AI’s Elusive Returns

While company after company reports benefits from AI, most have yet to realize any return on investment.

Sooner or later, the inevitable question about artificial intelligence will arise: Is the investment boom sweeping the corporate world a bubble?

The answer, predictably, lies in companies’ ability to demonstrate a return on their investment (ROI), where the benefits they realize from the technology outweigh the cost of employing it, expressed as a ratio in which the numerator exceeds the denominator. Yet a quotient larger than one has proven elusive, according to survey after survey.

Yes, companies are realizing benefits from AI, specifically generative AI (Gen AI), which can create text, images, and video from prompts using large language models trained on data. Companies including American Express, AstraZeneca, Bank of America, General Mills, Mass General, PayPal, Siemens, Unilever, and Walmart have reportedly seen improvements across research and development, manufacturing, logistics, and inventory management, as well as in customer and patient care, thanks to AI.

H&M Sees Improvements

Then there’s H&M. Facing high cart-abandonment rates and slow customer response times, the clothing retailer implemented an AI agent that offers personalized product recommendations, answers frequently asked questions, and guides customers through the purchasing process. The company reportedly found that 70% of customer queries were resolved autonomously, that conversion rates rose by 25% during chatbot interactions, and that response and resolution times fell threefold.

While H&M declined to confirm these numbers, a spokesperson says the company’s use of AI has had “positive effects on resource consumption, but also in terms of on inventory, raw materials, and emissions. AI furthermore helps us to create personalized customer experiences.”

As for PayPal, it reportedly saw an 11% reduction in losses in 2023, thanks in part to the employment of AI models to monitor fraud patterns. A company spokesperson declined to confirm the report but says that PayPal uses AI “to enhance our fraud prevention and detection capabilities alongside our manual risk controls.” The spokesperson adds that PayPal has seen a meaningful decrease in scam-related peer-to-peer transactions.

These types of gains are backed by data. In a McKinsey report published last November, a survey of 1,993 respondents from companies in 105 nations found that a majority reported either cost benefits or revenue gains from using AI. The most commonly reported of the savings were in software engineering, manufacturing, and IT, while revenue gains were commonly reported in sales and marketing, strategy and corporate finance, and product and service development.

The Big Disconnect

Yet despite the benefits companies have generated from AI, only a minority reports that the benefits outweigh the investment.

“While AI tools are now commonplace, most organizations have not yet embedded them deeply enough into their workflows and processes to realize material enterprise-level benefits,” the authors of the McKinsey study wrote. Senior partner Alex Singla wrote that most companies that have rolled out AI tools “have not yet productized use cases, redesigned workflows around AI and agentic capabilities, or built the platforms/guardrails needed to run them at scale.”

A study of more than 300 publicly disclosed Gen AI initiatives, published last July by the Massachusetts Institute of Technology, found that despite $30 billion to $40 billion in enterprise investment, 95% of the projects generated no return. A survey of 1,854 executives, published in October by the consultancy Deloitte, found that while 85% of the organizations had increased their AI investment in the previous 12 months and 91% were planning to do so again by year’s end, just 10% were realizing “significant” returns on their spend on agentic AI.

“Most organizations haven’t embedded AI deeply enough to realize enterprise-level benefits.”         —McKinsey

And those investments are likely to become more expensive as AI vendors shift from a subscription model, which offers unlimited use at a fixed price, to usage-based pricing.

McKinsey reported in November that “between 2015 and 2024, the number of consumption-based software companies more than doubled,” and that was before agentic AI changed the underlying logic of how software is used.

 In an interview with Global Finance, Nicolai von Bismarck, a partner and leader of McKinsey’s service operations practice, describes the shift as “structural and accelerating.” He adds, “The cost uncertainty this creates is real: It’s showing up in research as one of the top operational barriers to scaling AI.”

Small wonder, then, that banks are growing wary of lending to operators of data centers that support AI. JPMorgan Chase, Morgan Stanley, and Sumitomo Mitsui Banking Corp. are among those seeking to offload their ballooning data-center loan exposure to private funds and insurers. For example, banks have been trying to syndicate a $38 billion debt package tied to Oracle for six months and are offering it at a discount.

What explains the big disconnect between expectations and delivery in AI, compared with previous rounds of technological innovation?

The easy answer is the fear of missing out. Fueled by unprecedented hype, the corporate world is susceptible to a pervasive fear, leading to undisciplined investment. As von Bismarck says, “Companies invest heavily in experimentation but struggle to identify use cases with real impact. Many lack a clear road map connecting individual use cases to broader business strategy, leading to fragmented efforts, duplicative investments, and no single initiative achieving critical mass.”

What Will Accelerate ROI?

“Organizations that successfully cross the Gen AI Divide do three things differently,” the authors of the MIT report gathered from their survey results. “They buy rather than build, empower line managers rather than central labs, and select tools that integrate deeply while adapting over time.”

But a more fundamental issue may prevent even the most prudent organizations from realizing ROI from AI: context, or, in technology terms, metadata. A recent article in Modern Data 101, a publication of Modern Data Company, a venture-capital-funded firm with offices in Silicon Valley and India, argues that AI agents are ineffective unless they both understand and have access to data, and for that to happen, “the meaning of data must travel with the data itself.”

Von Bismarck sees the issue in terms of organization and culture. “Many of the companies that are breaking through share a common characteristic: They treat AI transformation the way they would treat any comprehensive operating-model transformation, with strategic discipline, executive accountability, and a clear theory of how value gets captured.”