--- title: "Why Do AI Projects Fail? Five Causes and a 12-Point Check" description: "The five causes behind most AI project failures, what published failure rates mean, and a twelve-point pre-mortem to run before your next build." canonical: https://shipshapedata.com/resources/ai-foundations/why-ai-projects-fail/ language: en-GB author: "Emile Van Zyl" author_role: "MarTech Automation Consultant" author_url: https://www.linkedin.com/in/emile-vanzyl-enhancing-your-human-advantage/ date_published: 2026-10-12 date_modified: 2026-10-12 --- # Why do AI projects fail, and the checks that predict it > The five causes behind most AI project failures, what published failure rates mean, and a twelve-point pre-mortem to run before your next build. Section: Home > Resources > AI foundations Canonical page: https://shipshapedata.com/resources/ai-foundations/why-ai-projects-fail/ By Emile Van Zyl, MarTech Automation Consultant (https://www.linkedin.com/in/emile-vanzyl-enhancing-your-human-advantage/) | published 2026-10-12 A pilot impresses in a demonstration and then goes nowhere. A model that scored well in testing is never switched on for real users. The budget is spent, the board asks what it bought, and nobody can point to a decision that changed. These outcomes are common, and they are rarely caused by the technology. Why do AI projects fail comes down, in most cases, to decisions made before the build starts: data that cannot carry the use case, no named owner, no measurable decision for the AI to improve, and no route from pilot to production. The model itself is rarely the cause, and each of these risks can be checked in advance. This guide is for the person who has been handed an AI mandate, or who has watched a pilot stall and needs to explain why. It sets out how often AI projects fail and what the published figures mean, the five causes that recur, a twelve-point pre-mortem you can run before the next build, and what the projects that reach production do differently. Shipshape Data builds data foundations and the AI that runs on them. Most of the failures we are asked to look at were decided before the build began, and most could have been seen in advance with a short set of honest checks. ## How often AI projects fail Published figures vary widely, and most of the variation comes from definitions. Three sources are worth knowing, with their dates. - Gartner, January 2026. Gartner's article on generative AI project failure is framed around a 50 per cent failure rate. It names poor use-case selection combined with a lack of business value as the reason that consistently tops the list. - Trullion, September 2025. Trullion's article is framed around the claim that 95 per cent of generative AI projects fail. It argues that AI fails when it is misapplied rather than because the technology is flawed, and that generic copilots and one-off pilots are not where the successful projects sit. - RAND, August 2024. RAND's research report on the root causes of AI project failure does not rest on one headline rate. It sets out several causes, including the case where the technology is applied to problems that are too difficult for AI to solve. Academic work points the same way. A 2022 review of AI project failure by Westenberger and colleagues, published on ScienceDirect, found that both organisational and technological issues can cause a project to fail. ### Why the figures disagree A failure rate is only as useful as its definition. Failure can mean a project that never reached production, a project that reached production but never returned its cost, or a project abandoned at any stage. Generative AI and older machine learning work are often reported separately, and the figures are not interchangeable. When you quote a figure to a board, quote its source, its date and its definition with it. The more useful point is that the serious sources agree on direction. Most failures come from the organisation and the data, not from the model. ## The five causes that recur The causes below appear again and again, in different combinations. Most projects that fail show at least two of them. ### 1. The data cannot carry the use case A use case needs data that is complete, current, consistently defined and reachable. Often it is none of these. Customer records disagree between the CRM and the finance system. The documents a chatbot should answer from are out of date. The history a forecast needs was never kept. Informatica's February 2025 article attributes much of the failure and low scalability of AI projects to the gap between traditional data management and AI-ready data management. Cutter's May 2025 article lists weak infrastructure for managing data and deploying models as a cause in its own right. A model trained or grounded on unreliable data produces unreliable output with confidence. A data quality framework gives you a way to define what good data means for the use case and to check it before anyone builds on it. ### 2. No one owns the outcome AI work often starts in a technical team and stays there. The business unit whose process it should change is consulted late or not at all. IBM's October 2025 article states that attempting enterprise AI in a vacuum, excluding strategic stakeholders and business units, is guaranteed to fail. Ownership means a named person in the business who is accountable for the result, who decides when it is good enough, and who will change the process to use it. Without that person, a working model has nowhere to go. ### 3. No measurable decision A use case described as 'use AI on our support data' is not yet a use case. It does not say which decision changes, who makes it, or how you will know it is better. Gartner's January 2026 article names poor use-case selection combined with a lack of business value as the most consistent reason generative AI projects fail. Every AI build should name one decision or task