AI investments soar, but ROI remains elusive.

([1])For CFOs, these sobering numbers signal that AI can no longer be a faith-based investment. In a new survey of 260 senior finance leaders (135 of them CFOs), 87% said they must tie AI spending to concrete business outcomes within the next year, yet only 22% can currently do so ([2]). Finance teams are now finding that every dollar spent on AI needs justification through measurable results – a stark departure from the recent 'blank check' era.

([3])That shift in mindset is being driven from the top as well. Roughly two-thirds of corporate boards now demand hard evidence of value before approving new AI investments ([4]). In fact, 22% of boards have gone so far as to halt any additional AI spending until existing projects prove their ROI, and 43% of finance chiefs report they are already being pressed by directors for AI performance metrics they currently lack ([5]).

([6])This push for accountability is forcing tough decisions. Many organizations are scaling back or canceling AI initiatives that cannot demonstrate tangible value. Among companies unable to show ROI, 75% of finance leaders say they have already pulled back or frozen AI investments, and 35% have outright killed at least one AI project mid-stream ([7]). By contrast, for those with proven results, these figures drop to just 38% and 11% ([8]). In short, CFOs are tapping the brakes on unproven AI efforts – and doubling down on projects that clearly deliver business outcomes.

([1])A stark example of AI’s value gap just emerged in healthcare. The Blue Cross Blue Shield Association (BCBSA) reports that AI-assisted hospital coding tools have driven $942 million in additional inpatient costs across 2024–2025, primarily by automatically adding extra 'complexity' diagnoses that pushed patient cases into higher-paying reimbursement tiers without any change in treatment or outcomes ([2]).

([3])Hospital leaders immediately pushed back on the implication that these costs were wasted. The American Hospital Association argued that patients truly are more complex and that better documentation (some of it enabled by AI) is simply capturing clinical reality, while suggesting insurers have their own incentive to minimize payments ([4]).

This clash highlights a broader lesson: AI can boost proxy metrics – whether billing codes or click-through rates – that don’t necessarily equate to real value. Enterprise leaders must rigorously audit AI-driven 'gains' to ensure they translate into tangible business outcomes, not just technical or paper improvements. If an AI tool shows a KPI jump, they need to verify it actually contributes to the bottom line.

([1])Another growing concern is the sheer cost of running AI at scale. In recent months, organizations that eagerly rolled out AI across their workforce have been caught off guard by steep cloud and compute bills, leading many to impose usage caps and stricter budget controls. Industry reports describe how companies from Amazon and Walmart to Cisco, Uber and Meta have set hard limits on employee AI tool consumption, shifted teams to more efficient models, and explicitly warned against deploying AI 'for the sake of AI' after seeing some annual budgets exhausted within mere months ([2]).

([3])In response to these cost shocks, some enterprises are rethinking the build-versus-buy equation. Nearly one-third of large firms in McKinsey’s study chose to develop certain AI applications – such as custom coding agents – in-house rather than pay hefty vendor fees ([4]). Building proprietary solutions can offer more control over costs and data, though it requires significant internal expertise and infrastructure.

([5])At the same time, competition is beginning to drive down prices in select AI services. Over the weekend, China’s Alibaba announced a new suite of five speech AI models (Qwen-Audio 3.1) with usage fees up to 95% lower for functions like transcription and voice generation ([6]). In concrete terms, Qwen-Audio 3.1 can transcribe an hour of audio for about $0.02 ([7]) – a fraction of the ~$0.15–$0.26 per hour charged by leading U.S. providers ([8]). These dramatic price cuts show that innovation isn’t just increasing AI’s capabilities – in some cases, it’s sharply reducing costs, a trend enterprise buyers should monitor closely.

([1])Finally, these sobering findings do not mean AI can’t deliver a return – they mean many organizations haven’t yet laid the groundwork to capture it. The eye-opening 95% 'no ROI' rate for generative AI pilots reflects how frequently companies underestimated the data, integration, and process changes required to turn impressive demos into reliable, scalable systems ([2]). In effect, many early AI initiatives were technology-driven experiments not grounded in operational reality, so it is no surprise that they often failed to produce real business outcomes.

([3])Notably, the small minority of companies that are achieving strong AI results tend to focus on narrow, well-defined use cases where success can be clearly measured, rather than chasing broad moonshots or deploying AI indiscriminately in hopes of striking gold ([4]). By securing quick wins in specific high-impact processes, these leaders build organizational buy-in and expertise before scaling up.

([5])Crucially, organizations seeing real ROI also invest far more in execution than in shiny tools. One study found unsuccessful AI programs devoted 78% of their budgets to vendor software and only 22% to integration, training, and measurement, whereas high-ROI achievers spent just 35% on licenses and the rest on the hard work of redesigning workflows, upskilling staff, and tracking performance ([6]). The lesson for the C-suite: extracting value from AI is less about algorithms and more about managing change – aligning projects with strategy, fostering the right talent and data foundations, and measuring impact relentlessly.