Why 95% of AI pilots fail — and what the other 5% share

MIT's 2025 research — naming the phenomenon the GenAI Divide — found that 95% of enterprise GenAI pilots show no measurable P&L impact. Not disappointing impact. No measurable impact. The study also found that externally built systems succeed at roughly twice the rate of internal builds. Organisations that bought judgement did better than organisations that bought technology and supplied their own. The conclusion is uncomfortable for a technology industry that has promised intelligence out of the box: the models are not the problem. The problem is AI doctrine — or rather, its absence.

What do the 95% of failed AI pilots have in common?

The 95% share a common cause: tools without doctrine. Capable AI systems dropped into organisations with nothing to teach them. An agent that has not been taught how a business thinks produces fluent, plausible, generic output. That is the most dangerous kind, because it reads like work.

The models themselves are extraordinary — and getting cheaper by the quarter. Every competitor now rents the same intelligence from the same handful of providers at the same falling price. When the intelligence is a commodity, value concentrates where differentiation is highest — and that is what you teach the agent.

Why the model is not the moat

Large language models commoditise. The moat is not the model — everyone has the model. The moat is the doctrine. Not your data. Data describes what happened; every organisation has warehouses of it, and most of it teaches nothing. What differentiates is the way to think: what a good strategy looks like, how it translates into the procurement process and into contracts, which deviations matter and which are noise, when a saving is real and when it is a report, what a rate actually buys.

That is doctrine. It does not come with the model. It cannot be scraped. It is the one input a competitor cannot rent. Cost and value intelligence provides the framework that gives the AI the DNA to make useful commercial decisions — not just generate plausible text.

How doctrine turns a general AI into a commercial asset

Doctrine is what transforms a general-purpose language model into a system that can contribute to commercial decisions. It is the structured body of principles, decision rules, and performance criteria that tells the AI:

• What a good procurement strategy looks like — not just what happened last time
• How to distinguish a genuine saving from a reported discount
• Which contract deviations matter and which are commercial noise
• What total cost means in the context of a specific category or contract
• How to evaluate a tender beyond the unit price

Without doctrine, the AI produces output that sounds right. With doctrine, it produces output that is right — because the principles against which correctness is judged are embedded in the system, not supplied by the user at the prompt. The difference is the same distinction explored in strategic sourcing versus a sourcing strategy: one is a discipline embedded in how you work. The other is a document you produce.

What powerabode built on this finding

powerabode spent five years writing commercial doctrine for the energy industry — tested in live engagements — and then structured it for machines. The result is The Workbench: AI agents that are taught the doctrine before they encounter the data. Three design choices matter:

• The agent is corporate memory. People rotate every few years. The agent stays. What it knows about the contract and the commercial logic that governs it stays with it.
• Collaborative, not policing. The agent works alongside the contract holder, not auditing from above. It supports decisions, not enforces compliance.
• Client data stays in the client's environment. The doctrine travels. The data does not.

Tools without doctrine vs doctrine-driven AI compared

Comparison of tools without doctrine and doctrine-driven AI systems

Dimension	Tools without doctrine	Doctrine-driven AI
Output quality	Plausible, generic, often misleading	Grounded in commercial principles
Differentiation	None — same model, same output quality	High — doctrine is proprietary
Business impact	None measurable (MIT finding)	Measurable — decisions improve
Learning source	General training data	Embedded commercial doctrine plus data
User experience	User must supply context and judgement	System supplies context; user supplies approval

What the GenAI Divide means for procurement leaders

The GenAI Divide is not about technology access — it is about whether the organisation has doctrine to teach the technology. For procurement and commercial leaders, the implication is clear: buying an AI tool without first defining the commercial doctrine it should follow is purchasing the 95% outcome. The successful 5% bought a way to think. The rest bought tools. The choice is available to every organisation. The difference is whether the doctrine exists before the procurement begins.

Frequently asked questions

What did the MIT 2025 GenAI study find?

The study found that 95% of enterprise GenAI pilots showed no measurable P&L impact. Externally built systems succeeded at roughly twice the rate of internal builds — organisations that bought judgement did better than those that bought technology and supplied their own.

What do the successful 5% share?

The successful 5% share doctrine. They taught their AI systems how to think — what good looks like, which deviations matter, when a saving is real — rather than just feeding them data and expecting intelligence to emerge.

What is AI doctrine?

AI doctrine is the structured body of commercial principles, decision rules, and performance criteria that tells an AI system what good procurement looks like — not just what happened in the past.

Why do externally built AI systems succeed more often?

Externally built systems arrive with embedded doctrine — the provider has already encoded what good looks like. Internal builds often start with the technology and attempt to develop the doctrine during implementation, which is much harder.

What is the GenAI Divide?

The GenAI Divide is MIT's term for the gap between the 95% of AI pilots that achieve no measurable business impact and the 5% that do. The divide is not about technology access or model quality — it is about whether the system was taught a way to think.

Summary

• 95% of enterprise AI pilots fail to deliver measurable P&L impact, according to MIT's 2025 research. The common cause is tools deployed without doctrine.
• The moat is not the model — everyone has the model. The moat is the doctrine: what the AI has been taught about good commercial decisions.
• powerabode's response: five years of commercial doctrine, structured for machines, delivered through The Workbench — agents taught to think before they encounter the data.

See it working — book a demonstration

Continue reading: The Workbench — AI for the complete contract lifecycle.

Related service: procurement and supply chain consulting for oil and gas.

References

1. MIT 2025 study on enterprise GenAI — the GenAI Divide.
2. Cost & Value Intelligence — the doctrine behind cost and value decisions. powerabode.