Far too many artificial intelligence projects are being pulled before they ever reach production. While corporate history is littered with cancelled technology initiatives, the scale and velocity of these failures set the current AI landscape apart.
S&P Global's 2025 enterprise survey of more than 1,000 business and IT leaders revealed that the average organisation has scrapped 46% of its AI proofs of concept. More concerningly, the proportion of companies abandoning the vast majority of their AI initiatives surged from 17% to 42% in a single year. This acceleration suggests that organisations are not settling into a mature rhythm; the pattern of failure is compounding.
The stated reasons for these failures are predictable. S&P's respondents cited compounding costs, data privacy liabilities, and security risks. Gartner's analysis of post-proof-of-concept abandonment adds poor data quality, inadequate risk controls, and ambiguous business value. Meanwhile, the RAND Corporation points to unclear problem definitions and an organisational obsession with the technology rather than the actual use case. Paradoxically, almost all of these variables are identifiable before work begins.
The critical question is why teams fail to flag them during the low-cost prototyping phase. Historically, the expense of software development forced early friction; companies had to justify investments to steering committees or Lean Portfolio Management reviews before writing code. Though imperfect, they forced organisations to answer whether a project should be built before capital was deployed.
Generative AI has removed that financial friction. When anyone can build a functional prototype in an afternoon, the urge to experiment outpaces the impulse to govern. Driven by rapid prototyping, teams rush forward into blind spots that traditional oversight would have exposed immediately.
Three of the failures — spiralling costs, ambiguous business value, and poorly defined problems — are inextricably linked. They are all symptoms of a reactive, "JFDI" approach driven by the industry bandwagon over the last twelve months. A company mandate of "we have to use AI" routinely lacks the strategy, operational goals, or financial controls needed to achieve anything before their budget has been spent.
The solution is straightforward, yet it is where I risk becoming unpopular: You need to get better at prioritisation.
Each pilot is individually defensible, backed by a protective sponsor, and justified in its own silo. Yet because generative AI has dismantled the initial financial barrier to entry, none of these projects are forced to compete for survival or prove they deliver more business value than the rest.
While building a prototype has become nearly costless, the downstream operational resources required to safely move an enterprise software project into production remain stubbornly finite and highly expensive. True implementation demands substantial engineering hours to build clean data pipelines, weeks of dedicated security review, specialised compute architecture, and executive focus to dismantle legacy roadblocks.
When eleven concurrent pilots scramble for the same data engineers and clog the same security sign-off queues, the operational maths fails. Every project receives only a useless fraction of the infrastructure and attention it requires to achieve a trustworthy, production-ready result. Consequently, no single initiative generates the definitive performance metrics needed to justify building a permanent data pool or writing a rigid risk framework.
This is what puts the technical fundamentals out of reach, starting with data infrastructure. When enterprises attempt to plug frontier models into unverified data lakes, failure is inevitable. True data utility demands a rigorous foundation: companies must systematically audit their internal datasets to prevent stagnation, and deploy targeted data pooling to provision the precise inputs their agents require to generate value.
Second, risk controls. Implementing effective governance requires moving beyond abstract policy into operational enforcement. Organisations must establish clear, actionable AI governance and assurance frameworks, ensuring that employees are explicitly trained on compliant usage. Safe deployment demands technical guardrails that actively regulate how internal data is handled within the business. Crucially, companies must map the full lifecycle of their data pipelines, ownership boundaries and liabilities before any models go live.
Open whatever tool holds your company's strategic backlog and count the number of items within. Next, how many have been successfully delivered in the last 12 months. Now take the first number, divide by the second — in very rough maths this is how many years it will take you to deliver everything your company feels is important. The worst I have seen is 102 years of work. If you land on 3–5 years you are fine.
The answer is enforcing comparative prioritisation at the point of entry. If a new project stream cannot prove it delivers strategic value then it should not be worked on. The risk is a company churning through productised "AI Slop" — just because it is fast to create and the words are in a compelling order doesn't mean it brings value to your customers.
Sarah McConnell is the founder of MML (Measure Model Lead), where she works with organisations on AI readiness. Find her on LinkedIn.