The return on intelligence (ROI)

The same AI budget can fund very different kinds of investment. Each has a different path to return, a different time horizon, and a different set of things that must go right.

Targeted AI applications for a single use case are often quick wins. When AI upgrades a task or basic process the company already does, with very little else having to change, the results are usually positive. Think of the notes a contact-center agent writes up after every call. AI can draft them in seconds. The company already knows how many calls it handles and how long each writeup takes, so the saving is easy to see and count. Implemented well, the investment can pay for itself quickly. High-volume repetitive human tasks are areas where returns can be calculated easily and should inform prioritization of funding before deployment.

Workforce enablement is one of the hardest returns to pin down. The company gives employees AI tools and expects them to work better. This can range from engineering teams leveraging agentic coding harnesses to individual workers leveraging desktop AI tools. There is intangible value in employees learning the art of the possible as well as more tangible value in the productivity gains, if used properly. The perceived upside can be visible at companies reporting over 10% cost-reduction and/or revenue uplift from AI: 92% of employees report productivity gains, and 44% say AI enables them to do at least twice as much work (1.6 times the share among employees at other companies). But even real gains may be spread thin across thousands of people: a few minutes saved here, a better draft there. In some cases, AI expenses will outweigh the return, at least in the near term.

AI foundations are the data, processes, technology, and governance investments that other AI investments depend on. They cover things such as data capture and quality, development of enterprise approaches to deploy AI safely and cheaply, and reimagining the process required to make sure AI isn’t solidifying poor practices and approaches. Many companies had identified these data and technology gaps long ago. Generative AI makes the cost of leaving them unresolved larger, and the potential payoff from fixing them greater. Much of the spending goes into repairing the underlying data and technology environment. The return shows up later through the AI applications and transformations those foundations make possible.

AI-enabled business transformation is the biggest investment and can lead to the biggest return. Transformation initiatives have the potential to produce the largest returns because they require the company to change how it operates and serves customers. This may include rethinking its offerings, how value is created, and how people and machines divide the work across a business line or the whole enterprise. That ambition also makes the return harder to deliver and slower to prove. Products, workflows, roles, technology, controls, incentives, and capacity may all have to change together. Building the AI capability can prove easier than making the organizational changes required to capture its value. The relevant question, therefore, is whether the company is making measurable progress toward its transformation objectives, not whether individual agents are being deployed. This involves retiring old processes, reallocating people and capacity, and ensuring that the potential long-term return justifies the full cost, time, and execution risk of the effort.

These four categories of investment aren’t mutually exclusive. A company should be making several kinds of AI investment at once, and a single initiative can expand over time. For instance, a targeted application might reveal that the underlying data needs to be rebuilt while broad workforce adoption may eventually require roles and workflows to change. The complexity demands that leaders be clear about which type of investment the company is funding at each point, because each one should be judged against a different path to return. We explore this notion of strategic portfolio discipline later.

The gain from AI is often easy to feel, but much of the cost is easy to miss. That combination can make the ROI case look cleaner than the economics really are. Consider a customer-service AI that resolves a larger share of inquiries without a human agent. In theory, this sounds promising, but if customers call back more often, escalations take longer, or remediation costs rise, some of that saving disappears. Automating more customer interactions creates value only if the actual cost and quality of resolution improve.

Many efforts to measure the return on AI fail to capture fully loaded costs associated not only with the one-time build but also the need to maintain and evolve the agentic infrastructure over time. Our employee data shows how material some of that surrounding work can be. Workplace AI users report spending an average of 22 minutes of checking, correcting, or redoing the output for every hour of AI-assisted work. This shows how the price of the AI tool alone can materially understate the cost of producing usable work.

Companies may overstate the benefit as easily as they understate the cost. Faster output creates less value if quality falls or the output doesn’t lead to actual impact. If a financial case assumes AI will save employee time, the company needs to know what that time will be used to improve: output, cost, service levels, speed, or another measurable result. A credible AI ROI calculation counts the full cost of producing the result and measures the business outcome that actually changed.

Some of the ROI work that matters most has to happen before the investment is approved.

In practice, we see many companies racing to experiment and demonstrate that they are adapting without giving full thought to where the value will truly be generated over the short and longer term. Failing to do so in advance risks leadership engaging in the practice of retroactive storytelling later on.

