Most AI ROI reports are too flattering.
They count hours saved. They count prompts submitted. They count tickets summarised, slides drafted, emails generated, and lines of code produced. Then someone converts those numbers into a productivity estimate and declares success.
Frankly, that is not ROI. It is activity dressed up as value.
The uncomfortable question is not whether AI made one task faster. The question is whether the enterprise became better off after the second-order costs arrived: review effort, rework, governance, security checking, change management, exception handling, training, integration, and the operational drag of poorly redesigned workflows.
That distinction matters in 2026 because AI is moving from personal productivity into core operations. If CIOs and CFOs keep approving AI programmes based on superficial efficiency metrics, they will overfund shiny tools and underfund the foundations that make AI useful.
The Hours-Saved Trap
The easiest AI metric is time saved. It is also the easiest to misuse.
If an analyst uses AI to draft a report in 30 minutes instead of two hours, the dashboard records 90 minutes saved. But what happened next? Did a manager spend another hour correcting weak assumptions? Did legal review take longer because source traceability was poor? Did the report change a decision, reduce risk, increase revenue, or improve customer experience?
I once advised a regional operations team that was proud of its AI-generated weekly summaries. The team had calculated thousands of hours of annual savings. When we walked the workflow, the picture changed. Managers were reading more summaries, not fewer. They were checking source documents because they did not fully trust the output. The AI had reduced drafting time but increased review load. The net benefit was still positive, but nowhere near the headline number.
That is the first rule of AI ROI: do not measure the task in isolation. Measure the system around the task.
The Evidence Is Already Warning Us
The market data is not saying that AI has no value. It is saying that value is uneven, and that weak measurement hides the gap.
Gartner reported in April 2026 that organisations with successful AI initiatives invest up to four times more, as a percentage of revenue, in foundations such as data quality, governance, AI-ready people, and change management than organisations with poor AI outcomes. The same Gartner survey found that only 39% of technology leaders were confident their current AI investments would have a positive impact on financial performance.
IBM’s 2025 CEO study tells a similar story from the top of the organisation. Surveyed CEOs said only 25% of AI initiatives had delivered expected ROI over the previous few years, and only 16% had scaled enterprise-wide. IBM also found that 72% of CEOs considered proprietary data key to unlocking generative AI value, while 50% said rapid investment had left them with disconnected technology.
BCG’s research is equally direct. In a survey of 1,000 senior executives, only 26% of companies had developed the capabilities to move beyond proofs of concept and generate tangible value. Just 4% had developed cutting-edge AI capabilities across functions and were consistently generating significant value. Seventy-four percent had yet to show tangible value from AI.
These are not anti-AI numbers. They are anti-theatre numbers. The winners are redesigning work, improving data, funding change, and measuring outcomes that matter.
Code Generated Is Not Software Delivered
The most obvious version of misleading AI ROI is in software engineering.
Counting code generated by AI is like counting bricks delivered to a construction site and calling the building finished. Some bricks are useful. Some are in the wrong place. Some create future maintenance problems. The economic question is not how much code appeared. It is whether reliable software reached users faster, with fewer defects and lower lifetime cost.
Google’s 2025 DORA research puts this in a useful frame. The report describes AI as an amplifier: it magnifies the strengths of high-performing organisations and the dysfunctions of struggling ones. That is exactly what I see in enterprise teams. Strong teams use AI to accelerate design exploration, test generation, documentation, refactoring, and operational analysis. Weak teams use AI to produce more work-in-progress that nobody has time to review.
Stack Overflow’s 2025 Developer Survey also shows why raw productivity metrics can be misleading. It found that 84% of developers use or plan to use AI tools, up from 76% in 2024, but 46% said they do not trust the accuracy of AI-tool output. Forty-five percent said debugging AI-generated code is time-consuming. Among developers using AI agents at work, 69% agreed they had experienced a productivity increase.
That is the tension. AI can increase individual productivity, but the enterprise still has to pay the review bill. If the ROI report counts generated code but ignores debugging, security review, architecture review, and maintenance, it is not measuring ROI. It is measuring inventory.
The Five-Cost Framework
CIOs need a more honest scorecard. I would start with five costs that most AI business cases undercount.
First, review debt. Every AI output that affects a customer, system of record, financial number, regulatory response, product design, or production code path needs some level of review. Review debt is the human effort needed to verify, correct, approve, or reject AI work.
Second, technical debt. AI can produce fast code, fast workflows, and fast integrations. It can also create brittle automation, duplicated logic, weak error handling, and unmaintainable prompts. The cost appears later, usually when the original productivity dashboard has already been circulated.
