routinely change how work is executed without redesigning who is authorized to decide at runtime. DAPM distinguishes between decision authority conditions and decision authority placement modes, and says there are different decision authority types: 1. Unplaced (Inherited) Authority describes the default condition in enterprise IT, where no actor has explicitly decided where decision authority should reside. Instead, authority is inherited from upstream platforms, vendor defaults, or historical operating assumptions. Intentional Decision Authority Placement Modes 2a. Platform-led Authority - decision authority is embedded within a centralized system that continuously evaluates and enforces runtime tradeoffs. The platform is authorized to decide outcomes related to efficiency, availability, performance, and cost within defined bounds. 2b. Governance-Coupled Authority - systems automate execution, but decision authority is escalated to human review under predefined conditions. Automation operates until thresholds are exceeded, ambiguity is detected, or risk criteria are met. 2c. Product-Aligned Authority - decision authority is deliberately delegated to product or service teams, bounded by explicit guardrails. Teams are authorized to make runtime tradeoffs within constraints defined by platform and governance functions. Townsend has defined an Authority Matrix to show where decision authority resides, where accountability is anchored, and how conflicts are resolved. He says: “Every decision is Retained, Delegated, or Ceded. Most enterprises have never mapped it. They find out in year two, when the system hallucinates on production data, nobody can explain why, and the judgment they need to fix it belongs to someone else.” His blog says: “You don’t have to take my word that this is where the value sits. Watch the vendors. The storage companies stopped selling storage. Pure Storage rebranded. VAST calls itself an operating system now. Google productized the exact layer I found was non-portable. They’re all climbing toward the reasoning plane, because that’s where the control lives. And the control is the lock-in.” For him the real AI infrastructure stack question is this: “When the board asks where the AI money went, don’t point at the model and don’t point at the data. Point one layer up. You don’t buy AI from a vendor. You buy decisions about which layers you own and which layers you let the platform own. The storage was never the decision that mattered. “If your organization can’t say who holds decision authority at the reasoning layer, you don’t have an AI strategy yet. You have someone else’s.” Comment Townsend draws a complicated picture with no easy way in. View it as a framework, a lens, for an AI team to use and, hopefully, get a method for making AI Infrastructure stack decisions explicit.