# Why most AI projects fail — and what hands-on support changes

The question of **why AI projects fail** is now answerable with hard numbers, and the answer is rarely the technology. The 2025 MIT study "The GenAI Divide: State of AI in Business 2025" found that 95% of generative artificial intelligence pilot projects produce no measurable economic return for the company. Between 30 and 40 billion dollars were spent in the first half of 2025 on generative AI, with most companies recording no concrete impact on their income statement. Italian SMEs sit inside this scenario: only 8% of companies with more than ten employees use AI technologies, against a European average of 20%, [a gap that leaves Italy eighteenth in Europe for AI adoption](/en/ecorner/2023/ai-adoption-slowdown.html).

The issue is not a lack of resources. It is that the vast majority of AI investments turn into failed projects, spending dispersed without generating the value promised.

## The numbers behind the failure

Data from authoritative sources converge on an uncomfortable reality. The RAND study, based on interviews with 65 AI experts from different industries, confirms that more than 80% of AI projects fail to meet their goals — twice the rate of traditional IT projects that do not involve AI. Gartner expected 30% of generative AI projects to be abandoned after the proof of concept by the end of 2025, because of high costs, poor data quality and implementation risks. Klecha & Co. estimates that 40% of autonomous AI projects will be abandoned within two years.

In Italy the picture is sharper. ISTAT 2025 data show 16.4% of companies with at least ten employees using at least one AI technology, double the 8.2% of 2024. But the figure crumbles by company size: only 7% of small companies and 15% of medium companies have active projects. 58% of companies that evaluated AI investments gave up for lack of adequate skills: the ability to govern technical and organisational complexity, rather than budget.

## The five structural causes of failure

## Leadership: when management does not know what it wants

84% of the experts interviewed by RAND point to company leadership as the main cause of AI project failure. The problem is strategic rather than technical. Many companies treat AI as software to install, not as a strategic lever that requires redesigning processes. AI is a tool to integrate into [a wider business strategy](/en/ecorner/2026/digital-strategy-business-growth.html), not an add-on.

The problem shows up in vague objectives. Launching projects to experiment or to innovate ignores the decisive question: which specific process and which measurable KPI are we expecting to improve? When metrics stay vague, even brilliant progress goes unnoticed because there are no parameters to assess it. There is also a systemic disconnect between those who lead the company and those who build the technical solutions. Leaders often lack a deep understanding of AI's capabilities and limits, which leads to poorly defined goals and a poor handover to technical teams.

## Data: the fragile foundation that brings everything down

52% of RAND experts cite the quality and limited usefulness of available data as a critical obstacle. This is particularly damaging for Italian SMEs. The Artificial Intelligence Observatory of the Politecnico di Milano highlights that the most cited limit among SMEs is immature data management: data exist, but they are fragmented, sporadic and ungoverned, and internal technologies and skills to make them usable by AI are missing.

It is not only a question of quantity. It concerns quality, structure and governance. Data must be complete, traceable and unambiguous, organised so that AI can learn meaningful patterns. The Politecnico notes that consumption-based pricing for generative AI makes total costs hard to predict, and more than 50% of active companies find managing GenAI costs complex for this very reason. Legacy technology adds another layer: years of under-investment have produced fragmented systems, legacy applications that do not talk to each other and architectures unsuited to the computing volumes AI requires.

## Distorted expectations: chasing technology instead of the problem

A third of the identified causes concern the wrong approach to the project. Attention is too often on using the newest and most advanced technologies rather than on solving real problems for end users. Companies chase the wow effect, spectacular solutions on paper that turn out to be irrelevant in the accounts.

The MIT describes this as a learning gap: AI systems, useful at an individual level, often remain shallow wrappers or scientific exercises unless embedded in well-designed processes. Tools such as ChatGPT or Copilot are implemented without rethinking operational flows, without training staff properly, without integrating context, memory and adaptation to the company's specifics. Unrealistic timelines make it worse. Many organisations push for fast results, but fast results are not the same as good results. AI implies substantial work on data, testing and verification before releasing models that meet the need.

## Skills: the gap that cancels technology

58% of Italian companies that evaluated AI adoption gave up because they lack staff able to manage it. This skill gap is the most urgent structural criticality. Technology availability exceeds companies' ability to absorb and integrate it into production processes.

The problem is not only data scientists and specialised engineers. It concerns overall company culture. AI changes roles and habits: cultural resistance can block technically perfect projects. A change management plan that involves people from the start, explains the benefits and offers continuous training becomes indispensable. Without internal talent, the company depends entirely on external vendors and loses critical control: it cannot assess the impact of proposed solutions, guide development according to its specifics or supervise daily operations.

