Three returns, one ruler

Every AI investment committee eventually asks the same question: what is the return? It is the right question asked in the wrong singular. AI produces three different kinds of value, and only one of them answers to a payback period.

The other two are real, but they arrive in a different currency and on a different clock. Measure them with the investment ruler and they look like failures, right up until the moment they are cancelled. This insight sets out the three returns, what the 2026 evidence says about each, where the measurement goes wrong, and how to govern a portfolio that runs all three at once.

It is a companion to The number nobody decks, which asked how far the evidence of AI value has travelled, from the user to the P&L. This piece asks the question that comes before it: what kind of value is it in the first place?

1.Three business cases, three currencies

Gartner's framework, The 3 Business Cases of Generative AI Value (Nate Suda and Hung LeHong, January 2025), argues that AI outcomes range from everyday to game changing and cannot be described by a single currency of value. It separates three business case types. Defend cases yield employee returns. Extend cases create a return on investment. Upend cases create a return on the future. We abbreviate them here as ROE, ROI and ROF.

The important word is parallel. These are not three stages that an initiative climbs through. A healthy portfolio runs all three at the same time, for different reasons, measured in different ways.

2.Return on employee: the Defend case

The return on employee is the value of augmenting individual people: a relationship manager who drafts client correspondence in half the time, an underwriter who reads a policy file faster, an analyst who spends less of the week summarising. It is the most common case and the most diffuse. The value is spread thinly across thousands of working days rather than concentrated in one line of the accounts.

Adoption is now broad but uneven. Eurostat reports that 20 percent of EU enterprises with ten or more employees used at least one AI technology in 2025, rising to 55 percent of large enterprises against 17 percent of small ones. In a survey of nearly 6,000 executives across the US, UK, Germany and Australia, 69 percent of firms actively use AI, and executives who use it personally average about 1.5 hours a week. McKinsey's 2026 survey finds that 80 percent of its 1,719 respondents say AI has improved their productivity.

Wide, felt and shallow is exactly the profile of a return on employee. It is a real return. It is simply not a financial one yet, and demanding that it become one at the first review misreads what was bought. The leading indicator here is not payback. It is whether use is active, repeated and deepening, because without that there is no return of any kind.

3.Return on investment: the Extend case

The return on investment comes from re-engineering a specific process with a metric that already exists and already has an owner. It is the only one of the three that behaves like a conventional investment case, which makes it the easiest to fund and the easiest to over weight.

The clearest evidence comes from a single re-engineered process. Brynjolfsson, Li and Raymond, in The Quarterly Journal of Economics, studied a generative AI assistant rolled out to 5,179 customer support agents.

Two things make this an Extend case rather than a Defend one. The process had a metric before the AI arrived, and the change was measured on it. That is what lets a return on investment be proven. It also shows where the value concentrates: in compressing the gap between novice and expert, which is a design choice about who the tool is for.

Extend cases fail differently. The risk is not that value goes unmeasured, but that credit is claimed against a metric the initiative did not move. The executive survey from the Federal Reserve Bank of Atlanta shows how wide that gap can run: executives reported a 1.8 percent labour productivity gain from AI for 2025, while their own revenue and employment figures implied 0.6 percent. A return on investment has to be read in the revenue-implied number, not in the survey.

4.Return on the future: the Upend case

The return on the future is a strategic bet: a new market, a new product, or a public facing model that changes the firm's position rather than its cost base. For a bank it might be an advice service delivered largely by AI to a segment that was never profitable to serve. For an insurer, distribution through partner platforms. The right question is not what the payback is, but what position the bet buys and what it costs to keep the option open.

This is where the evidence is thinnest, and that is itself informative. In McKinsey's 2026 survey, 6 percent of respondents attribute 5 percent or more of EBIT to AI, unchanged from the year before. The top performers among them report pursuing growth or innovation alongside efficiency, and nearly three quarters report fundamentally redesigning workflows, against about a quarter of everyone else.

That tells us what the firms furthest along say they do. It does not measure a return on the future as a realised outcome, and no source we hold does. The honest position is that this return is governed by intent and milestones today, and that any number offered as its payback should be treated with suspicion.

5.The delay trap

The failure that the three returns are designed to prevent has a simple mechanism. An initiative that is really a return on employee, or a return on the future, is approved through the investment process because that is the only process there is. It is then reviewed against a payback curve it was never going to follow. It shows nothing on the P&L by the second quarter, and it is cut.

The 2026 data shows how much value currently sits on that slower clock. Across the four country panel, 89 percent of executives report no productivity impact from AI over the past three years, yet the same executives expect AI to raise labour productivity by 1.4 percent over the next three. Whatever one makes of the forecast, the value they describe is expected later than it is being judged.

European firm data points the same way. The Bank for International Settlements finds AI adoption associated with a one-off productivity level shift of about 4 percent among non-financial firms, an effect amplified most by investment in training. Training is a cost incurred now for a return that arrives later, which is precisely the kind of spend a payback gate is built to reject.

Judged on an investment timeline, almost every employee or future return looks like a failure, right up until it is cancelled.

6.Three gates, three rulers

The fix is a governance discipline, not a better model or a better survey. Auron applies it at the three gates every initiative passes through.

1. At the Fund gate, name the return before the number. Every initiative is tagged ROE, ROI or ROF, and Defend, Extend or Upend, before a business case is written. The tag decides which ruler applies for the rest of its life.
2. At the Scale gate, judge each return on its own measures. A return on employee scales on active and deepening use. A return on investment scales only if its owned metric has moved against baseline at full volume. A return on the future scales when its milestones are hit and the option is still worth holding.
3. At the Run gate, keep separate funding pools. Employee returns are run as operating cost, investment returns are measured net of the AI spend that produces them, and future returns are held to a venture budget with a defined exit. The investment lens never sets the bar for the other two.

7.Which return is this initiative chasing?

Before any AI business case is written, five questions settle which ruler it should be held to. Answer them for the initiative, not for the technology.

8.The limits of this evidence

• The three-way split is a governance discipline to impose, not a description of how firms currently measure AI. No source we hold shows firms separating the three returns in practice.
• Only the support study measures a return directly, in one firm and on an early model. The BIS result is derived from firm accounts. Every other figure here is reported by executives or respondents.
• No source measures a return on the future as a realised outcome. The evidence for it is about intent and practice among the firms furthest along.
• The executive surveys are anchored in the US, UK, Germany and Australia, and the European firm data excludes financial firms and contains no Swiss firms. Read across to Swiss banks, insurers and MedTech firms is a hypothesis to test, not a finding.
• The named cases rely on what each organisation has published or told the press. Only the GI Genius trial is independently peer reviewed. The bank, insurer and asset manager scenarios in the text are illustrations, not documented cases.

Before the next AI business case reaches your committee, ask which of the three returns it is chasing. Then ask whether the ruler you are about to hold it to was built for that return.

Auron Advisory is formed to build the intelligent enterprise: organisations where AI and disruptive technologies not only increase productivity but open new markets and materially improve the ones already there.

Sources