it improves and one metric that shows it, such as time to resolve a ticket, forecast error or hours spent on manual review. If the metric cannot be measured today, the baseline cannot be set, and success cannot be shown later. ### 4. No path from pilot to production A pilot runs on an extract, in a notebook, with a small group of friendly users. Production needs live data feeds, access controls, integration with the systems people already use, monitoring, and someone responsible when it breaks. When none of this is planned at the start, the pilot succeeds on its own terms and is never deployed. The steps from a trained model to a live service are covered in our guide to model deployment. ### 5. The problem is the wrong shape for AI Some problems are too hard for current AI, and some are better solved without it. RAND's August 2024 report notes that AI is not a magic solution and that some projects fail because the problem is too difficult for AI to solve. Cutter's May 2025 article lists overreach among its causes. A narrow, well-defined task with good examples is a far safer first build than a broad ambition. Our explainer on narrow AI sets out why focused systems tend to succeed. > Most AI projects are decided before the build starts. The model is rarely the reason they fail. ## A twelve-point pre-mortem A pre-mortem asks you to assume the project has already failed and to work out why, before you spend the budget. The twelve checks below cover the five causes. Answer each one yes or no, and be strict: a partial answer is a no. ### Data - The data the use case needs exists, and you can name the systems it lives in. - Key business terms, such as customer, order or active account, have one agreed definition across those systems. - The data is current enough for the decision, and you know how often it refreshes. - Someone has checked its quality against this use case, not only in general. ### Ownership and value - A named person in the business owns the outcome and has the authority to change the process. - The decision or task the AI will improve is written down in one sentence. - There is a metric for that decision, and you have measured today's baseline. - You know what result would make you stop, as well as what result would make you continue. ### Production - The people who will use the output have been involved, and you know where in their work it will appear. - There is a planned route from the pilot environment to live data, live systems and governed access. - Monitoring, support and retraining have an owner after go-live. - The task is narrow enough that a first version can ship and prove its value within a bounded scope. Count one point for each yes. Twelve out of twelve is rare and not required. Each no is a risk you can name and fix. Several no answers in the data group mean the build should wait until the foundation is ready. Any no in the ownership group means the work has no one to receive it, however good the model is. It also helps to know where you stand overall. Our guide to AI maturity models describes how organisations move from ad hoc experiments to managed production, and which checks matter most at each stage. ## What a successful project does differently Projects that reach production and stay there tend to share a sequence rather than a technology. - They fix the data first. The data foundation, with one governed definition of the business, is in place before the use case is built on it. - They start narrow. The first use case is chosen because it is measurable and the data supports it, not because it is the most ambitious. - They name the owner and the metric at the start, and they agree the baseline before building. - They plan for production from the first day. Access, integration, monitoring and support are part of the scope. - They stay past go-live. Models drift and data changes, so someone watches performance and acts on it, as described in our guide to model monitoring in production. ### Scope, build, stay We work in that order. Scope means assessing readiness and sequencing the work. Build means the foundation first, then governance, then the use case, tied to a real metric before anything ships. Stay means we remain past go-live, because that is when much of the value is decided. At 1NCE, AI automation for customer engagement reached 60% of tickets handled to a human standard. That result depended on the foundation underneath it, not on the model alone. ### A first win that pays for the second One use case in production, tied to a metric the business cares about, is a better plan than a programme of pilots. It builds the case, the data and the confidence for the next use case. Sometimes the honest answer is to stop. If the data is not ready or no one will own the result, the right decision is to fix that first or not to build. We have turned down work on that basis. ## How to rescue a stalled AI project A stalled project is usually recoverable, and the work it produced is rarely wasted. The rescue uses the same checks, in a fixed order. - Run the pre-mortem on the project as it stands, and mark each no honestly. - Find the owner. If no one in the business will take the result, pause until someone does. - Restate the decision and the metric, and measure the baseline if it was never taken. - Compare the data the pilot used with the live data production will use. The gap is often where the stall began. - Narrow the scope to the smallest version that can ship and show the metric moving. - Plan the route to production before resuming work on the model. In our experience, AI engagements tend to become data engagements, and stalled projects are no exception. Once the basics are in place, our guide to AI implementation strategy covers how to choose and sequence the next use case. ## Where to start Start with an honest view of readiness. The AI readiness assessment is a free, 16-question self-assessment