I want to be upfront about something. I am not going to offer a formula for calculating the return on AI investment, because no reliable, universal one exists. What I can offer is a different way of framing the question, because I think most organisations are currently answering it with the wrong data, against the wrong benchmarks, over the wrong timeframe.
The evidence on AI’s value is, in fact, far more contested and more interesting than most corporate conversations acknowledge. Understanding the debate properly, I would argue, is a prerequisite for making good decisions about AI investment right now.
A genuinely contested question
The sceptical case deserves to be heard first. Goldman Sachs chief economist Jan Hatzius stated recently that AI’s contribution to US GDP growth in 2025 was ‘basically zero’, adding that there had been significant misreporting of its macroeconomic impact. A separate Goldman analysis of corporate earnings found no meaningful relationship between AI adoption and productivity at the economy-wide level.
Nobel Prize-winning MIT economist Daron Acemoglu has argued in a peer-reviewed paper in Economic Policy that AI is likely to increase total factor productivity by less than 0.66% over the next decade. His deeper concern is what he calls ‘so-so automation’: AI deployments that allow companies to cut headcount without generating real productivity gains, eroding human capability while delivering nothing genuinely transformative.
These are serious arguments, and they should not be waved away. But they are, importantly, arguments about macroeconomic measurement over a 10-year horizon. They tell us less than is often assumed about what is happening inside individual organisations right now.
The measurement gap
Here is where the picture becomes more interesting. The same Goldman Sachs analysis that found no economy-wide signal also found that organisations successfully integrating AI into specific functions reported median productivity gains of around 30%, primarily in software development and customer service. The absence of an aggregate signal is not evidence that AI is not working. It is evidence that the gains are concentrated and localised, not yet widespread enough to move national statistics.
A consistent pattern emerges across major industry surveys. McKinsey’s 2025 State of AI report, drawing on nearly 2000 executives across 105 countries, found that meaningful bottom-line impact remains rare: only 6% of organisations qualify as high performers with an EBIT impact of 5% or more. But those organisations share a common approach. They redesign workflows rather than bolting AI on top. They invest in people and process, not just platforms. And they are led by executives who actively champion adoption rather than delegating it to IT.
Meanwhile, the Wharton Human-AI Research annual study found that three out of every four business leaders already report positive returns on their generative AI investments, while Google Cloud’s second annual ROI of AI study reported that 74% of executives achieve ROI within the first year.
How do we reconcile these conflicting narratives? Researchers at UC Berkeley’s Sutardja Centre put it well in a 2025 analysis of the AI measurement problem: we are not experiencing an AI failure. We are experiencing a measurement failure. The instruments we use to evaluate technology investments were built to count things that are large, visible, and land clearly on a balance sheet. What AI is doing, for most organisations, is none of those things.
Where the returns are actually accumulating
In my view, AI return on investment is accruing in four distinct areas, none of which are well captured by conventional measurement.
The first is the accumulation of small, daily recoveries of human time. Consider what ordinary knowledge work looked like before AI became genuinely capable. Drafting a carefully worded email to a difficult client, summarising months of data for a board presentation, producing a first draft of a lengthy report: each took significantly longer than it does today. According to the Federal Reserve Bank of St Louis, employees using generative AI are saving an average of 5.4% of their working hours each week, with frequent users recovering more than nine hours. IBM’s 2025 Race for ROI study of 3500 senior executives found that 66% reported significant operational productivity improvements. None of this appears on a P&L, but it is compounding, every working day, across every person in an organisation.
The second source of return is what people do with the time AI returns to them. When the tasks consuming someone’s afternoon are handled more efficiently, the recovered hours become available for the kind of thinking that operational load has been quietly crowding out: the strategic problem that keeps getting deferred, the client relationship that deserves more attention, the creative work that requires uninterrupted concentration. A senior partner at a professional services firm described it to me this way. For years, she had arrived at the office at seven in the morning to secure two hours of genuine thinking before the emails began. AI gave her those two hours back in the middle of the day. The work did not change. The quality of her strategic thinking did.
The third source is a genuine capability shift. AI is enabling organisations to do things that were previously beyond their practical reach, not just do existing things faster. At my company, The Strategy Group, we recently migrated to a new project management platform and encountered a data compatibility issue that would have typically required two weeks of painstaking manual restructuring.
