A recurring paradox

AI adoption is growing, but economic results lag behind. According to McKinsey's The State of AI report cited by AI4Business, 44% of organizations say they are scaling AI enterprise-wide, up from 38% a year earlier, while the share of those attributing at least some impact on operating margin to the technology remains at 37%, practically unchanged. Companies deriving at least 5% of EBIT from AI account for 6%, the same percentage as in 2025.

PwC's survey of 4,454 CEOs, presented in January, reports that 56% have not seen significant financial benefits and that only 12% have achieved both cost reduction and revenue growth. In Italy, the AI market reached €1.8 billion in 2025, with 71% of large companies engaged in at least one project, and about one company in three struggles to estimate costs and benefits before starting, according to the Artificial Intelligence Observatory of the Politecnico di Milano.

The discipline of stopping

The Gartner survey published on September 1, covering 1,303 companies with at least $50 million in revenue, shifts attention to governance. Only 22% have extended AI across multiple business units, and 11% do not know how much their own function spent on AI in 2025. Organizations that track returns and stop weak projects report positive results on 81% of initiatives, while the least effective ignore the return on nearly one project in three.

Gartner had forecast that more than 40% of agentic AI projects will be canceled by the end of 2027 due to rising costs, uncertain value or insufficient risk controls. According to the analysis, each of these causes is visible in advance if the project starts with a written baseline and exit criteria. The weak point would therefore not be the technology chosen, but the absence of a documented starting point and a closing condition.

Measuring the cost of the work performed

The article cites two opposite cases. Commonwealth Bank of Australia had justified 45 job cuts with a voice assistant, only to backtrack in August 2025 after the union showed call volumes were rising. Klarna had announced an assistant capable of handling two-thirds of customer conversations, with an estimated $40 million improvement in profits, but in May 2025 the CEO admitted that the cost drive had hurt quality and the company resumed hiring.

On the measurement front, the analysis notes that model price lists are quoted in dollars per million tokens, whereas companies pay for tasks performed. Artificial Analysis introduced cost per task: in the tables consulted in October, GPT-6 Astra and Claude Fable 5.1 achieve the same score of 53 but cost $3.26 and $7.63 per task respectively, while DeepSeek V4 Flash falls to 22 cents with 18 fewer points. For a business case, the calculation should be extended to cost per completed case, including failed attempts and human review time.

The most cited figure of the past twelve months—95% of pilot projects with no measurable impact on the income statement, attributed to the MIT NANDA report—should be handled with caution: according to the material, the report defines success as marked and sustained impact and is based on interviews, not financial statements, so much so that the authors themselves speak of indicative figures. The causal links between measurement discipline and results therefore remain to be verified: the surveys cited describe correlations between governance practices and declared success rates; they do not prove that one produces the other.

Source and attribution: this original summary was prepared by AICronaca based on the report by AI4Business ↗. The statements are attributed to the source; any information limitations: The figures come from surveys and statements reported by the source, not from audited financial statements. The MIT NANDA report is described as based on interviews and with indicative numbers. Gartner's forecasts on the cancellation of agentic projects by 2027 are estimates. The material contains no data on Italy beyond those from the Politecnico di Milano Observatory..