Custom AI software: built to fit or bought off the shelf?

The Italian artificial intelligence market reached 1.8 billion euros in 2025, growing 50% on 2024. Custom AI software is now within reach of any budget, yet the gap between those who adopt it and those who stand still keeps widening: 71% of large companies have started at least one AI project, while among SMEs the share falls to 8%. The gap points to a difficulty that comes before the choice of product. It lies in how the question is framed.

The gap between the AI market and adoption in SMEs

According to the Digital Innovation in SMEs Observatory of Politecnico di Milano, 76% of Italian small and medium-sized enterprises have neither invested nor plan to invest in artificial intelligence, and only 7% have started structured training. Claudio Rorato, director of the Observatory, traces the root of the problem to readiness: «AI is not a shortcut. It requires vision, skills, adequate processes and a data culture». His reading shifts attention from the technology to the conditions that make it usable. Companies that experiment without a stated purpose end up with one more tool and an unchanged organisation of work. The budget is spent, the process stays as it was.

The friction between generic tools and real processes

Adopting a tool and transforming a process are different operations. Off-the-shelf software is designed for generic use cases, built around what many customers have in common, and it works well when the business process is standard. When the operational flow carries specificities of its own, the product does not bend to the company: the company adapts its way of working. A report by Anitec-Assinform with Politecnico di Torino describes the outcome as a lock-in effect: the company learns to use a given solution without developing strategic autonomy over AI. The consequences accumulate, between training redone at every vendor update and features that stay outside the contractual perimeter. The risks of automation projects without preliminary analysis follow the same mechanism, and so does the theme of training on the tool rather than the process. One indicator measures the distance between product and real work: only 19% of users say they rely exclusively on the tools provided by the company. The remaining share signals that the software adopted does not cover the work as people perform it.

The proprietary perimeter as a competitive advantage

Building a tailored platform has changed order of magnitude. Generative AI applied to code production cuts development costs by 15-30% on repetitive tasks, according to an analysis by SynSphere, a software house in the sector. The gap against subscription software has narrowed without closing, and this changes the calculations for those who had ruled out the option on budget grounds. More mature organisations are abandoning ideological approaches, all custom or all subscription, in favour of pragmatic choices: that is what lab51 finds in a May 2026 analysis. A platform built on the proprietary perimeter embeds the company's business logic, becomes an asset owned by the client and stays under its control. lab51 describes it as the shift from software treated as an operating cost to software treated as an industrial asset.

Distributing the investment over time completes the reasoning. A platform designed with intelligence is built in modules: the first release covers the most urgent needs with a contained, proportionate commitment, and later phases expand the solution following the company's real evolution. The first module focuses on what is needed now; later modules answer what matures afterwards. Three concrete consequences follow. The company pays for the features it uses, with no fees for what it does not need and no waiting for a third-party vendor to build what is missing. Every phase consolidates part of an asset, as document management built on the company's own operating logic shows. The solution grows with the company's operational specificities, while the software market for SMEs keeps pushing towards standardised models. The right economic comparison is the three-to-five-year TCO, where custom software is often more convenient from the third year, once subscription fees, process adaptations and hours spent on manual operations the generic product does not cover are included.

Choosing between custom software and SaaS

The picture does not justify a preference in advance. When business processes follow a widespread standard and the sector requires no significant operational specificity, subscription software is the right and faster choice: activation takes days and the initial cost stays low. Custom work emerges under three recognisable conditions. The first is an operational flow a manual cannot replicate, because it comes from the company's history and its exceptions. The second is the presence of heterogeneous systems already in use that must talk to each other. The third is a process that constitutes a competitive advantage, to be protected inside the company's boundaries instead of exported to third-party infrastructure. The right vantage point is the three-to-five-year TCO, where the initial cost weighs less than it appears.

The question to start from

Before assessing a product, the useful question concerns the perimeter. Which processes, in this specific company, with these data and these systems, justify a dedicated investment, and which belong to common ground where a standard works better. The answer points to the type of solution to look for: a subscription for shared areas, a proprietary platform built in modules for what sets the company apart. Custom AI software then stops being a product to pick from a catalogue and becomes the consequence of reading the processes. That reading is where every conversation with EBM starts.