Instead, management should start by asking a simple question: If this investment works, what return do we expect, and what has to be true to get it? The answer usually reaches well beyond the AI tool. A new AI pricing tool, for example, may pay off only if the company first cleans up its pricing data, switches off the old pricing system, and gets sales teams to act on the tool’s recommendations. The expected return plus everything that has to be true to get it is the investment’s return thesis. If nobody has a credible plan for one of those requirements, the expected return should come down before the money is committed.

A return thesis has to involve much more than the finance department because change is complicated and has many downstream dependencies. Technology may own the AI capability while operations changes the workflow, human resources changes roles or staffing, risk approves new controls, and the business at large acts on the resulting capacity or customer opportunity. The investment case is credible only when the people who control those decisions understand what the return depends on and are prepared to act.

As Chapter 2 on the bionic workforce shows, the majority of executives lack a clear, shared view of where their company’s AI is headed, let alone what each return depends on.

A business case for ROI can be written before anyone knows the return precisely. Management can start with a range of likely returns, release money in stages, and agree in advance what result at each stage would justify scaling up, redesigning, pausing, or stopping. Each stage should test whether one of the requirements can be met. For the pricing tool, a first stage might test whether sales teams act on its recommendations before the company commits to switching off the old system.

An experiment also can create value through learning. The company should still specify what it needs to learn, which assumption the experiment is testing, and which investment decision will change if the evidence comes back differently. By the time the investment reaches scale, management should know what the return depends on, which assumptions remain uncertain, who controls the critical decisions, and what evidence will determine whether more capital follows. That is how the company improves the odds of creating the return before it has to measure it.

AI can create capacity without creating value. Turning one into the other requires an ROI harvester: a senior owner who is accountable from the outset for the intended business result and has end-to-end authority to make the business decisions to capture it.

A company can correctly classify its AI investments, make a full cost-benefit analysis of each one, and assign the accountability for delivering returns, yet still end up with a disappointing result. What’s needed is the discipline to prioritize investment in areas that are critical to the enterprise’s long-term success.

We call this portfolio discipline, and it begins with leaders looking at their broad business strategy. Are we prioritizing for growth? Margin? Customer experience? Resilience? Risk reduction? All of these call for different combinations of targeted applications, workforce adoption, foundational and transformation investments. A company competing on margin might concentrate on targeted applications in its highest-volume operations, along with a process redesign to turn faster work into lower unit costs. One competing on growth is more likely to need strong customer data foundations and product and sales enhancements that bring new offers to market. A priority on resilience or risk reduction would shift weight toward governance, controls, and data quality.

Without that strategic anchor, the portfolio tends to reflect the use cases that arrive first, have the strongest sponsor, or offer the easiest payback to demonstrate. The result is a sum of separate functional decisions rather than a coherent set of investments that compound advantage.

In practice, many companies begin with near-term use cases, and that can be helpful in mobilizing the organization toward consistent ROI. They can free up cash for later bets, build confidence, and show what the enterprise needs next. But they also can become distorting when their short payback becomes the portfolio rule. Applications with clear baselines and near-term savings can readily clear a conventional business-case hurdle. Data, governance, workforce capability, and operating-model redesign may struggle to clear that bar because their returns are more diffuse, arrive later, or show up in other teams’ results. A short-term approach steers capital toward the returns that are easiest to prove quickly and away from the investments that the business strategy depends on.

Quick wins carry a secondary problem: Benefits from tools competitors can easily buy are unlikely to distinguish a company for long. Returns that create lasting advantage are more likely to come from growth, which is what our CEO data shows. Nearly four years after the launch of ChatGPT, only about one in eight CEOs surveyed report AI-driven cost reduction or revenue uplift above 10%. Yet those CEOs are 4.4 times as likely as CEOs who cannot yet point to returns to have deployed AI at scale in new products, services, and revenue streams. A portfolio of AI investments designed around easy-to-prove gains is unlikely to generate outsized returns.

Aaron Fine is a Managing Partner and Head of Americas for Oliver Wyman

Ana Kreacic is Partner and Chief Knowledge Officer of Oliver Wyman, and COO of the Oliver Wyman Forum

Ashik Ardeshna is a Partner and Co-Head of Quotient, AI by Oliver Wyman in the Americas, overseeing the firm's AI and analytics initiatives

Edwin Anderson is a Partner in Oliver Wyman's Banking and Financial Services practice

Eric Czervionke is a Partner in Oliver Wyman's Banking and Financial Services practice

John Romeo is a Managing Partner at Oliver Wyman and CEO of the Oliver Wyman Forum

Simon Luong is a Program Director at the Oliver Wyman Forum