95% of AI Pilots Fail, the Other 5% Are Worth Understanding

We previously looked into the evidence the world has run out of the physical infrastructure AI needs. This post looks at the demand side of AI. What are enterprises doing with AI? When you look closely, it is considerably less than the earnings calls suggest.

Broad use, narrow value

McKinsey’s 2025 State of AI survey is the most comprehensive of its kind. 1,993 respondents across 105 countries. The headline: 88% of organisations now use AI somewhere, up from 78% the year before. Two-thirds use it in multiple functions. [1]

Transformation? The devil is in the details. Only 39% of respondents attribute any impact on earnings to AI. Most of those who could said the impact was less than 5%. Just 6% of organisations qualified as “high performers” with meaningful earnings impact.

Nearly two-thirds had not yet begun scaling AI across the enterprise.

The MIT number

The most cited source of AI scepticism comes from MIT’s GenAI Divide report, published mid-2025. Based on executive interviews, leadership surveys and 300 documented deployments. MIT found that of $30 to $40 billion in enterprise generative AI investment, 95% of pilot programmes delivered zero measurable return. Only 5% of integrated systems generated significant value. [2]

The cause was not the technology. Generic tools like ChatGPT see heavy adoption. The failures were in workflow integration, organisational inertia, and what the researchers called a “learning gap” between models and the institutions trying to use them.

The MIT figure has been criticised. Some argue the methodology over-weighted large enterprise rollouts and missed grassroots employee adoption. But the broad pattern shows up in adjacent surveys. Gartner’s research has tracked similar attrition rates.

AI washing: the regulators step in

This gap, between rhetoric and reality, has caught the eye of regulators. In February 2025, the US Securities and Exchange Commission (SEC) created a dedicated unit with “rooting out AI washing” as a stated priority. AI washing is the practice of overstating AI capabilities to investors. [3]

Several cases are instructive. Presto Automation marketed a drive-through AI product as eliminating human order-takers. The SEC found that most orders required human intervention and the underlying AI was owned by a third party. Nate Inc raised $42 million by claiming AI completed online purchases, when nearly all orders were processed manually by contractors. [4]

AI washing remains an explicit regulatory priority for 2026. [3]

Gartner makes it official

Gartner’s 2025 Hype Cycle for Artificial Intelligence placed generative AI in the ‘Trough of Disillusionment’. This is the phase where original excitement wears off and early adopters report performance issues and low returns.

Meanwhile, AI agents have been pushed up to the ‘Peak of Inflated Expectations’. Gartner notes that 57% of organisations admit their data is not AI-ready and that organisations lack genuine trust in AI agents’ ability to operate without human oversight. [5]

Agentic AI: the hardest reality check

The next wave of corporate AI rhetoric is ‘agentic AI.’ Autonomous software that does not just answer questions but plans tasks, calls tools, executes workflows and makes decisions across multiple steps. Sundar Pichai, Satya Nadella, Marc Benioff and dozens of other CEOs spent 2025 telling shareholders that agents would replace large parts of office work in 2026.

The reality is that it is early days for production-grade agentic AI.

McKinsey found that while 62% of organisations report experimenting with AI agents, only 23% are scaling them in even one function. In any given business function, no more than 10% of organisations have scaled agent deployment. [1] The successful use cases are tightly bounded: IT service desks, internal knowledge retrieval, software engineering copilots. Single-domain, structured tools where errors can be caught.

Gary Marcus, the cognitive scientist and prominent AI sceptic, predicted at the start of 2025 that AI agents would be:

“Endlessly hyped but far from reliable, except possibly in very narrow use cases.”

His year-end review concluded that prediction held. [6] The structural problem is compounding error. Agentic tasks involve multiple steps. In systems like Large Language Models (LLMs) that are fundamentally probabilistic, multiple steps mean multiplying the chances of failure.

The CPU surprise

Here is where the supply-side argument from the previous post gets worse.

In April 2026, Morgan Stanley published a 73-page report arguing that the next phase of AI infrastructure is not bottlenecked by GPUs at all. It is bottlenecked by CPUs, memory and substrates. [7]

A chatbot serves one query and the GPU does the work. The CPU is a supporting player. Morgan Stanley estimates a traditional chatbot AI server uses roughly one CPU per twelve GPUs. An agent, by contrast, must plan, call APIs, retrieve data, run code, evaluate the output and iterate. A joint Georgia Tech and Intel study found that CPU-side processing accounts for 50 to 90% of end-to-end latency in agent workloads. [7]

Morgan Stanley estimates the CPU-to-GPU ratio in Nvidia’s next-generation Rubin platform falls to roughly 1:2. Agentic workloads will generate 15 to 45 exabytes of additional DRAM demand by 2030. Morgan Stanley pegs the incremental CPU market from agentic AI at $32.5 to $60 billion by 2030. [8]

The implication is that the supply problem does not go away when GPU production catches up. Agentic AI moves the bottleneck sideways, from GPUs to CPUs, substrates, foundry capacity and DRAM. All of which have even slower expansion cycles than GPUs themselves.

Asked when production-grade agentic AI will exist at scale, expert answers cluster between “two to five years” and “we don’t yet know how.” Gartner classifies agentic AI as five to ten years from mainstream maturity. [5] Morgan Stanley says the agentic cycle will redefine infrastructure priorities over the next five years. [7] Marcus argues that without robust, trustworthy AI, agents will not work at all in the general case. [6]

The gap between agentic AI demos and agentic AI in production is currently enormous. Nobody has yet shipped the solution to unreliability.

The next and final post in this series covers the bull case, the bear case, what it means for smaller companies, and the three things non-technical managers should plan around.

References

[1] McKinsey & Company, “The State of AI: Global Survey 2025.” https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

[2] Fortune, “MIT report: 95% of generative AI pilots at companies are failing” (August 2025). https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

[3] StoneTurn, “Next-Generation Compliance: Preparing for Continued SEC AI Washing Enforcement.” https://stoneturn.com/insight/next-generation-compliance-preparing-for-continued-sec-ai-washing-enforcement/

[4] Darrow, “AI Washing Sparks Investor Suits and SEC Scrutiny.” https://www.darrow.ai/resources/ai-washing

[5] Gartner, “Hype Cycle for Agentic AI.” https://www.gartner.com/en/documents/7671861

[6] Gary Marcus (Substack), “Six (or Seven) Predictions for AI 2026 from a Generative AI Realist.” https://garymarcus.substack.com/p/six-or-seven-predictions-for-ai-2026

[7] Bitget News, “Morgan Stanley Major Research Report: The Rise of AI Agents and Why the Bottleneck Has Shifted from GPU to CPU.” https://www.bitget.com/amp/news/detail/12560605375963

[8] ANI News, “Morgan Stanley: Agentic AI Shifts Value from GPUs to CPUs and Memory, Creating Up to $60bn Incremental CPU TAM by 2030.” https://aninews.in/news/business/morgan-stanley-agentic-ai-shifts-value-from-gpus-to-cpus-and-memory-creating-up-to-60bn-incremental-cpu-tam-by-203020260422131744/