AI ROI belongs on the same page as every other operating number the COO already reviews. If an AI system is working, it shows up in cost per case, cycle time, backlog and error rates for the process it runs. If it does not show up there, the return is a story, however good the demo was.
Most AI programs cannot show a return yet
The large surveys agree on the scale of the gap.
- IBM's 2025 study of 2,000 CEOs found that "only 25% of AI initiatives have delivered expected ROI over the last few years, and only 16% have scaled enterprise wide."
- MIT NANDA's "The GenAI Divide" (July 2025) estimated enterprise investment in generative AI at $30 billion to $40 billion and concluded that "95% of organizations are getting zero return."
- BCG's October 2024 research was published under the headline "74% of Companies Struggle to Achieve and Scale Value."
Each study defines success its own way, and none of them should be read as a precise failure rate. Read together, they say something simpler: most companies spending on AI cannot point to a number that moved.
Model metrics do not add up to ROI
The numbers AI programs most often report are the ones easiest to collect: accuracy on a test set, number of users, prompts per week, survey estimates of hours saved. They describe the tool. None of them tells a finance team whether the operation got cheaper, faster or better.
Hours saved is the most misleading. An agent who saves twenty minutes a day on drafting does not reduce the cost of the process unless the work is redesigned around that time: fewer hand-offs, a shorter queue, a smaller backlog, or the same team absorbing more volume. Without that redesign, the saved minutes dissolve into the day.
MIT NANDA chose a stricter bar for its own research. It counted a task-specific tool as successfully implemented only when users or executives described "a marked and sustained productivity and/or P&L impact." That is the right bar for an operations leader, and it can only be met by measuring the operation.
The return lives in the process owner's numbers
| Operating metric | What it captures | Where the baseline comes from | Trap to avoid |
|---|---|---|---|
| Cost per case | Labor, tools and rework divided by cases closed | Finance and workforce data for the process, before the build | Leaving out the time people spend on exceptions and reviews |
| Cycle time | Time from case opened to case resolved | Timestamps in the system of record | Measuring only the AI's step, not the whole case |
| Backlog | Cases waiting beyond target | Queue reports | A falling backlog from a seasonal dip |
| Error and rework rate | Cases reopened, corrected or escalated by customers | Quality audits, reopen rates, complaints | Sampling only the cases the AI handled confidently |
| Exception rate | Share of cases routed to people | The new system's logs | Treating a low rate as good when the system is guessing |
| Cash or revenue protected | Credits avoided, disputes resolved in time, churn prevented | Billing and retention data | Claiming revenue the process never influenced |
Two rules make the table work. Take the baseline before anything is built, because once a pilot starts, nobody remembers what normal looked like; the first step of the AI implementation guide produces it. And measure on live volume, because a test set is a sample somebody chose.
Attribution needs a comparison, not a before-and-after story
Before-and-after numbers flatter any change. Volumes move with the season, a policy change cuts complaints, a new hire clears a backlog. To know what the AI system did, compare like with like: run a controlled share of live volume through the new process while the rest follows the old one, or compare the same weeks year on year, and keep the comparison running until the pattern is stable.
This is where many organizations believe they are better prepared than they are. In IBM's CEO study, 68% of respondents said their organization has "clear metrics to measure innovation ROI effectively," yet the same respondents reported that only a quarter of AI initiatives delivered the expected return. Clear metrics are not enough without a baseline and a comparison.
Back-office returns are underrated because they are harder to attribute
Where AI money goes says a lot about how ROI gets measured. MIT NANDA found that "Budgets favor visible, top-line functions over high-ROI back office," and argued that this pattern "reflects easier metric attribution, not actual value." Sales and marketing pilots produce numbers that are easy to claim. Back-office processes produce savings that sit in operating budgets and take baselines to see.
The same report noted that the highest-performing organizations "report measurable savings from reduced BPO spending and external agency use, particularly in back-office operations." For a COO, the lesson is to start where the numbers already exist. Processes in back-office operations, such as billing inquiries, reconciliations and exception queues, usually have volume data, cost data and service levels. That makes their return measurable from the first week, as described in back-office automation.
The full running cost belongs in the calculation
An honest ROI subtracts everything it takes to keep the system working:
- model and infrastructure usage, which rises with volume;
- engineering to maintain integrations as connected systems change;
- monitoring, sampling and periodic review of decisions;
- the time people spend on exceptions, approvals and quality checks;
- the effort of changing roles and habits in the team.
A pilot that leaves these out will look better than production ever does. A deployed system has to earn its return after all of them, every month.
A pilot that cannot measure its return should not end in a yes
The practical test is simple. Before a pilot starts, write down the operating metrics it must move, the baseline for each, the running cost it will carry and the decision those numbers will trigger. If the pilot cannot be measured that way, it is not ready to start. If it can, the decision at the end is arithmetic, and the argument for production makes itself. The pilot-to-production post covers how to scope that pilot.
Put the same test to any partner, including the engineers behind AI implementation services: which process numbers will the pilot answer to, and who agreed them?