“We hope, at least, to have paved a path for the future,” says Mario Venturi, the CEO of KeyCrime, reflectíng on the closure of his company, which is now in liquidation.

A former police officer in Milan, Venturi founded the company after years of service. His speciality was identifying patterns in specific crimes, such as robberies. He introduced systematic data collection and standardized the interviews conducted with victims. Using statistical methods, "the application takes data from the latest crime and compares it with other crimes on the search for similarities. If similarities are found, they are presented to the end users." Venturi initiated the development of Italy's first predictive policing software, DELIA, which stands for “Dynamic Evolving Learning Integrated Algorithm.” The Milan police department started trials in 2008, making the software one of the most established digital tools in Europe in this field. It was managed by KeyCrime.

Italian media began speculating early on that KeyCrime might not survive implementation of the Artificial Intelligence Act. Wired Italia, for instance, described the company as being "on the brink of collapse," attributing its downfall to the new regulation, officially adopted by the European Union this year. The AI Act has imposed particularly strict requirements on predictive policing systems, classifying them as "high-risk" technologies.

Venturi insists that the system he developed fully complies with the AI Act's standards. However, to be able to continue developing the software and selling it to police departments across Italy and Europe, the company would have to wait for approval from the European Union. Earlier this year, when debates about the AI Act's provisions were still ongoing, and as the European Parliament and the European Council reached a provisional agreement on the precise wording, Venturi grew increasingly concerned about a potential block on his technology. "The regulatory proposal was released in July, then postponed in December. We would have had to wait for approval from the European Council and then the Parliament. We are a startup: If we don't sell, how can we survive?," the former police officer asks, disillusioned. 

Still, the company’s financial struggles are hardly new. Even before the drafting of the AI Act got underway, KeyCrime had already begun experiencing economic losses.

A Failed Project

Founded in 2009, KeyCrime started with a capital of €100,000. During its first three years, it incurred only losses, bringing in virtually no revenue. After initially shutting down in 2011, the company relocated to Milan and resumed operations in 2019, securing around €1.2 million from the Italian venture capital firm Oltre, the investor Sdg Group, and the entrepreneur Giorgio Gandini. "Unfortunately, we weren't very lucky," Venturi explains, "because we spent the first year and a half engineering the software. We were about to start the sales process when COVID hit."

COVID-19 hit Italy hard in early 2020, with the pandemic ultimately spreading to all parts of the globe. The world came to a standstill, and so did KeyCrime. However, this phase also coincided with growing hype around predictive policing algorithms. So much so that, in 2020, the Department of Public Security inside Italy’s Interior Ministry hinted that it planned to develop a national predictive policing system based on KeyCrime's experience, called Giove (Jupiter). By 2023, the Central Anticrime Directorate of Italian National Police stated that the Giove project was not yet operational and still in the "feasibility study" phase, according to information provided after a transparency request was filed.

Testing the System

KeyCrime claimed to have partnered with the Spanish firm PSS in an effort to enter the South American market. However, there is no indication that DELIA  has been used in Spain or any South American countries. Spain relies on its own, in-house predictive policing algorithms.

According to Venturi, DELIA has been particularly “successful” since its implementation. A series of press releases and statements from the Milan Police Headquarters praise the effectiveness of the technology, but neither the law enforcement body nor the Ministry of Interior have provided data to back the efficacy of the software. Unlike other systems of its kind, such as PredPol, which is used in the U.S. and relies on hotspot analysis (a technique that looks for clusters of crime locations), KeyCrime developed a technique known as crime linking, designed to identify recurring patterns in serial crimes.

The growing interest from Italian state agencies in Artificial Intelligence applications across various sectors − including security, education, welfare, and taxation − highlights a significant gap between the academic and corporate worlds. Companies with more resources can afford to develop these applications, but their systems remain closed to scientific scrutiny. This concern is shared by Serena Quattrocolo, professor of criminal procedural law at the University of Eastern Piedmont. "Unfortunately, we lack the resources and sufficient pressure to incorporate legal training into the design and testing of these applications. There is a missing dialogue between the two sides − and companies, with their greater firepower, have the freedom to do as they please, especially in the Italian context where predictive policing is not yet regulated by law."

Lots of Data, No Evidence

The official KeyCrime website, now offline, claimed that use of the software led to a reduction in retail robberies of about 58 percent from 2008 to 2017 in Milan, with a decrease in bank robberies of 89 percent from 2009 to 2017 in the Milan area. The site also stated that in the Milan area trial, DELIA "was evaluated by the Italian Ministry of Justice," although this claim could not be verified. An independent 2019 study by Giovanni Mastrobuoni, professor of economics at the University of Essex, suggested that when comparing efforts between the two Italian law enforcement agencies – the Polizia, which used DELIA, and the Carabinieri, which did not – "the solve rate of subsequent robberies in the Polizia sector was 8 percentage points higher than in the Carabinieri sector.”

It is worth noting that robbery and theft rates in Italy have been steadily declining for the past decade. From 2008 to 2017, overall crime rates fluctuated, but they dropped significantly starting in 2014. By 2017, 280,000 fewer crimes were committed than in 2008, a 10.3 percent decrease. In 2017, robbery and theft rates dropped by 7.1 percent and 6.0 percent, respectively, compared to 2015 – a trend observed across Europe. 

In 2022, Italy's National Institute of Statistics reported that in 2020 and 2021, public perception of safety improved in local areas, with the restrictive measures imposed during the pandemic leading to a sharp decline in predatory crimes (home burglaries, pickpocketing and robberies). These crimes reached their lowest levels ever recorded in 2020. In 2021, as mobility and social contact restrictions eased, there was a slight increase, but the numbers remained well below pre-pandemic levels. From our research and due to the lack of transparency in the release of independent data on the system’s use, it was not possible to ascertain how DELIA has influenced the decline in crime rates in the Milan area. Given the global trend of crime rates remaining more or less steady, the effectiveness of any predictive policing system remains questionable.

The KeyCrime experiment, which grew into a highly anticipated project over the years, driven by the narrative that AI could solve every problem in the world, including crime, ultimately turned out to be a failure. It did, however, pave the way for the future. Other predictive policing software is being used in Italy, including Pelta Suite, deployed in small northern Italian cities since 2021, and Vigilium, tested in Trento and then adopted in Naples since 2019 and in various municipalities in the Emilia-Romagna region between 2020 and 2021. These experiments also take advantage of a gray area in Italian law, as there is no official definition of "predictive policing," nor any legislation governing the practice. The fact that these systems are still in operation could indicate that KeyCrime has just fewer means to survive the new AI Act’s rules.

This article was developed with the support of Journalismfund Europe and a grant from Investigative Journalism for Europe (IJ4EU).

A correction was made on Nov. 21, 2024: The article was updated to remove inaccurate information about early KeyCrime investors and to specify the type of crimes DELIA assesses, following a request from Mario Venturi.