AI investments soar as ROI reality sets in.

New data shows that AI spending is blowing past corporate budgets – and executive patience is wearing thin. In a survey released just this week, 46.9% of enterprises reported their AI expenditures in the second half of 2026 exceeded what they had planned ([1]). This isn’t a few outliers; it’s nearly half of big organizations busting their AI budgets. The result? CFOs and boards are hitting the brakes on freewheeling AI spending and demanding to see value for every dollar.

For many companies, the conversation around AI is shifting dramatically. As one analyst put it, the AI budget question has moved from “whether to adopt” to “how to fund it” ([2]). In practice, most firms haven’t been reining in overruns by canceling projects – only about one in six actually slow their AI initiatives when costs climb ([3]). Instead, 47.6% of over-budget organizations simply asked for more money, and a further 43.3% absorbed the extra costs and planned to sort it out later ([4]). This essentially means the usual budget guardrails failed; as the same expert noted, the budget has “stopped being a control” in many AI programs ([5]).

The scramble to cover ballooning AI bills is revealing priorities – and pain points. To fund AI overruns, nearly 61% of firms that shifted funds within their IT budgets did it by cutting external consultants and contractors first ([6]). In some cases, business units outside of IT are now chipping in – 23.1% of organizations over budget tapped non-IT departmental funds to fill the gap ([7]). External service vendors are taking a hit as companies rob Peter to pay Paul. These measures show how urgent – and unsustainable – the situation has become. Board members are growing wary, and moving forward they are likely to demand much tighter governance on AI spend, clearer cost-benefit analysis, and defined budget limits. The honeymoon period for “spend now, figure out value later” is coming to an end.

The surge in AI investment has not been matched by commensurate business returns for most organizations. Roughly 70% of companies worldwide now report using generative AI in at least one business function ([1]), but much of this usage remains stuck at the pilot or single-use-case stage. Industry analysts are openly talking about an emerging "AI backlash" as the initial wave of generative AI deployments often failed to deliver the promised transformation to the bottom line ([2]). There’s a pronounced “demo gap” between what AI vendors and innovators have showcased and what companies are actually able to scale in production to drive results ([3]).

The pharmaceutical sector’s experience is a cautionary example of this ROI gap. A new Citi Research survey of US pharma R&D leaders found that 72% of these companies have already scaled AI across their research and drug development efforts ([4]) – essentially a near-universal adoption. Yet only 24% of pharma executives expect AI to significantly improve their new drug success rates ([5]). Over half anticipate merely modest improvements, not the game-changing breakthroughs they’d hoped for ([6]). As the head of Asia healthcare research at Citi bluntly warned, the biggest risk isn’t that AI won’t speed up drug discovery – it’s that all that acceleration "does not translate into better drugs or higher probability of success" ([7]). In other words, faster R&D is futile if it doesn’t yield real outcomes. This tempered outlook from a high-tech industry like pharma speaks volumes. It suggests that even as companies invest heavily in AI, they are sobering up about the returns – recognizing that simply deploying algorithms doesn’t guarantee a payoff in practice.

Similar signs of an adoption-to-outcome disconnect are appearing across industries. Survey data summarized this week exposed that while AI budgets and usage have exploded, many enterprises have little to show on the balance sheet so far ([8]). Even among organizations that *are* measuring the impact of AI, the median positive return on these projects is only around 10% ([9]) – a far cry from the dramatic productivity leaps often touted at conferences. In fact, one recent analysis of multiple 2026 surveys noted that corporate AI spending and adoption are surging, but capabilities like governance, security, and ROI measurement are lagging behind, pushing the market out of its hype phase into an "accountability" phase ([10]). The message for executives is clear: high adoption alone means little unless those investments yield tangible improvements in efficiency, revenue, or profit.

Why are so many AI initiatives under-delivering? One fundamental problem is that the success of AI projects depends on much more than algorithms – and those “mundane” prerequisites have been underestimated. Chief among them is data. Only a tiny fraction of companies have truly "AI-ready" data at scale, with consistent, well-integrated, high-quality information across the enterprise. Most organizations still describe their data as only partially ready for AI use, and this shortfall is putting many projects in peril. In fact, Gartner analysts now predict that 60% of AI projects could be abandoned by the end of 2026 due to a lack of AI-ready data to support them ([1]). This should be a flashing warning light for any CEO: no matter how powerful your model, poor data quality and siloed systems will stall your AI results.

