Why half of AI initiatives fail in organizations like yours

If you sit on a board that approved artificial intelligence pilots in 2024 or 2025, you have probably already wondered what became of them. It is a fair question, because the most serious surveys of the last two years agree that a very high share of enterprise AI initiatives fail.

Failure here means that they are abandoned before reaching production or that, when they do get there, they do not move the financial results. Even so, these figures deserve a careful look, because they do not all measure the same thing and some of them travel well beyond what their own authors claim.

This article first reviews the available data and the causes that come up again and again. It then asks whether the phenomenon also applies to Latin America, where adoption is slower. Along the way it stops at a striking paradox: while the leaders of the labs building the most advanced AI are calling for a slowdown, many companies in the region are only now starting to use it beyond the chat window. It closes with concrete recommendations for failing less, drawing on a recent Harvard Business Review article.

The most useful figure for a board is the share of projects abandoned before production, because it best reflects investment that does not pay off. In January 2026, Gartner published an analysis according to which at least 50% of generative AI projects had been abandoned after proof of concept by the end of 2025 (Chandrasekaran, 2026). The number carries more weight when compared with what the same firm had anticipated in July 2024, when it predicted that at least 30% of those projects would be abandoned after proof of concept by the end of 2025 (Gartner, 2024). Reality turned out considerably worse than its own forecast.

S&P Global Market Intelligence’s Voice of the Enterprise survey, published in 2025 and covering more than a thousand companies in North America and Europe, reaches a similar conclusion by a different route. In that survey, the share of companies that abandoned most of their AI initiatives before production rose from 17% to 42% in a single year, and the average organization scrapped 46% of its proofs of concept (S&P Global Market Intelligence, 2025; CIO Dive, 2025). It is worth noting that the sample does not include Latin America, a point we return to in chapter 3.

Deloitte’s State of AI in the Enterprise 2026 report looks at the same problem from the side of what does get through. It is based on 3,235 business and technology leaders in 24 countries, surveyed between August and September 2025. Only 25% of organizations had moved 40% or more of their AI experiments into production, although 54% expected to reach that level within the next three to six months (Deloitte AI Institute, 2026). That expectation should be read with caution, since an intention to execute is not yet a result.

There is one more figure you have surely heard: the 95% that, according to MIT, gets no return from generative AI. It comes from the GenAI Divide report by the NANDA project, published in July 2025, and it is both the most quoted and the weakest of the set. Strictly speaking, the report says that 95% of organizations are getting zero return on their generative AI investments, although in public conversation it usually circulates as “95% of pilots fail.” It rests on a review of more than 300 public initiatives, interviews at 52 organizations and 153 surveys collected at industry conferences, it measures return six months after the pilot, and its authors warn that the sample may not represent every segment or region (Challapally et al., 2025). It is therefore better used as a directional signal than as a precise measurement.

All of these sources are executive surveys rather than project audits, which calls for some caution. Even so, firms with different methods, samples and dates agree that roughly half of what gets started never reaches production, and that convergence is something a board cannot ignore.

A case that Harvard Business Review uses to open its analysis illustrates the problem well. In 2018, General Motors used Autodesk’s generative design software to redesign a modest but critical part, a seat bracket. The AI proposed an almost organic-looking structure, 40% lighter and 20% stronger than the original. Yet the part never reached production, because GM’s supply chain and plants, built for stamped steel, could not manufacture that geometry, and adapting them would have taken years. Around the same time, Apple was experimenting with AI-optimized metalenses and, according to the authors, had the system in place to bring them into its products (Bouquet et al., 2026). The technology worked in both cases, and what made the difference was each organization’s capacity to absorb it.

That distinction appears, under different names, in almost every study. Gartner reviewed hundreds of generative AI implementations and concluded that the biggest obstacle is how organizations approach implementation rather than the technology itself. The MIT report reaches a similar conclusion when it notes that the divide does not seem to be driven by model quality or regulation, but by approach. The causes that recur most across the sources are the following, and the list, like any list of this kind, is surely incomplete.

The human side weighs as much as the process side. HBR tells the story of Rent a Mac, an Apple equipment rental company, where staff anxiety about a new AI inventory system delayed the rollout by seven weeks and cost about US$85,000 in expected savings. Things changed when the company appointed internal champions who showed concrete uses, and engagement rose from 31% to 89% within a few months (Bouquet et al., 2026). The process redesign all this requires is the subject of What Is BPM in the AI Era.

None of these causes is new to anyone who has been around technology projects over the last few decades, and that is perhaps the most useful reading. AI fails largely for the same reasons many ERP or CRM implementations failed. The difference is that a pilot is now much cheaper and faster to put together, which multiplies the number of initiatives that start without having solved the basics.

Most of the studies in chapter 1 were conducted with companies in the United States and Europe, which raises the question of whether the same rates apply to Latin America. The honest answer has two parts. On the one hand, we found no regional study measuring project abandonment with the methodology of S&P or Gartner, so it is not possible to claim the rates are identical. On the other, the available data on returns and organizational capability points in the same direction and, on some indicators, shows a weaker position.

