Investment Surges While Organizational Reality Lags Behind
Global AI spending is on track to reach $2.5 trillion in 2026 – a 44% jump from the previous year – and yet the majority of enterprises are not growing revenue through AI or fundamentally changing how they operate. The money is moving. The organizations, largely, are not.
What’s emerging from this gap is a structural problem that no additional budget resolves on its own. Model capabilities are advancing faster than most organizations can absorb them, and the result inside many large companies is not progress – it’s fragmentation.
When AI Gets Good at the Wrong Things
The fragmentation problem looks like this in practice: a sales agent operates without visibility into open support tickets, or a marketing system personalizes content with no access to what the finance team already knows about that same customer. Each function may perform well in isolation. The enterprise as a whole, though, accumulates little usable intelligence and has less to act on as a result.
This is the core diagnosis in a new report from MIT Technology Review Insights, which describes the current moment as an “agentic shift” – a move away from AI as a discrete tool toward AI as an operating model. The distinction matters because the solution isn’t a better model or faster infrastructure. It’s the connective tissue between people, processes, and data, along with the governance structures that allow organizations to act on that intelligence without things going wrong.
Companies that are generating sustained returns from AI share a specific discipline, according to the report: they treat process redesign as work that happens before model selection, not after deployment. They build for how the technology will evolve, rather than retrofitting roles and workflows once a system is already running. For those companies, the architectural question comes second. The operating model question comes first.
The Data Problem No One Wants to Admit
Most enterprises discover too late that having data and having AI-ready data are entirely different conditions. The volume of stored data is not the constraint. Accessibility is.
The report argues that what makes AI “compoundable” – capable of building on its own outputs over time – is a sovereign, composable foundation that can query and prepare data where it already resides, without requiring migration or centralization. That framing is increasingly practical rather than theoretical, given that data residency laws, multicloud environments, and organizational complexity are making centralization harder to achieve, not easier. Sovereign control over where models run and where data lives becomes the thing that preserves adaptability when the legal or technical landscape shifts.
Three Infrastructure Problems That Require Simultaneous Solutions
The report identifies three architectural shifts that organizations need to pursue at the same time, not in sequence. First, rebuilding data infrastructure for accessibility rather than volume – a meaningful reversal of the data-hoarding logic that drove enterprise strategy for the past decade. Second, replacing fixed tech stacks with composable architectures that can evolve as models and tools change, rather than locking organizations into configurations that made sense at deployment but age poorly. Third, resolving questions of AI sovereignty: where intelligence runs, who controls it, and how it operates across organizational and jurisdictional boundaries.
That last category – sovereignty – is the one with the least established playbook. Jurisdictional complexity around AI is building faster than regulatory frameworks can clarify it, and enterprises operating across borders are already encountering situations where the legal answer to “where can this model run?” is genuinely unresolved. The organizations best positioned to navigate that aren’t the ones waiting for clarity; they’re the ones that built flexibility into their architecture before the question became urgent.
The companies falling behind, by contrast, are what the report calls process-last companies – organizations that selected models first and are now trying to retrofit their operations to fit. That sequencing problem is expensive to reverse. Workflows built around a specific model’s limitations tend to calcify, and the people operating within them develop incentives to defend rather than redesign. The gap between process-first and process-last organizations is widening, not closing, even as overall AI investment rises sharply across the board.
None of this is an argument against AI investment. The $2.5 trillion figure reflects genuine institutional conviction that AI will reshape how large organizations function. But conviction without architectural preparation is producing a specific and identifiable failure mode – intelligence locked in silos, data estates that can’t be queried, and operating models that were designed for a world where AI was still mostly a future ambition. That world is gone. The question is whether the infrastructure built to replace it was designed for the technology that’s actually arriving, or the technology that was easy to imagine three years ago.
Process-first companies are pulling ahead, and the gap between them and everyone else is structural – which means it won’t close just because spending increases. The enterprises still retrofitting workflows around models they deployed last year are running an experiment in organizational inertia, and the results are already visible in the numbers.