## Integration: when technology does not fit real work

MIT identifies the inability of AI systems to adapt to business workflows as one of the main causes of failure. This is not a model or infrastructure problem: the knot is the weak ability to store feedback, improve alongside real activities and integrate with existing processes. Superficial integration produces tools that stay outside the workflow, used occasionally by individual employees but never embedded in structural processes.

A significant phenomenon then emerges: unofficial use of AI tools by employees, so-called shadow AI. Tools such as ChatGPT are used informally, often beyond official company licences. This indirect use is changing work processes more than official projects, while staying outside IT and managerial control. It reflects a real demand for tools that deliver personal value, and it highlights the gap between internal demand for everyday intelligence and the company's ability to orchestrate effective solutions institutionally.

## The Italian case: what makes SMEs different

Italian SMEs face specific challenges. A widespread culture of under-investing in digital weighs heavily. Many owners, especially over 50, see digital as a cost to compress rather than a strategic investment. Past negative experiences with improvised consultants feed structural distrust.

The Politecnico Observatories show that about 45% of integrations between ERP systems and digital platforms fail because of misalignment between pre-existing business logic and the platform's technical constraints. The problem is accurate mapping of the information flows between systems that have layered up over the years without a coherent design, not the availability of APIs or connectors.

The budget framework of Italian SMEs is often unstructured: the issue is difficulty in assessing the costs and benefits of complex technology investments, not an absolute lack of resources. Consumption-based AI pricing makes forecasting even more opaque. This generates decision paralysis. The gap with Europe widens when looking at intensity of use: the difference between large companies and Italian SMEs grew from 20 percentage points in 2023 to 25 in 2025. Large Italian companies reach 53% AI adoption, SMEs stay at 15.7%.

## Where hands-on advisory changes everything

Strategic hands-on support in AI implementation for SMEs is not one option among many; it is the difference between joining the MIT 95% statistic and being part of the 5% of projects that generate measurable value. The complexity described above cannot be managed by improvising.

## Pre-implementation assessment: avoid the wrong project

Before choosing technologies or vendors, a thorough analysis of existing processes is needed. This is not about computerising what already exists, an expression that betrays a failing approach. It is about identifying real bottlenecks, mapping operational flows and understanding where time and margin are lost today.

A specialist partner runs this phase with established methods: interviews with key operators, analysis of process data where available, identification of gaps between the current situation and business goals. The result is a business case with measurable ROI, not generic references to efficiency or innovation: a concrete reduction in operating time on specific activities, hours per month recovered for higher-value work, a cut in errors on a critical process. This phase prevents the most common mistake: investing in AI for problems that do not exist or that need organisational rather than technological work.

## Data complexity management: building the foundation

Data is the main failure factor for Italian SMEs. Professional support starts with a thorough audit of the quality of available data: completeness, accuracy, structure and accessibility. Many companies discover at this stage that critical data is spread across unintegrated systems, incomplete and non-standardised.

Data governance becomes a priority. This includes procedures for collecting, validating and updating data, compliance with the GDPR and the AI Act, and architectures that integrate with legacy systems without replacing the whole infrastructure. A representative case: a manufacturing company with an old ERP, separate order management software and spreadsheets for analysis. Direct AI implementation would have failed for lack of coherent data. Support planned an intermediate phase of partial integration, extraction of critical data, cleaning and standardisation. Only afterwards came the gradual introduction of predictive algorithms on usable datasets.

## Gradual implementation: quick wins and scalability

The big bang approach, trying to transform the whole organisation with a single ambitious AI project, is an almost certain route to failure. Professional support favours quick wins: projects of value deliverable in three to six months, with reasonable effort, on specific processes. These pilots need precise characteristics: measurable goals, visible impact for stakeholders, manageable complexity. Their success builds internal credibility, unlocks further budget and creates organisational momentum. Scalability is designed from the start but implemented in successive phases, with a realistic horizon and verifiable milestones.

## Skills transfer: building gradual autonomy

The final goal of support is to build progressive autonomy in using and managing the implemented AI solutions, not to create dependence on the external vendor. This requires operational training of the internal team, with concrete work on the specific tools adopted rather than generic theory. Cultural change management accompanies the technical training: people need to understand why the AI tool is useful for their daily work, and how it works. Building a core of internal skills, even a small one, lets the company keep governance over the project and reduces the disconnect identified as a primary cause of failure.

