AI ROI and Value Realization

Between August and September 2025, Deloitte surveyed 1,854 senior executives in 14 countries across Europe and the Middle East about their returns on AI. Technology investments are usually expected to pay back within seven to 12 months. Before reading on, guess how long most of these executives said a typical AI use case took to deliver a satisfactory return, and what share had seen payback within a year.

Most said two to four years. Only 6 percent reported payback in under a year. In the same survey, 85 percent had increased their AI investment over the previous 12 months, and 91 percent planned to increase it again¹.

The two headline figures measure slightly different things. Two to four years is how long most respondents said a typical use case took to reach a return they judged satisfactory; 6 percent is the share reporting full payback within a year. These are self-reported figures, and the gap they describe is not proof that AI fails to pay. It is proof that the return arrives later and less predictably than the approval slide implied. The business case that won the funding contained a forecast. The organization then spent two to four years finding out what the result was. Whether that second number was ever measured, explained and acted upon is the subject of value realization.

The core idea

Return on investment is calculated twice. The forecast ROI is computed at approval from expected benefit and expected cost; every input is an assumption. The realized ROI is computed after the money is spent, from benefit that was measured and cost that was paid. Approval is not capture. Between the two calculations, people have to adopt the tool, the workflow has to change, quality has to hold and freed capacity has to turn into money. Each step can move the second number away from the first.

This chapter uses the base case from Building the AI Business Case throughout, rather than a new example. That case expected 10 million of benefit over three years, after discounting potential value for an adoption ramp and for how much freed capacity would be realized. Its full cost was 5 million: 2 million once, then 1 million a year to run. With a 70 percent chance of delivery, its risk-adjusted value was 7 million minus 5 million, or 2 million. The previous chapter built that forecast. This one turns it into a return and then sets the return against what happens.

Four measures of one base case

Four measures answer four different questions about the same cash flows. Net value asks how much is left after cost: 10 million minus 5 million, or 5 million over three years. ROI asks how large the return is per unit spent: net value divided by full cost, here 100 percent. Payback asks how long it takes to recover the money. The base case loses 1 million in the first year and earns 2.5 million in the second, so cumulative net value turns positive about 0.4 of the way through year two, a payback of roughly 1.4 years. Net present value asks what the flows are worth today once later money is discounted at the organization’s required rate of return².

Each measure hides something the others show. Payback ignores the time value of money and everything that happens after the money is back³, so a project that pays back quickly and then stops earning looks better than it is. ROI ignores timing altogether: 5 million of net value earned in year one and the same amount earned in year five give the same percentage. Net present value corrects for timing. At an illustrative 10 percent rate, the base case’s net flows of minus 1, plus 2.5 and plus 3.5 million are worth about 3.8 million today, not 5 million. Your finance team may define these measures differently, for example dividing by the initial investment rather than full cost. That is acceptable. What is not acceptable is a review that uses a different definition from the forecast.

Set the payback against the survey in the opening. The base case recovers its money in about 1.4 years, while most executives in that survey reported two to four years to a satisfactory return on a typical use case. That does not make the case wrong. It makes the adoption and realization assumptions behind it the first things to measure.

Risk-adjusted ROI describes a portfolio, not a project

The risk-adjusted value of 2 million gives a risk-adjusted ROI of 40 percent: 2 million divided by 5 million. That is the right number for comparing this bet with others. It is the wrong number to expect from this one. A single initiative does not return 40 percent. If it is delivered, it is expected to return 100 percent. If it fails after the full cost is spent, it returns minus 100 percent. Weighting the two outcomes by 70 and 30 percent gives 40 percent, but no individual project lands there.

This matters at review. Once the system is live and working, delivery risk has resolved, and the realized result should be compared with the forecast for a delivered system: 10 million of benefit, 100 percent ROI. Judging a working system against the risk-adjusted 40 percent flatters every shortfall. Risk-adjusted forecasts are tested across a portfolio of initiatives, which Module 09 treats in full.

Why ROI falls faster than benefit

Suppose the base case is delivered on cost, but only 6 million of the 10 million expected benefit arrives. Many people guess that ROI falls to about 60 percent. It falls to 20 percent: 6 million minus 5 million, divided by 5 million. The organization captured 60 percent of the value and lost 80 percent of the return. Cost does not shrink when benefit does, so every shortfall in benefit comes straight out of the net value.

At half the expected benefit, the base case only breaks even. That is the same point the previous chapter found as a switching value, the level an assumption must reach before an option stops being worth doing⁴: realization falling from half of freed capacity to a quarter. The leverage works in both directions. Every point of benefit recovered after launch is worth twice as much in ROI, which is why managing realization often pays better than renegotiating the license.

The share of expected benefit actually captured, 60 percent in this example, is a useful number in its own right; the AI Value Scorecard, later in this module, reports it as the realization rate. It is not ROI, and a review that reports one as if it were the other will either panic or relax at the wrong moment.

Explaining the gap: four variances

When realized value falls short, the instinct is to declare the project a success or a failure. A finance team would do neither. It would explain the variance, the difference between plan and actual, by cause, because each cause has a different owner and a different remedy. Four variances cover the usual causes of an AI shortfall.

The adoption variance is the gap in volume: fewer users, desks or cases than the ramp assumed. The benefit per use variance is the gap in rate: each use saves less time or earns less than planned, often because quality concerns add review. The conversion variance is the gap between freed capacity and money; as Productivity vs Realized Capacity showed, hours that are absorbed into the day are capacity, not cash. The cost variance is the gap between planned and actual full cost. The first three reduce benefit; the fourth raises cost. Measured separately, they tell the organization whether to train, redesign, cut a budget line or renegotiate.

