AI Must Prove Value, Not Just Capability: The 2025 ROI Reckoning

Two and a half years after ChatGPT, the question every CFO is finally asking out loud is no longer "can it do this?" but "is it worth what we're paying for it?" The capability race produced models so good they have effectively commoditised. The value race, proving an enterprise's AI spend translates into measurable productivity, revenue, or margin, has barely started, and H1 2025 made the gap impossible to ignore.

For three years, AI strategy at most large companies has been judged on capability metrics: which model, which context window, which benchmark. By mid-2025, capability has stopped being a meaningful differentiator. Three different vendors and at least one open-weight family can serve nearly any enterprise use case. What hasn't been settled is whether all that capability translates into ROI, and the data so far is uncomfortable.

The widening capability-versus-value gap

McKinsey's State of AI 2025 survey, published in May, found that 78 percent of organisations now use generative AI in at least one business function, up from 55 percent the year before. The same survey reported that fewer than one in five organisations could point to a meaningful EBIT impact from those programs at the enterprise level. Adoption is up sharply; measurable financial impact is not.

The Stanford AI Index 2025, released in April, told a similar story from a different angle: corporate AI investment hit a record in 2024, productivity gains in narrow tasks are now well-documented, but most companies still cannot answer the basic question of how their AI spend converted into operating results.

IBM's Institute for Business Value 2025 CEO study, published in May, went further: only about a quarter of AI initiatives had delivered the expected ROI, and CEOs were openly conceding that AI investments hadn't yet met expectations. The polite phrasing in the press releases was "the gap between ambition and execution." The blunter reading is that the bill is now arriving.

Klarna as the canary

In 2024, Klarna boasted that an AI assistant was doing the work of 700 customer-service agents. In May 2025, the company's CEO Sebastian Siemiatkowski reversed course publicly: Klarna had cut too many human agents too quickly, customer-experience scores had suffered, and the company was reintegrating humans into the loop.

Klarna isn't a canary because the AI didn't work, it did. It's a canary because the company optimised for cost cuts instead of customer value, and the bill came due in churn and brand damage. By July 2025 Klarna had become a reference point in nearly every enterprise AI procurement conversation: capability is necessary, but not sufficient.

The spend-versus-impact gap at hyperscale

The pattern repeats one tier up. In early July 2025 Microsoft announced roughly 9,000 layoffs, its third major round in twelve months, even as it continued to spend tens of billions of dollars on AI infrastructure. Investors stopped accepting "AI capex" as a self-justifying narrative; analysts started asking out loud how Copilot revenue compares to the data-centre build-out.

Meta took the opposite tack. In June 2025 it invested roughly $14 billion for a 49 percent stake in Scale AI, brought Alexandr Wang in to lead a new "Meta Superintelligence Labs", and aggressively hired senior researchers from OpenAI, Anthropic, and Google DeepMind. The bet is that getting to superintelligence first is worth almost any price. The shareholders' question is what intermediate value gets delivered along the way.

Apple, at WWDC on 9 June, conceded the inverse problem: it had announced AI features in 2024 that it could not ship in 2025. The more personalised Siri promised at WWDC 2024 slipped publicly, and Apple Intelligence's enterprise relevance remains modest. The capability gap is real, but the deeper issue is that none of these companies, incumbents or challengers, have a clean answer to the ROI question.

What this means for European enterprises

For organisations in Switzerland, Germany, and the broader EU, the H1 2025 picture has three operational consequences.

Procurement now has to score value, not just capability. A two-year licence committed in 2024 because Vendor X had the best benchmark looks fragile in 2025 when an open-weight competitor matches it at a fraction of the cost. AI procurement has to bake in periodic value reassessment, multi-vendor optionality, and clear exit clauses.

Pilots without measurement are now liabilities. The McKinsey data shows that companies treating gen AI as a portfolio of measured experiments, with pre-stated KPIs, baselines, and review gates, outperform those running open-ended exploration. The EU AI Act's risk-management obligations push the same direction: documented, auditable AI use is governance, but it is also the only honest way to know what is working.

Governance is increasingly inseparable from value. Data-loss prevention, role-aware access, and audit trails are usually framed as compliance costs. In a 2025 procurement conversation, they are also the controls that let an enterprise actually deploy AI to its sensitive workflows, the workflows where the ROI is. Governance is no longer a tax on AI value; in regulated industries, it is the precondition for it.

The next phase

The capability-first phase of enterprise AI is closing. The next phase is value-first: programs that survive 2026 procurement reviews will be the ones with documented productivity uplift, defensible cost reductions, or revenue contribution traceable to specific deployments, and the governance posture that makes those deployments safe in regulated environments.

The vendors that win the next two years won't be the ones with the highest benchmark scores. They will be the ones that make ROI the default measurement, not an afterthought.