Why do AI projects fail?
The short answer: AI projects fail mainly for organisational reasons, not technical ones. The RAND study identifies five root causes of failure: problems with leadership and problem definition, poor data quality, a focus on technology rather than the business, insufficient infrastructure for deployment, and applying AI to problems beyond its current capabilities. Only one of the five is purely technical. In fact, 84% of private-sector interviewees pointed to leadership problems as the main cause.
The most recent data point in the same direction. MIT's The GenAI Divide report, analysed by Forbes, concludes that 95% of generative AI pilots in companies generate no measurable impact on the profit and loss account. And Gartner predicted that at least 30% of generative AI projects would be abandoned after the proof of concept due to poor data quality, inadequate risk controls, escalating costs or unclear business value. The Spanish context adds urgency: according to the INE ICT survey, 21.1% of companies with 10 or more employees already use AI, 8.7 points more than the previous year. More and more companies are investing, and more and more are exposed to these mistakes.
Strategy and leadership mistakes
1. Starting with the technology rather than the problem
This is the most common mistake. The project is born from the question "what can we do with AI?" instead of "what business problem do we need to solve?". RAND describes it as chasing the latest technology instead of solving users' real problems. The result is usually an eye-catching demonstration that fits into no process. At LurNova we always start with AI consultancy: before talking about models, we identify which processes have friction, what data exists and where AI can add measurable value.
2. Not defining objectives or success metrics
A project without clear indicators cannot officially fail, but nor can it succeed. Gartner points to unclear business value as one of the four main reasons generative AI projects are abandoned. If nobody has defined what "working" means (reduced response time, hours saved, resolution rate, quality of output), the initiative is judged on impressions. And impressions, once the initial enthusiasm fades, tend to decline. Define metrics before you start and measure them from day one. It is also worth distinguishing between technical metrics (accuracy of answers, processing times) and business metrics (hours saved, cases resolved, cost per operation). A project can do well on the former and contribute nothing on the latter, which are the ones that justify the investment to a management committee.
3. Unrealistic expectations from management
Leadership can sink a project through excess or through neglect. Through excess, when it expects transformative results in weeks or asks AI to solve problems beyond the current state of the technology, another of the root causes identified by RAND. Through neglect, when it delegates the initiative entirely to the technical team and does not take part in the business decisions the project requires. Well-applied AI can improve processes significantly, but it requires realistic timelines, sustained budget and engaged management.
Data and infrastructure mistakes
4. Underestimating data quality and availability
Many projects fail because the organisation does not have the data it needs, or because the data it has is incomplete, scattered or ungoverned. It is the second root cause in the RAND study and the leading reason for abandonment according to Gartner. An AI agent using RAG technology, for example, is only as good as the documentation it works on: if the knowledge base is out of date or poorly structured, the answers will reflect it. Auditing the data before committing to a timeline is one of the most profitable investments in the whole project.
5. Getting stuck in the eternal pilot
Some companies chain proofs of concept without ever taking anything into production. The MIT report captures it well: high adoption, low transformation. The pilots that do generate value are those integrated into real workflows, with memory, learning and connection to the company's systems. Moving from pilot to production requires planning from the outset for integration with existing systems, scalability and maintenance. A useful proof of concept is not one that shows the technology works in the abstract, but one that validates the use case in conditions close to real ones, with real data and real users. If the project plan ends at the demo, so does the project.
6. Not planning the deployment infrastructure
RAND identifies a lack of investment in infrastructure as a root cause of failure: the organisation builds a model it then cannot deploy, maintain or monitor. This decision includes an increasingly relevant question: where the AI runs and who controls that environment. Depending entirely on an external platform means accepting its changes in pricing, terms and availability. That is why at LurNova we also work with private AI models and the client's own infrastructure, so that the company retains operational control of a technology that is becoming critical.
People mistakes
7. Not involving end users
The purchasing decision is usually made by general management, IT or digital transformation. The people who will use the tool eight hours a day rarely take part in its design. The typical outcome is a technically correct system that nobody uses. RAND links this to its first root cause: misunderstandings about the intent and purpose of the project between those who commission it and those who will use it. Involving users from the definition phase, and not just in the final training, multiplies the chances of real adoption.
8. Neglecting training and change management
An AI tool without people trained to use it with judgement is an expensive licence. AI training is not just about teaching people to write prompts: it includes understanding what the system can and cannot do, how to verify its answers and how to integrate it into daily work. The EU AI Act, moreover, has required since February 2025 that organisations ensure a sufficient level of AI literacy among their staff. What we have seen with our clients is that projects with a structured training plan achieve adoption much sooner than those that trust that "people will pick it up".
Security and compliance mistakes
9. Ignoring data security and sovereignty
When the company does not provide secure tools, employees use whatever they find. This is the phenomenon of shadow AI: the use of AI tools without the organisation's approval or oversight. According to IBM's Cost of a Data Breach 2025 report, 20% of organisations suffered security breaches linked to this unauthorised use, and those with a high level of shadow AI incurred on average an additional 670,000 dollars per breach compared with those that barely use it. Entering corporate data into generic tools without guarantees of traceability or access control is a risk no serious project can afford. Our premise at LurNova is precisely that: artificial intelligence, if it is not secure, is not intelligent.
10. Forgetting regulatory compliance
Regulation (EU) 2024/1689, known as the EU AI Act, has applied generally since 2 August 2026. Transparency obligations require informing customers, workers and users when they interact with AI systems or receive artificially generated content, and high-risk systems require specific quality management systems (obligations whose general application has been postponed to 2 December 2027). Penalties for prohibited practices can reach 35 million euros or 7% of annual worldwide turnover. Designing the project without taking this framework into account means redoing work, or worse: withdrawing it. Compliance is not a final formality, it is a design requirement.
How to turn the statistics around
The optimistic reading of all this data is that the causes of failure are known and avoidable. McKinsey's State of AI report confirms it from the opposite angle: although 88% of organisations already use AI, those achieving real impact on results share a pattern, which is that they redesign their workflows instead of layering the technology on top of existing processes. In other words, they do exactly the opposite of the mistakes on this list: they start from the problem, involve people, look after their data and treat AI as a business transformation, not a software purchase.
There is an eleventh habit that separates the projects that last from those that fizzle out: measuring, iterating and maintaining after launch. AI systems are not software you install and forget. Data changes, processes change and the system's performance degrades if nobody watches it. A project that does not set aside budget and owners for the post-deployment phase is repeating, with a delay, the infrastructure mistake: reaching production with no plan to stay there.
At LurNova we support that whole journey: consultancy to identify use cases with a real return, training so that teams use AI with judgement, agents with RAG technology that work on your company's knowledge and private models deployed on infrastructure you control. If you are planning an AI project and want to avoid these ten mistakes from the design stage, let's talk. A timely conversation costs far less than a failed pilot.
CEO of LurNova, with extensive experience in technology and digital transformation. Involved in developing technology projects and applying artificial intelligence in organisations, he combines technical knowledge with business vision.