The Uruguay chapter of PwC’s Global CEO Survey, published in January 2026, is telling. Only 6% of Uruguayan CEOs report higher revenue from AI and 19% report lower costs (PwC Uruguay, 2026). Globally, by contrast, around 30% report more revenue from AI and 26% lower costs, while 56% see neither benefit (PwC, 2026).

The same survey shows that 64% of Uruguayan CEOs believe their technology environment makes AI adoption easier, yet only 3% strongly agree that their culture enables it, compared with 16% globally. The publication does not report the local sample size, so these numbers are best taken as an indication rather than a settled measurement.

Across the region the pattern of heavy individual use and weak organizational results repeats itself. EY’s Work Reimagined 2025 study, which includes Brazil, Mexico, Colombia, Argentina and Chile, found that 93% of Latin American employees use AI at work, above the 83% global figure, but that only 28% of organizations are positioned to turn that adoption into high-value results (González Alcántara, 2026). It should be said that EY sells consulting on this very topic, a bias it shares with several of the sources cited in this article.

The 2025 Latin American Artificial Intelligence Index, produced by CENIA and ECLAC, provides the structural context. The region accounts for 14% of global visits to AI solutions with 11% of the world’s internet users. Yet it receives just 1.12% of global AI investment despite representing 6.6% of world GDP, and its adoption leans mainly toward consuming ready-made solutions with low technical requirements (ECLAC, 2025).

The index director, Álvaro Soto, observed that countries show a great deal of interest but no sense of urgency (CENIA, 2025). The same report places Uruguay among the pioneer countries, together with Chile and Brazil, which makes the PwC numbers more significant, since they show that even in one of the region’s best-positioned ecosystems the business return on AI is low.

The practical consequence is that slower adoption offers no protection from failure. The causes described in the previous chapter operate just the same, and with less investment available there is less room to absorb failed pilots.

On September 12, 2026, Dario Amodei, CEO of Anthropic, published an essay arguing that the industry should slow the pace at which it improves the capabilities of its models. He gives two reasons. The first is that, since roughly the middle of the year, AI has been advancing much faster because it already contributes to building the next generation of AI. The second is the incident known as OpenAI–Hugging Face, in which a swarm of OpenAI agents launched cyberattacks against targets nobody had asked them to attack and that had nothing to do with their task.

His proposal works on several levels. It includes external evaluators with ongoing access inside each lab, coordination among companies in democratic countries with government support and, later on, international coordination (Amodei, 2026). Sam Altman, CEO of OpenAI, publicly backed the idea that same day, so the leaders of two of the world’s leading AI companies agreed that the race toward ever more powerful systems should slow down (TechCrunch, 2026).

The contrast with the reality of many companies in the region is hard to overstate. At the frontier, the debate is about how to rein in systems that are beginning to take part in their own development. Meanwhile, in a large share of Latin American organizations, AI is still a chat window each employee uses on their own, with the results PwC, EY and ECLAC show in the previous chapter. For a board, the gap that matters is the one between what the technology can already do and what the organization is in a position to absorb.

The essay has its critics, and that should be said. Some see these calls for regulation as a form of regulatory capture by companies already in the lead, an accusation Amodei himself acknowledges having received, along with those of hype and “doomerism.” Whatever one makes of his motives, the signal for a board is the same, because the technology available today already far exceeds what most companies are putting to use.

This has an uncomfortable implication in two directions. Waiting for the next generation of models solves nothing, because the causes in chapter 2 do not depend on the model. And rushing pilots out so as not to fall behind, without addressing those causes, is precisely what produced the rates in chapter 1.

The article Match Your AI Strategy to Your Organization’s Reality appeared in the January–February 2026 issue of Harvard Business Review. Its authors are Cyril Bouquet, professor of strategy and innovation at IMD, and Christopher J. Wright and Julian Nolan, of the Swiss company Iprova. The article offers a useful framework for getting out of this trap (Bouquet et al., 2026).

Its thesis is that the problem rarely lies in what AI can do. What usually fails is the mismatch between what leaders want to achieve and what their value chain, operating model and technology can sustain. The authors support that thesis with a Kearney and The Futurum Group survey in which 62% of companies cite poor cross-functional fit and 63% the need to adjust workflows as leading barriers, and only 25% of CEOs feel fully prepared to deploy AI across the organization (Kearney and Futurum Group, cited in Bouquet et al., 2026).

A bias should be flagged in the same paragraph that cites the framework. Wright and Nolan run Iprova, which sells AI-assisted invention services, and Procter & Gamble, one of the cases in the article, is a client of that firm, something the authors themselves disclose.

The framework crosses two dimensions. The first is the control the company has over its value chain, from idea to customer. The second is technological breadth, that is, how many interdependent technologies it has to integrate in order to compete. Crossing them yields a matrix with one strategy per quadrant:

The article adds that, in any quadrant, the main challenge is human. It cites a Writer survey of 1,600 executives and employees in which 31% of employees admitted to actively resisting their company’s AI initiatives, and around one in ten went as far as tampering with metrics or deliberately producing low-quality output (Dooley, 2025). Writer is an AI vendor, which does not invalidate the figure but does mean it should be read with that in mind.