Somewhere in a glass-walled conference room, a chief executive is proud of their AI rollout. They know exactly how many Copilot licenses they distributed. They can tell you how many tokens their teams consumed last quarter. What they probably cannot tell you is whether any of it made the company more money.
That is the uncomfortable observation at the centre of a Bloomberg Radio interview published on June 4, 2026, featuring Dr. Rebecca Homkes - an economist and lecturer at the London Business School and faculty member at Duke Corporate Education. Her argument, stated plainly: AI adoption is high. AI integration is not. And the gap between those two things is eating up an enormous amount of corporate investment with very little to show for it.
The headline figure is striking. Only around 10 to 15% of business work is being done in any form with AI assistance, Homkes said. In a technology described by its proponents as a once-in-a-generation transformation - somewhere above the internet and below electricity on the hype index - that is a remarkably thin slice. Nearly every large company has bought the tools. A small fraction has changed how the work actually gets done.
The background
To understand why this matters economically, it helps to understand what companies actually did over the past three years.
When generative AI - the kind that produces text, code, images, and analysis from a prompt - became widely available to the public in late 2022, it spread faster than almost any consumer technology in history. ChatGPT reached 100 million users in two months. The internet took about seven years to reach 50% of global adoption. Generative AI got there in roughly three, a pace that gave investors and corporate strategists the impression that transformation was imminent and enormous.
The corporate response followed a predictable arc. In 2023, most executives were skeptical - reasonably so, given that tech hype cycles have a long and expensive history of disappointing forecasts. Through 2024, the mood shifted to what Homkes calls "dabbling": companies started throwing tools at their workforces. Microsoft Copilot - an AI assistant built into Word, Excel, Teams, and Outlook that costs businesses a monthly per-user fee on top of their existing Microsoft 365 subscriptions - became one of the most widely deployed examples of this approach. Buy licenses. Distribute them. Call it AI strategy.
By 2025, the spending had become enormous. Goldman Sachs estimated global AI investment at roughly $200 billion for the year, with overall AI-related enterprise spending projected at $1.5 trillion when hardware, cloud infrastructure, and software are included.
The problem is that spending money on tools is not the same as redesigning how work happens. And redesigning how work happens is, it turns out, the actual hard part.
GDP - the total value of everything a country produces in a year - is where the macro stakes of this conversation live. If AI delivers on even a fraction of the productivity gains promised, GDP growth would accelerate significantly. If it does not, trillions of dollars of investment will have produced an expensive automation of tasks that were not blocking growth in the first place.
What is actually happening
Homkes draws a sharp distinction between the two kinds of productivity gains AI can deliver, and the difference explains most of what has gone wrong with the current wave of deployment.
The first kind is team-level or function-level productivity. An engineering team that uses AI coding assistants can plausibly see 30% more output. A customer service operation deploying AI for ticket handling and response drafting can see gains of 30 to 50%. These gains are real, measurable at a departmental level, and relatively easy to achieve with off-the-shelf tools. They are also, in Homkes's framing, not what investors and CEOs are hoping to see when they talk about AI transforming their businesses.
The second kind - cross-functional, organisation-wide productivity - is where the real economic value would come from. And that, she argues, is not happening in any meaningful way yet. The reason is structural: achieving gains at that level does not mean adding AI tools on top of existing workflows. It requires actually redesigning the workflows. That is slow, politically complex work that involves rethinking how different departments hand off tasks to each other, how decisions get made, and how accountability is measured.
Meanwhile, companies are measuring the wrong things. Homkes notes that most organisations are tracking outputs - how many AI use cases are active, how many licenses are in use, how many tokens are consumed. Those metrics say nothing about whether the organisation is actually more capable or more profitable. According to her analysis, organisations get essentially nothing from those output measures. What would matter is measuring outcomes: whether AI is increasing revenue, reducing time to market, or improving customer retention in measurable ways.
McKinsey's 2025 State of AI survey of executives across more than 100 countries found that 88% of respondents reported regular AI use in at least one business function. Yet only around one in three had begun scaling AI programs across the organisation, and meaningful bottom-line impact - an earnings improvement of 5% or more - was confined to about 6% of companies. The numbers match Homkes's diagnosis almost exactly.
The gap between widespread access and meaningful integration is sometimes called the Solow Paradox - a reference to the economist Robert Solow's 1987 observation that computers were visible everywhere except in the productivity statistics. The same dynamic took years to resolve after the PC rollout of the 1980s and the internet boom of the 1990s. AI may follow the same curve, delayed by the same structural friction.
The money trail
The economics here are uncomfortable for the companies doing the selling and the companies doing the buying.
On the vendor side, the business model depends on volume. Microsoft charges businesses for Copilot access as an add-on to existing Microsoft 365 subscriptions. OpenAI, Anthropic, Google, and others charge primarily on usage - often measured in tokens (the unit by which AI models process text: roughly three-quarters of a word). As Homkes noted, AI infrastructure pricing has been in flux, with vendors repricing compute costs and token rates in ways that are creating volatility for corporate buyers trying to model their costs.
The demand side is also unusual. Unlike software that replaces a specific tool or process, generative AI is being sold as a general-purpose capability layer. That means the ROI depends almost entirely on what the buyer does with it - and most buyers, as the data shows, are not doing very much. A UNSW Business School analysis of AI investment patterns found that Goldman Sachs chief economist Jan Hatzius stated AI's contribution to US GDP growth in 2025 was "basically zero," while MIT economist Daron Acemoglu has suggested AI will likely increase total factor productivity - a measure of economic efficiency that accounts for all inputs, not just labor - by less than 0.66% over the next decade. Hatzius's own team expects meaningful productivity gains to begin materialising from 2027, compounding slowly through the late 2030s.