Owning the benefit

Variances can only be explained if the benefit was defined, baselined and owned before launch, and organizations have long been weak at this. The UK’s National Audit Office found that government “does not routinely look at what happens after major projects are completed,” and that experts saw more effort spent identifying potential benefits to make the case for investment than ensuring projects achieved their purpose⁵.

The same report explains why ownership cannot sit with the delivery team. Value “is generated by how organisations or stakeholders adopt and make use of project outputs in practice,” and it is “rarely the case that a project delivery team alone can deliver the full value of a project”⁵. For AI, the team that builds the system cannot change a newsroom’s rota or cut a freelance budget. A business leader can. That is also why the benefit and the money are kept apart: the business owns the benefit, and finance validates the money.

Put one such line in a register for each benefit and the register becomes the instrument for the review. Its targets should follow the year-by-year profile of the business case, so that a first-year shortfall is visible in the first year and not averaged into a three-year total. The evidence column matters more than it looks. A benefit described as “hours saved” has no place in the accounts; a benefit described as “freelance spend on production work, from the invoice ledger” does, and finance can confirm it.

Review forward, not backward

A realization review asks two different questions, and it is easy to merge them. The first looks back: what return did the money spent produce, and why did it differ from the forecast? The answer improves the next business case, because a measured result is a better outside view than any supplier’s figure, the point Building the AI Business Case made about borrowed evidence. The second looks forward: from today, is the value still to come worth the cost still to be spent?

Only the second question decides whether to continue. Money already spent cannot be recovered by stopping, so it should not weigh on the choice, in either direction. A project with a poor backward-looking ROI can be well worth continuing, and a project that looks fine on paper can deserve to stop if its remaining value has collapsed. The AI Value Scorecard sets out decision rules for scaling, adjusting, pausing and stopping. What this chapter adds is the arithmetic those rules rest on: decide on forward value, and explain backward variance.

Story: a publisher’s first year

The case that follows is an illustrative composite; the organization is unnamed and the figures are the base case from the previous chapter, carried forward so the arithmetic can be followed.

A regional publisher of daily newspapers, local news sites and a handful of magazines approved an AI production tool. It drafts story summaries, newsletter versions, headline options, search metadata and archive tags, and editors review every draft. The case was the base case: 10 million of expected benefit over three years, mostly from lower freelance and contractor spend on production work and from hiring avoided for a planned newsletter expansion, against 5 million of full cost. The digital product team that built the tool was named as owner. Its register tracked weekly active users and estimated hours saved, and the first review was set for month 12.

At month 12 the tool worked, so delivery risk had resolved, but the numbers did not match the plan. Finance measured 1.2 million of benefit against the 2.0 million planned for year one, and 3.4 million of cost against 3.0 million, because the archive’s rights metadata had to be cleaned before the tool could tag it. The first-year net was minus 2.2 million instead of minus 1.0 million. Split into variances, the benefit gap told a clear story. Adoption reached 30 percent of production staff, not the 40 percent planned, because the sports and features desks kept their old workflow: 0.5 million. After one summary misstated a criminal charge, the standards editor required a second human read of every summary, which cut the time saved on each: 0.2 million. And two desks kept their freelance budgets unchanged, so freed hours were absorbed rather than banked: 0.1 million.

The board then looked forward. The 3.4 million already spent was gone whatever it decided. Stopping would save the remaining 2 million of running cost and forgo all remaining benefit, a forward value of zero. Continuing unchanged, at 60 percent of the planned 8 million for years two and three, would bring 4.8 million of benefit for 2 million of cost: plus 2.8 million. A fix costing 0.2 million, a desk-by-desk adoption plan, a second-read rule limited to court and crime stories with a checklist, and freelance budget lines cut by the managing editor, now the benefit owner, was forecast to lift capture to 85 percent: 6.8 million of benefit for 2.2 million of cost, plus 4.6 million.

The board continued and fixed. It also held the planned extension to the magazine division until monthly adoption reached the 70 percent the case assumed for year two. The re-forecast was honest: 8.0 million of benefit for 5.6 million of cost over three years, an ROI of about 43 percent rather than 100, a payback of about 2.2 years rather than 1.4, and a net present value of about 1.6 million at 10 percent rather than 3.8. Without the fix, the three-year ROI would have been about 11 percent, close to the conservative scenario the case had shown at approval. The 85 percent is itself a new forecast, and the managing editor now reports it monthly, desk by desk.

What this means for leaders

The business case is a hypothesis; realized ROI is the result. A leader’s job after approval is to make sure the result is measured, explained and used. That means insisting on a benefits register with a baseline, a target by year, a named business owner and evidence finance can verify; comparing a working system with its delivered forecast, not its risk-adjusted one; explaining each shortfall as adoption, benefit per use, conversion or cost; and deciding on the value still to come, never on the money already gone.

Two habits make a large difference. Review early, because a first-year variance found in month three can still be fixed in year one. And feed the realized numbers into the next case, so that the organization’s forecasts improve over time instead of repeating the same optimism with new technology.

Check yourself

1. If 60 percent of the expected benefit arrives and cost is on plan, a 100 percent forecast ROI falls to about 60 percent.MythFact
2. A short payback period shows that an investment creates value.MythFact
3. A risk-adjusted ROI of 40 percent is the return a single delivered initiative should be expected to make.MythFact
4. Money already spent on an AI initiative should not decide whether to continue it.MythFact
5. The team that built the AI system is the right owner of its benefits.MythFact
6. An initiative with a disappointing realized ROI can still be worth continuing.MythFact

Reflection: an investment already live

What comes next

This chapter followed one benefit line, the money a case promised and the money it delivered. Many AI initiatives do more than that: the publisher’s tool also changed how fast stories reached readers, how editors spent their time and how exposed the newsroom was to errors. The next chapter, Total Business Impact, brings those effects into one picture and shows how to count each of them once.