Integration and process change pose equally thorny hurdles. Many early AI pilots were treated as isolated tech experiments, bolted onto legacy processes without rethinking workflows or incentives. In hindsight, this was a recipe for failure – the majority of effort in scaling AI is not in tweaking algorithms, but in the heavy lifting of data engineering, systems integration, and change management. These are the less glamorous tasks (cleaning up data, integrating AI into existing software, retraining staff), and they’ve often been under-funded in the AI gold rush. When that work doesn’t get done, even the most advanced AI models end up underutilized or producing outputs that never translate into business action. As a result, we see many proof-of-concept projects lingering on the shelf and a growing number of AI investments being put on hold.

The sheer computational cost of cutting-edge AI has also taken many by surprise. Nearly half of all enterprise AI spending this year will go into computing infrastructure like specialized chips and cloud services ([2]). Training and running large models – especially generative AI – is extremely expensive, with one forecast showing spending on these models skyrocketing 117% year-over-year ([3]). These costs add up quickly when usage isn’t carefully controlled. Many CIOs have stories of eye-popping cloud bills from unrestrained AI experiments. This is forcing hard choices: some organizations are delaying AI projects or scaling back model usage until they can optimize costs. Even AI providers are starting to adjust their pricing strategies in response – for example, one leading AI vendor recently slashed a key usage fee by 75% to appease customers’ cost concerns ([4]). Clearly, the economics of AI are evolving, and both buyers and sellers are seeking ways to make AI’s ongoing costs more sustainable.

If there’s a silver lining in this cold shower of AI ROI data, it’s that we now have a clearer idea of what it takes to capture value from AI – and what mistakes to avoid. A critical success factor is disciplined management. Organizations that approach AI as a strategic investment – setting concrete objectives, tracking results, and being willing to pull the plug on underperformers – vastly outperform those that don’t. In Gartner’s study, the companies labeled “high performers” were those that continuously measured the ROI of their AI initiatives and managed them as a portfolio of bets; they reported positive returns on the great majority of their AI projects ([1]). By contrast, many laggards admitted they simply didn’t know the ROI of a large share of their AI experiments, flying blind on value delivered ([2]).

Enterprise leaders are also learning to focus on the right use cases. Practical, lower-profile applications – like automating IT operations, customer service processes, or quality control in manufacturing – are emerging as reliable sources of ROI, often delivering cost savings in the 20–30% range and faster cycle times. Meanwhile, some of the flashier ambitions (like fully AI-driven product innovation) remain unproven in terms of bottom-line impact ([3]). For example, in the pharma industry, AI is already yielding returns by streamlining clinical trial processes and optimizing supply chains, but using AI to discover brand-new blockbuster drugs at scale has yet to show clear ROI ([4]). The takeaway: doubling down on known high-value use cases while approaching moonshots with careful expectations can help balance long-term vision with short-term wins.

Finally, companies are starting to equip themselves to manage AI’s impact. New tools and frameworks are emerging to help quantify and accelerate AI-driven improvements. Just yesterday, productivity analytics firm ActivTrak launched a “workflow optimization” platform to help enterprises pinpoint where AI and automation can save the most time and money – and to measure the results ([5]). As ActivTrak’s Chief Product Officer put it, the goal is to give executives data to “make those decisions and prove what changed,” turning AI initiatives from a collection of assumptions and experiments into “a measurable system for improvement” ([6]). In the same vein, successful companies are investing heavily in employee training and process re-engineering to ensure that AI tools are actually being used effectively on the front lines of the business. And they are not shy about cutting losses: when an AI pilot doesn’t perform, it’s rapidly retooled or shut down, so resources can be redirected to more promising initiatives. This pragmatic, value-first mindset is what separates the AI winners from those stuck in perpetual experiments. With boards and investors now zeroed in on ROI, adopting this disciplined approach is becoming non-negotiable.