The false apocalypse of AI and the real systemic risk

Between 28 January and 13 February 2026, more than USD 2 trillion of capitalisation evaporated from the enterprise software sector. This is the figure that Andrea Pignataro, founder and CEO of ION Group, starts from in his talk 'The Wrong Apocalypse' . Triggering the sell-off were the new plugins of Anthropic's Claude Cowork and the advancement of Claude Code, tools that showed how an artificial intelligence system is now capable of drafting legal documents, managing accounting flows and automating tasks typical of structured cognitive work. In more brutal words: how an AI system can replace much of the software we have known and used for years.

The market reaction was linear: if an AI agent can do what software does, that software becomes superfluous. Reinforcing this reading was Dario Amodei's essay, The Adolescence of Technology, which hypothesises a 'country of geniuses in a data centre' capable of disrupting employment and economic structures.

Pignataro argues that this interpretation hits the wrong target. The mistake lies in confusing capability and coordination. Enterprise software, he argues, does not serve primarily to perform cognitive tasks, but to coordinate cognitive work between different actors, with shared rules, authorisations, traceability, standards. It is an institutional infrastructure before being technological. Replacing a management system is not the same as replacing a function: it means rewriting the 'language' with which an organisation operates.

This is where Wittgenstein's notion of a 'language game' comes in: companies do not simply use Salesforce or an ERP, they talk through those platforms. Processes, metrics and hierarchies are embedded in the software. Therefore, according to Pignataro, erosion will mainly affect the more 'commodity' layer of software, while solutions that are deeply embedded in organisational processes will be more resilient.

The biggest risk, however, is not the rapid replacement of tools. It is the paradox whereby each firm, by adopting AI to remain competitive, helps to train the platform that could disintermediate it. Through aggregate and longitudinal patterns, platforms learn the 'grammar' of entire sectors: consulting, legal, insurance, financial advisory. No proprietary data is exposed, but the structure of work is. The result is a tragedy of the commons dynamic: rational individual decisions produce a collectively destructive outcome.