that scores strategy, people, data and change on the spot, with no email gate. It takes the pre-mortem further and gives you a starting point you can defend to your leadership team. If the score points to gaps in data, start there, because the foundation decides what every later use case can do. If it points to gaps in ownership or value, start with the use case and the metric. In either case, choose one decision to improve, prove it in production, and use that result to fund the next one. ## Where this leaves you AI projects rarely fail because the model was wrong. They fail because the data could not carry the use case, no one owned the result, the decision was never measured, or there was no route from pilot to production. Each of these is visible before the build starts, which means each can be checked and fixed. Run the twelve checks before your next commitment. A no is useful information. Fix it, or decide not to build yet. Both are better than a second pilot that stalls for the same reasons as the first. Our AI strategy service turns these checks into a readiness and maturity assessment, a sequenced roadmap and a business case, designed by the team that delivers the build. We name what each result depends on, and we say if the foundation needs work first. For a quick view on your own, take the readiness assessment. If you would rather work through your own project with the people who would build it, Talk to us and you will get a personal reply, usually within one working day. ## Frequently asked questions Most AI projects fail for reasons decided before the build starts. The data cannot carry the use case, no one in the business owns the result, there is no measurable decision for the AI to improve, or there is no plan to move from pilot to production. Some fail because the problem is too difficult for AI. The model itself is rarely the cause, and each risk can be checked in advance. The 95 per cent figure is widely quoted. Trullion's September 2025 article, for example, is framed around the claim that 95 per cent of generative AI projects fail. Other sources give lower figures: Gartner's January 2026 article is framed around 50 per cent of generative AI projects. The difference is mostly definition, whether failure means never reaching production, never returning value, or being abandoned. Check the definition before quoting any figure. AI magnifies whatever it sits on. Over reliable, well-defined data with a clear owner and metric, it works. Over inconsistent data, vague goals and no route to production, it fails faster. Pilots often succeed on a clean extract with friendly users, then stall when they meet live data, real users and governance requirements that were never planned. There is no single agreed rate. Gartner's January 2026 article on generative AI is framed around a 50 per cent failure rate, while Trullion's September 2025 article uses 95 per cent. RAND's August 2024 report focuses on root causes rather than one figure. Each source defines failure differently, so quote the source, the date and the definition together. AI fails in business when it is treated as a technology purchase rather than a change to how a decision is made. Gartner's January 2026 article names poor use-case selection combined with a lack of business value as the most consistent reason. IBM's October 2025 article warns that AI built without business stakeholders is guaranteed to fail. A named owner, a measured baseline and data that supports the use case address both. ## Frequently asked questions ### Why do AI projects fail? Most AI projects fail for reasons decided before the build starts. The data cannot carry the use case, no one in the business owns the result, there is no measurable decision for the AI to improve, or there is no plan to move from pilot to production. Some fail because the problem is too difficult for AI. The model itself is rarely the cause, and each risk can be checked in advance. ### Is it true that 95% of AI pilot projects fail? The 95 per cent figure is widely quoted. Trullion's September 2025 article, for example, is framed around the claim that 95 per cent of generative AI projects fail. Other sources give lower figures: Gartner's January 2026 article is framed around 50 per cent of generative AI projects. The difference is mostly definition, whether failure means never reaching production, never returning value, or being abandoned. Check the definition before quoting any figure. ### Why does AI fail so much? AI magnifies whatever it sits on. Over reliable, well-defined data with a clear owner and metric, it works. Over inconsistent data, vague goals and no route to production, it fails faster. Pilots often succeed on a clean extract with friendly users, then stall when they meet live data, real users and governance requirements that were never planned. ### What is the AI project failure rate? There is no single agreed rate. Gartner's January 2026 article on generative AI is framed around a 50 per cent failure rate, while Trullion's September 2025 article uses 95 per cent. RAND's August 2024 report focuses on root causes rather than one figure. Each source defines failure differently, so quote the source, the date and the definition together. ### Why does AI fail in business? AI fails in business when it is treated as a technology purchase rather than a change to how a decision is made. Gartner's January 2026 article names poor use-case selection combined with a lack of business value as the most consistent reason. IBM's October 2025 article warns that AI built without business stakeholders is guaranteed to fail. A named owner, a measured baseline and data that supports the use case address both. --- Shipshape Data is a London AI consultancy: we build the data foundation your AI depends on, then the AI on top. Site guide for agents: https://shipshapedata.com/llms.txt | Contact: hello@shipshapedata.com