The range of expert forecasts is dizzying. Homkes herself cited estimates running from a 0.1% productivity boost all the way to 80% - three orders of magnitude apart. That spread is not a sign of healthy scientific debate. It is a sign that nobody actually knows, and that the underlying economic mechanism is still poorly understood even by the people paid to model it.
Who is benefiting in the meantime? Primarily the infrastructure layer. Nvidia's chips power most AI model training and inference - the computational process of generating a response. Cloud providers - Amazon, Microsoft, and Google - earn margin on every AI workload running on their servers. The model developers themselves, including OpenAI and Anthropic, are burning significant capital building and maintaining models at scale. The corporate users paying for access are, in most cases, getting productivity improvements confined to individual teams, not the organisation-wide transformation the pitch decks promised.
There is also a governance cost that rarely shows up in AI investment discussions. Homkes identifies four things organisations need to pursue simultaneously to get real value from AI: workforce upskilling, a responsible AI and ethics framework, a functional data infrastructure, and a clear link to business outcomes. Most established companies do not have their data in a state where AI tools can actually use it effectively. Legacy systems, siloed databases, and inconsistent data formats are common across large organisations, and solving them is expensive and unglamorous work that rarely gets budgeted alongside the headline AI spending.
What people are doing about it
The companies making the most progress, according to Homkes, share a counterintuitive characteristic: they have the clearest rules. Organisations that defined strict governance frameworks and responsible AI policies early have moved faster, because employees understand what they are allowed to do and what falls outside boundaries. When the rules are ambiguous, people hesitate. That hesitation compounds across an organisation into stagnation.
The Boston Consulting Group - where Homkes also serves as an advisor - has made board-level and C-suite sign-off on ethical AI implementation a prerequisite before engaging with corporate AI projects. Younger employees, it turns out, care as much about this as senior leadership does.
On the education side, business schools are trying to fill the gap that vendor training has left open. What most leadership teams need is not how to use a tool - plenty of that exists already - but how to redesign their organisations around AI outputs. How to retrain and redeploy staff. How to set accountability structures for AI-assisted decisions. How to measure outcomes rather than activity. Duke Corporate Education and the London Business School, where Homkes teaches, are among the institutions working on that layer of knowledge transfer.
Corporate buyers are also starting to push back. Goldman Sachs has signalled that AI budgets will increasingly go to teams that can demonstrate measurable productivity gains, while pilots without clear ROI face tighter scrutiny. The era of buying AI tools because everyone else is buying AI tools is showing early signs of ending - or at least slowing down.
The broader workforce adjustment is happening unevenly. Function-level gains in engineering and customer service have been real, but they have not translated into visible wage or employment changes at scale yet. Deloitte's 2025 survey of senior executives found that most organisations expect to achieve satisfactory return on AI investment over two to four years - significantly longer than the seven-to-twelve-month payback period typically expected from technology investments.
The bottom line
The AI story is not that companies are failing to adopt the technology. They have adopted it broadly and spent heavily to do so. The real story is that adoption without workflow redesign produces very little economic value - and workflow redesign is slower, harder, and more expensive than buying a software license. The distance between "we have Copilot" and "AI is driving our growth" is not a few months of training. It is, as Homkes's analysis suggests, a fundamental restructuring of how organisations operate, probably measured in years. Most companies are not doing that work yet. And until they do, the gap between AI hype and AI returns is going to stay exactly where it is.
Timeline
- November 2022 - OpenAI releases ChatGPT publicly, triggering the current wave of generative AI investment and corporate interest
- 2023 - Most C-suite executives remain skeptical; AI experimentation is confined to a small number of early adopters
- 2024 - Corporate AI spending accelerates; enterprises begin distributing tools such as Microsoft Copilot widely across workforces; the "dabbling" phase intensifies
- Early 2025 - Microsoft bundles Copilot with E3 and E5 enterprise license tiers, removing the explicit cost barrier and sharply increasing the number of employees with access to AI tools
- 2025 - Goldman Sachs estimates global AI investment at approximately $200 billion; McKinsey's State of AI survey finds 88% of organisations use AI in at least one function, but only 6% report meaningful bottom-line impact
- 2025 - Goldman Sachs chief economist Jan Hatzius states AI's contribution to US GDP growth in 2025 was "basically zero"; MIT economist Daron Acemoglu projects AI will increase total factor productivity by less than 0.66% over the next decade
- April 15, 2026 - Microsoft restricts full Copilot functionality in Word, Excel, PowerPoint, and OneNote to paid licensees, formalising a two-tier AI access model in the enterprise
- June 4, 2026 - Dr. Rebecca Homkes speaks on Bloomberg Radio, arguing that AI adoption is high but "incredibly shallow," with only 10-15% of business work touching AI in any meaningful way, and that cross-functional productivity gains require workflow redesign that most companies have not started
Summary
Who: Dr. Rebecca Homkes, economist and lecturer at the London Business School and faculty at Duke Corporate Education, speaking on Bloomberg Radio
What: A detailed assessment of the state of enterprise AI adoption, arguing that near-universal tool deployment has not translated into meaningful productivity or revenue gains, with only 10-15% of business work currently involving AI, and cross-functional gains absent because they require workflow redesign rather than tool distribution
When: The interview was published June 4, 2026, drawing on data from across 2023-2026
Where: Global, with particular focus on large enterprises in the United States and Europe deploying tools such as Microsoft Copilot and products from OpenAI, Anthropic, and Google
Why: The economic stakes are significant: if AI delivers its promised productivity gains, global GDP growth accelerates substantially; if adoption remains shallow because organisations do not redesign their workflows, trillions of dollars of investment will have generated only marginal returns, and the timeline for meaningful impact shifts from years to potentially a decade or more
