AI Spending Soars, ROI Stalls – A Reality Check for Business Leaders

Businesses have poured unprecedented funds into artificial intelligence, spurred by competitive pressure and high expectations. Global private AI investment reached a record $285.9 billion in 2025 ([1]), and major enterprises like JPMorgan Chase and Bank of America now commit billions annually to AI in pursuit of efficiency and growth ([2]). Yet early results have underwhelmed. A recent PwC global survey of 4,500 CEOs found 56% report no increase in revenue or decrease in costs from their AI investments over the past year ([3]). Only about one in eight CEOs saw AI deliver both cost and revenue benefits in that time ([4]). In Deloitte’s 2025 study, most firms said a typical AI project takes 2–4 years to pay back – far longer than the 7–12 month payback period business leaders expect for technology investments ([5]). In short, the promised quick wins of AI have largely failed to materialize for the majority.

This growing realization is prompting a backlash from the C-suite and investors. After years of experimentation with elusive returns, CEOs and CIOs are under intensifying pressure to show measurable value. According to a Kyndryl survey, 61% of senior executives feel more pressure to prove AI ROI now than they did a year ago ([6]). Boardrooms, once swept up in AI enthusiasm, are now asking hard questions about the bottom-line impact of these initiatives. Investor expectations are also proving aggressive – a Teneo study found 53% of investors want to see positive returns from new AI projects in under six months ([7]). Yet 84% of CEOs acknowledge that achieving returns will take longer than two quarters ([8]). This mismatch is creating palpable tension. The honeymoon period for unchecked AI spending is over; today, AI projects are judged by the same standards as any other investment.

Even in sectors that have led in AI adoption, tangible payoffs remain murky. Big banks, for example, have been quick to bet on AI: JPMorgan earmarked $1.2 billion of its $19 billion tech budget to AI, and Bank of America is investing $4 billion in AI-related “strategic growth” initiatives ([9]). Goldman Sachs and BNY Mellon likewise spend billions on technology, much of it on AI and automation ([10]). Yet according to analysts, these institutions have so far seen little meaningful impact on revenue or profitability from AI – no fundamental change to their business outcomes ([11]). This reflects a broader “AI impact gap” affecting many industries ([12]): lofty ambitions and spending have not translated into the expected improvements on the income statement. Senior leaders who once feared missing out on the AI gold rush are now grappling with an uncomfortable question: was the rush into AI justified without clearer returns?

While ROI remains uncertain, the costs of deploying AI at scale are very real – and rising. Many companies have been caught off guard by how expensive widespread AI usage can become. Uber, for example, blew through its entire 2026 AI budget for an advanced text-generation model by April ([1]). A number of other tech-driven firms saw usage skyrocket unexpectedly, leading Microsoft to revoke employees’ access to a costly third-party coding assistant after just months, and one company reportedly faced a $500 million bill when it failed to put usage limits on an autonomous AI agent ([2]). Even though leading AI providers have dropped per-unit prices, the push for “AI everywhere” – especially with powerful generative models and agents – has driven overall consumption (and cloud bills) through the roof ([3]). What started as small pilot projects have turned into significant new cost centers as usage scales beyond initial forecasts.

This sticker shock has set off a wave of cost-cutting measures. TechCrunch and other outlets report that companies across industries are now imposing strict budget limits and governance on AI usage ([4]) ([5]). Microsoft, for instance, has begun shifting its Office 365 suite to rely more on its own internally developed AI models, reducing calls to OpenAI’s and Anthropic’s systems to save on hefty licensing fees ([6]). Other major players like Amazon, Meta, and Uber have likewise curtailed employee access to expensive third-party AI tools in recent months ([7]). In response to enterprise concerns, the Linux Foundation even launched a "Tokenomics Foundation" to establish standards for tracking and controlling AI compute and usage costs – introducing cloud-style cost discipline to AI deployments ([8]). In short, CFOs are now as worried about AI costs as they once were about cloud spending.