Third, exception cost. Automation changes the shape of human work. Routine cases may move faster, but exceptions often become more complex. If AI handles simple requests and humans receive only ambiguous, angry, risky, or high-value cases, the average handling time for human work may rise.
Fourth, governance overhead. Responsible AI needs inventory, risk tiering, testing, data controls, audit trails, vendor review, monitoring, and incident response. These are not blockers. They are operating costs. Pretending they do not exist only makes the business case fragile.
Fifth, integration drag. A chatbot that works in a demo is not the same as an AI capability embedded into order management, claims handling, credit operations, procurement, software delivery, or customer support. Integration is where many AI pilots lose their economics.
The hard truth is that AI value is often real, but narrower than the first dashboard suggests. A credible ROI model subtracts these costs before claiming victory.
Measure Business Outcomes, Not AI Activity
The right unit of measurement is the business workflow.
For customer service, measure containment quality, repeat contact rate, complaint escalation, customer satisfaction, refund leakage, regulatory complaints, and cost per resolved case. For software engineering, measure lead time, change failure rate, escaped defects, security findings, maintainability, and developer satisfaction. For finance operations, measure close-cycle time, reconciliation breaks, audit adjustments, exception backlog, and control failures.
Notice what is missing: number of AI interactions.
AI usage can be a useful adoption signal, but it is not the outcome. A business can have heavy AI usage and no improvement in margin, risk, speed, or customer experience. It can also have modest usage concentrated in a few high-value workflows and produce a better return.
McKinsey’s 2025 global AI survey reinforces this point. It found that high performers were nearly three times as likely as others to have fundamentally redesigned individual workflows when deploying AI. High performers were also more likely to have senior leaders who demonstrated ownership and commitment, and to have defined processes for determining when model outputs need human validation.
That is the pattern: workflow redesign, leadership ownership, human validation, and disciplined scaling. Not more prompts.
A Practical AI ROI Scorecard
Here is the scorecard I would ask every enterprise AI programme to use before it claims impact.
- Baseline outcome: What was the business metric before AI?
- Gross productivity gain: What work became faster or cheaper?
- Review and correction cost: How much human verification did the AI create?
- Exception impact: Did the remaining human work become easier or harder?
- Quality impact: Did defects, complaints, reversals, or rework change?
- Risk impact: Did data exposure, policy violations, audit findings, or security issues change?
- Adoption durability: Are people still using the capability after the novelty fades?
- Integration cost: What did it cost to embed the capability into real workflows?
- Net P&L impact: Did revenue, cost, working capital, risk loss, or service capacity improve after all costs?
- Strategic option value: Did the organisation gain a reusable capability it can apply elsewhere?
This is more demanding than a productivity calculator. It forces the business to prove that AI changed an operating result, not merely a task duration.
The CFO Test
A simple way to test an AI ROI report is to ask whether the CFO would recognise the value in the financial statements or risk profile.
If the report says “10,000 hours saved”, the next question is: where did those hours go? Did headcount growth slow? Did revenue capacity increase? Did service levels improve without extra staff? Did risk losses fall? Did project throughput rise? Did customer churn decline?
If the answer is “people are doing more strategic work”, ask for evidence. Which work? Which metric moved? Which backlog fell? Which decision improved?
I am not arguing that every AI benefit must show up immediately as cost take-out. Some of the best AI investments create option value: better data foundations, faster experimentation, stronger decision support, reusable agent patterns, and new product capabilities. But even option value needs a thesis, milestones, and leading indicators. Otherwise, it becomes a polite name for hope.
What Good Looks Like
The organisations getting AI ROI right tend to behave differently.
They choose fewer use cases and go deeper. They fund data and process work, not just licences. They put business owners, engineers, risk teams, and frontline users in the same delivery loop. They measure before-and-after workflow outcomes. They treat governance as part of the operating model. They stop pilots that do not move a real metric.
They also accept that productivity is not evenly distributed. Some roles get immediate gains. Some teams inherit more review work. Some processes need redesign before AI can help. Some use cases should not be automated at all.
The bottom line is this: AI ROI is not found in the tool. It is found in the redesigned work around the tool.
In 2026, executives should be suspicious of any AI report that celebrates activity without subtracting the cost of trust. Hours saved are interesting. Code generated is interesting. Prompts used are interesting. But enterprise impact begins only when the organisation can show that AI improved an outcome after review debt, technical debt, exception cost, governance overhead, and integration drag have been counted.
That is the difference between AI theatre and AI value.