Instead, we used Claude to build a bespoke conversion application in approximately 30 minutes, without a developer or an IT request. A task that previously would have required two weeks of skilled labour was finalised before lunch. Every organisation has a version of this story: a problem sitting in the ‘too difficult, not worth the cost’ category for years. AI is making those problems tractable, and the value created is real even when it is entirely invisible to traditional ROI measurement.
Human-centred Intelligence: the fourth return
The fourth source of AI ROI is the one I find most compelling, and the one I believe will prove most consequential over time. It requires introducing a concept I have been developing: Human-Centred Intelligence.
Most practitioners are familiar with human-centred design, the principle that the best products and experiences are built around a genuine understanding of the person at their centre. Human-Centred Intelligence extends that principle into the age of AI. It describes the point at which AI and the human professional meet, not to replace one another, but to give the human something they have rarely had in adequate supply: the time, the headspace, and the capacity to genuinely focus on the customer.
For most people in client-facing roles, authentic customer focus has always been in competition with operational reality. The desire to prepare thoroughly for every meeting, to follow up thoughtfully, to anticipate what a client needs before they have articulated it: all of this gets crowded out. When AI handles preparation, analysis, follow-up, and reporting, the professional can be genuinely present with their client in a way that the operational load previously made impossible. From that position, it becomes feasible to deliver experiences that were not previously available: personalisation that previously required an entire team, and the kind of anticipatory service that builds lasting loyalty.
The commercial consequences are measurable. McKinsey’s research on AI-powered customer experience found that AI-enabled approaches can enhance customer satisfaction by 15 to 20%, increase revenue by 5 to 8%, and reduce cost-to-serve by 20 to 30%. Customer retention research consistently shows that a 5% improvement in retention can lift profits by 25 to 95%, and that acquiring a new customer costs five to seven times as much as keeping an existing one. Human-Centred Intelligence is what happens when AI gives professionals the freedom to be at their human best with the people who matter most to their business. That is not a soft benefit. It is a compounding commercial advantage.
The layoff headlines, the payback curve, and what it all means
There is one obvious objection to the argument that AI’s value is distributed and hard to measure. If that is true, how have the Block, Duolingo, Atlassian, and others found it clearly enough to cut hundreds of roles and, in the main, cite AI as the reason? The answer is that those companies are measuring something different: headcount reduction.
When AI handles work that previously required 50 people in a contact centre, that saving is large enough to appear directly on the wage bill. That is genuine ROI, but it is a specific kind, available to organisations running large volumes of repeatable process-driven work at scale. It is not the same as what is accumulating across every desk in organisations that are using AI more broadly. The deeper risk, as Prof. Acemoglu warns, is that organisations automate the wrong things, cutting costs at the margin while destroying the human capability that determines long-term competitive advantage.
Underlying all of this is a timeline problem. Deloitte’s 2025 survey of 1854 senior executives found that most organisations are achieving satisfactory ROI on AI over 2-4 years, significantly longer than the seven-to-twelve-month payback conventionally expected from technology investments. That gap exists because AI is not software you deploy; it is a change in how people work. Such changes have always required longer periods to compound into measurable outcomes.
McKinsey estimates the long-term opportunity at $4.4 trillion in additional productivity growth from corporate use cases. Goldman Sachs, in its more conservative modelling, expects AI-driven productivity gains to begin materialising meaningfully from 2027, building through the late 2030s. The bulls and the bears are largely debating timing, not direction.
The proof that AI is working will not arrive as a clean number on a quarterly report. It will show up as time reappearing in people’s days, problems getting solved that previously sat untouched, customers experiencing something qualitatively different, and capabilities that simply did not exist six months ago. The task for leaders is not to find a better formula for calculating AI ROI. It is to develop organisational literacy to recognise value in forms that existing instruments were never designed to see.
Jeffrey Tobias is an accomplished and prominent innovation thought leader and strategist, drawing expertise from the academic, government, entrepreneurial and corporate worlds. He serves as an Adjunct Professor and Fellow at AGSM @ UNSW Business School, and holds a B.Sc (Hons), University Medal and PhD from UNSW Sydney. In 2003, he founded The Strategy Group, where he currently serves as Managing Director.
This article was originally published by UNSW’s BusinessThink research platform.
Disclaimer: This article is for informational purposes only and does not constitute financial advice. Past performance is not indicative of future results, and investors should consider their own circumstances before making investment decisions.