The drive to rein in AI expenses is also spurring a reexamination of the “build vs. buy” equation. Some organizations are exploring open-source and regional AI models as cheaper alternatives to Big Tech offerings. In an ironic twist, Chinese AI labs have recently produced cutting-edge large language models that not only rival some of the best US systems in capability, but also come at a fraction of the price ([9]) ([10]). One new Chinese model, GLM-5.2, has impressed Silicon Valley engineers with its advanced coding skills while reportedly costing significantly less to run than OpenAI’s or Anthropic’s equivalents ([11]) ([12]). The arrival of such low-cost competitors is timely: Coinbase claims it has nearly halved its AI expenses by defaulting some workloads to two Chinese-developed AI models, dramatically reducing its reliance on pricier US cloud AI APIs ([13]). Although most mainstream companies still spend only modestly on AI (the median business using just $11 per employee monthly on AI tools) ([14]), those at the forefront – the top 1% of “AI-pilled” firms – are now burning up to $7,500 per employee each month on AI solutions ([15]). For these heavy adopters, finding cost-efficient models and optimizing token usage isn’t just an IT concern; it’s becoming a board-level mandate to protect margins.

If AI is so powerful, what explains the dearth of ROI in practice? One issue is where the money has been going. Companies often jumped into glamorous or trendy use cases that didn’t address pressing business needs ([1]). MIT’s “State of AI in Business” study found that 50–70% of many firms’ AI budgets were poured into customer-facing pilots – like generative AI for marketing content or chatbots – which were easy to pitch but delivered superficial benefits ([2]). Meanwhile, higher-impact opportunities in operations, finance, and supply chain saw comparatively less investment, even though these areas are where early adopters are actually realizing cost savings and efficiency gains ([3]). In short, misplaced priorities and "fear of missing out" led to a flurry of AI projects without clear ROI, a dynamic one analyst dubbed an "AI hangover" once the initial euphoria wore off ([4]) ([5]).

Another common pitfall is the gap between experimentation and execution. Many AI projects never made it out of the lab into production, creating lots of prototypes but few usable solutions. Interviews by MIT researchers revealed that only about 5% of AI pilot projects successfully scaled up to deliver measurable value – the rest fizzled out ([6]). Often the culprit was not the algorithms themselves but the surrounding infrastructure and data. Companies frequently overestimate their data readiness and the ease of integration ([7]). A proof-of-concept might perform well on clean sample data, but once deployed, hidden data silos, quality issues, and outdated systems stall progress ([8]). Many organizations also attempted too many projects at once, spreading resources thin. BCG found that average companies juggled around six different AI use cases, whereas the top performers focused on just 3–4, allowing them to execute those few initiatives much more fully ([9]). This “spray and pray” approach – trying a bit of everything – left many projects under-resourced and unable to achieve impact.

Human factors further explain the shortfall in outcomes. Without enough investment in training and change management, even promising AI tools can languish unused. Yet only 28% of organizations have upskilled more than a quarter of their workforce to use AI, according to BCG’s survey ([10]). Employees often resist or underutilize AI solutions they don’t understand or trust, especially when poorly communicated or when AI output quality is inconsistent (as seen with generative AI’s tendency to "hallucinate" incorrect information ([11])). This not only limits the technology’s benefits but can create new inefficiencies as workers double-check and rework AI outputs. Cultural and organizational misalignment can likewise derail ROI – deploying AI in a silo, without broad buy-in and adaptation of processes, is a recipe for minimal impact ([12]).

Lastly, a simple but critical reason many AI initiatives underwhelm is that success isn’t being measured properly. Astonishingly few companies set concrete performance indicators for their AI projects. Deloitte found that most organizations do not even track the financial metrics of their AI implementations in a rigorous way ([13]). In the absence of clear KPIs, projects can drift without accountability, and any benefits they produce may go unnoticed and unquantified. This lack of ROI measurement means that even potentially successful AI solutions can be deemed failures — or vice versa — because leaders lack data to know the difference. It’s hard to achieve returns when you don’t define what success looks like.