## Continuous monitoring: from implementation to iterative improvement

An AI model launched and left alone quickly becomes obsolete. Data change, the market evolves, business processes shift. Treating AI as a one-shot project means losing contact with operational reality and watching performance degrade. Professional support includes continuous monitoring of implemented solutions: verifiable KPIs tracked over time, iterative improvement cycles, feedback collection, dataset updates, model refinement and adaptation to process changes. The ability to intervene promptly when performance falls, scale what works and abandon what does not is a decisive competitive advantage.

## The signals that call for support

Some situations make professional support indispensable rather than advisable. A budget under 50,000 euros on a project that affects critical core processes: the temptation to handle it internally hides high risks, because an error on a critical process costs far more than the investment in external skills. Total absence of data scientists or internal AI skills: improvising almost certainly means failing. Fragmented IT infrastructure, with three or more management systems that do not talk to each other: integration requires system integration skills that SMEs rarely have. Management that cannot distinguish machine learning from generative AI: that knowledge gap will lead to wrong choices and dispersed investment. Projects launched to experiment without a defined problem, measurable KPIs or verifiable hypotheses: these always end among the abandoned. A previous failed AI implementation with significant sunk costs: repeating the same mistakes without a deep analysis of the causes wastes further resources.

## Beyond the statistics: from theory to practice

A concrete example clarifies the point better than any statistic. A mid-sized retail company in food, 15 million euros in revenue, faces growing inefficiencies in warehouse logistics. Orders, supplier management, stock rotation: everything works but with continuous loss of time and margin. The owner considers AI to optimise the warehouse.

A do-it-yourself approach would have meant buying warehouse software with built-in AI, a fast rollout and patchy training. The predictable result: software that does not talk to existing systems, dirty data feeding the algorithm with unusable suggestions, frustrated staff abandoning the tool after a few weeks. Project closed after six months and 30,000 euros spent with no results.

The supported approach: three months mapping processes. The real bottlenecks emerge: badly designed operational flows, not technology, with manual procedures repeated because it has always been done that way, and critical data recorded in personal spreadsheets that are not shared. Four months of pilot on a specific department: partial integration of existing systems, cleaning of the historical order dataset, a predictive algorithm on stock rotation limited to high-turnover categories, operator training and daily performance monitoring with progressive adjustments.

Result after nine months: 30% reduction in operating time on order planning activities, measured and documented. Over-stock errors down 40% on the pilot categories. Operators actively using the system because they see concrete benefit in their daily work. Budget invested: 45,000 euros, ROI reached in 14 months.

The difference is not technological. Both approaches could have used similar tools. The difference is methodological: skill in mapping processes, experience in data management, the ability to size the project correctly, change management support and continuous monitoring. The question is no longer whether to implement AI. It is how to avoid becoming part of the MIT 95%. The 2025 data show that investing capital is not enough: it takes method, skills and structured support.

## Frequently asked questions

**Why do 95% of AI projects fail according to MIT?**
The MIT study "The GenAI Divide" found that 95% of generative AI pilots produce no measurable return. The causes are organisational rather than technological: weak leadership, poor data quality, distorted expectations, skill gaps and superficial integration into workflows.

**What are the five main causes of AI project failure?**
Leadership that cannot define measurable objectives, data that is fragmented and ungoverned, expectations focused on technology rather than on the problem, lack of internal skills, and AI tools that do not integrate with real business processes.

**What does hands-on advisory change in an AI project?**
It adds a pre-implementation assessment with a measurable business case, an audit of data quality and data governance, a gradual implementation based on quick wins, training that builds internal autonomy and continuous monitoring with iterative improvement.

**Why do Italian SMEs struggle more with AI adoption?**
Only 7% of small Italian companies have active AI projects, against 53% of large companies. The main barriers are immature data management, a culture of under-investing in digital, difficulty in assessing costs and benefits, and a lack of internal skills.

## Sources

MIT, "The GenAI Divide: State of AI in Business 2025" — https://www.media.mit.edu/projects/nanodegree/overview/
RAND Corporation, "The Root Causes of Failure for Artificial Intelligence Projects" — https://www.rand.org/pubs/research_reports/RRA2680-1.html
Gartner — generative AI project abandonment and proof-of-concept data — https://www.gartner.com/en/newsroom
ISTAT — enterprises and ICT, 2025 data on artificial intelligence — https://www.istat.it/it/archivio/imprese
Osservatori Digital Innovation, Politecnico di Milano — artificial intelligence and data maturity in Italian companies — https://www.osservatori.net