A small minority of organizations are bucking these trends and realizing substantial returns from AI – and they offer a blueprint for others under pressure to deliver value. What are they doing differently? For starters, successful AI programs have strong executive ownership and clear business alignment. Deloitte found that in about 10% of organizations, the CEO is directly leading the AI agenda, which signals AI’s strategic importance and ensures initiatives target top business priorities ([1]). These companies treat AI projects like any other major business program, with defined objectives and accountability, rather than open-ended R&D. A senior leader’s sponsorship also helps break down silos and align AI efforts with enterprise-wide transformation goals.

Another differentiator is focusing on quality over quantity in AI use cases. Rather than dabbling in dozens of experiments, top performers double down on a few high-impact applications. BCG’s global survey revealed that leading companies concentrate on an average of just 3.5 AI use cases, whereas other firms spread efforts across 6 or more ([2]). By prioritizing depth, these leaders were able to reengineer processes and scale solutions around those use cases, and they anticipate roughly 2.1 times higher ROI on their AI investments compared to peers ([3]). One example comes from the insurance sector: New York Life’s technology arm rigorously evaluates AI initiatives based on tangible impact on the company’s earnings plan, funding only projects with clear cost or revenue benefits. This disciplined approach has yielded positive returns on AI efforts to date ([4]), reinforcing the value of selective investment.

Crucially, organizations that get ROI from AI invest heavily beyond the algorithms. They recognize that turning pilot projects into profit requires spending more on people and processes than on technology alone. BCG researchers describe a “10-20-70” rule among AI high-achievers: roughly 10% of the effort goes into developing AI algorithms, 20% into the supporting data and IT infrastructure, and 70% into training employees, adapting workflows, and driving cultural change to utilize AI effectively ([5]). In practice, this means extensive upskilling programs, redesigning workflows around AI capabilities, and involving end-users early so they trust and adopt new tools. When Palo Alto Networks implemented AI to automate IT operations, for example, it paired the technology deployment with process changes and staff training – resulting in automated tasks jumping from 12% to 75% of IT operations over 18 months, which cut those operating costs in half ([6]). Such results underscore that AI’s value isn’t plug-and-play; it flourishes when humans know how to leverage it and processes are restructured to capitalize on AI’s strengths.

Measuring and iterating for success is the final piece of the puzzle. Leading organizations establish clear metrics for AI outcomes from the outset – whether it’s reduction in processing time, decrease in call center volumes, increased conversion rates, or other KPI improvements. They move beyond vanity metrics like number of users, focusing instead on “auditable outcomes” that tie AI usage to business results ([7]). Some forward-looking CFOs are even introducing dedicated “AI ROI” tracking in financial reports (essentially an AI P&L) to monitor value creation versus costs ([8]). On the technology side, vendors are responding by providing more transparency: for instance, Google’s enterprise tools now let companies track AI consumption by department in real time ([9]), helping correlate usage with productivity gains or lack thereof. Equipped with better data, management can identify which AI applications are truly driving performance and which are not pulling their weight. By systematically measuring impact and course-correcting, these leaders ensure that resources support the AI initiatives with the strongest returns – and they cut bait on projects that don’t meet ROI targets.

The upshot for executives is clear. Realizing value from AI at scale is possible, but it demands a shift from indiscriminate experimentation to disciplined execution. The companies gaining a competitive edge are those that have moved past the "AI as magic" mindset and are treating AI as a strategic business tool – one that requires upfront planning, cross-functional change, and rigorous accountability. As one industry expert put it, the era of groundless AI hype is ending, and the era of AI accountability has begun ([10]). For senior leaders, that means now is the time to recalibrate AI programs around tangible outcomes. The message is simple: stop chasing every shiny new AI idea, and start doubling down on the applications that deliver real, measurable impact.