| In short: To measure the ROI of AI spend, compare the full cost of each AI use case (models, infrastructure, data work, licences and people) with a measured change in a business metric against a baseline you recorded before launch. Track both as a unit metric, such as cost and value per completed task, and review them monthly so you can scale, fix or stop each use case. |
Most organizations can now tell you what they spend on AI. Far fewer can tell you what they get back. In the FinOps Foundation’s State of FinOps 2026 survey of 1,192 practitioners, 98% said they now manage AI spend, up from 31% two years earlier. Yet KPMG’s Global AI Pulse for Q3 2026, a survey of 2,131 senior leaders, found that only 12% consistently assess AI value against costs across the organization.
That gap matters because the money is real and the scrutiny is growing. McKinsey’s 2026 State of AI survey found that 37% of respondents see at least some EBIT impact from AI, essentially unchanged from 2025, and only about 6% qualify as “AI high performers” attributing 5% or more of EBIT to AI.
This article gives you a practical way to measure AI ROI per use case: what to count as cost, how to measure value, a worked example and a monthly review that finance and engineering can both trust.
| IN THIS ARTICLE
1. Why is AI ROI so hard to measure? 2. What should you count as the cost of AI? 3. How do you measure the value of an AI use case? 4. A worked example: calculating ROI for one use case 5. How to build a monthly AI value review 6. How Sela helps 7. Frequently asked questions 8. Key takeaways |
1. Why is AI ROI so hard to measure?
Three problems come up in almost every organization.
Costs are scattered. AI spend sits across cloud bills, model API invoices, SaaS licences such as copilots, data platform costs and staff time. No single report shows the full cost of one use case.
Value has no baseline. Teams launch an AI feature and then try to measure its impact, but nobody recorded how long the task took, how much it cost or how often it went wrong before. Without a baseline, any value claim is a guess.
Productivity doesn’t automatically become profit. In McKinsey’s survey, 80% of respondents said AI had improved their individual productivity, but that hasn’t yet shown up as broad financial impact. Saving an employee an hour a week is only a financial return if that hour turns into more output, lower cost or avoided hiring.
The organizations that get this right measure on purpose. KPMG found that among organizations reporting established AI returns, 48% consistently assess value against costs, about four times the overall rate. McKinsey’s high performers are twice as likely as others to have defined processes for measuring the impact of their AI initiatives.
2. What should you count as the cost of AI?
Measure the total cost of each use case, not just the model bill. A complete view usually has five parts:
| Cost category | What it includes |
| Model usage | Tokens or API calls, provisioned throughput, fine-tuning and evaluation runs |
| Infrastructure | GPUs or compute, vector databases, storage, networking, logging and monitoring |
| Data work | Pipelines, data preparation, labelling and the share of the data platform the use case relies on |
| Licences | Seats for copilots and AI-enabled SaaS tools, counted for the people who actually use them |
| People | Build and maintenance time, prompt and evaluation work, human review of AI output, training |
Two practices make this workable. First, tag cloud and AI resources by use case and team from day one, so costs can be allocated without manual spreadsheets. Second, separate one-time build costs from run costs. A use case can be a good investment with a long payback and still have run costs that grow too fast as usage scales.
Cost visibility on its own goes a long way. KPMG found that 77% of organizations reporting established AI returns use cost-monitoring dashboards, compared with 43% of those still experimenting.
3. How do you measure the value of an AI use case?
Value should be a change in a business metric that someone outside the AI team already cares about. Most use cases fit one of four types:
● Efficiency: time or cost per task, such as handling time per support case or hours per report.
● Revenue: conversion rate, average deal size, retention or new product revenue.
● Quality and risk: error rates, rework, compliance exceptions or fraud losses avoided.
● Capacity: more work done with the same team, such as cases closed per agent or releases per month.
For each use case, follow four steps:
1. Pick one primary metric and at most two secondary ones. More than that and nobody reads the report.
2. Record a baseline for at least a few weeks before launch, or keep a control group that doesn’t use the AI tool.
3. Agree how value converts to money with finance before launch. For example, decide whether saved hours count only when they reduce overtime or contractor spend.
4. Measure adoption alongside the outcome. Low usage is the most common reason a good use case shows poor ROI.
4. A worked example: calculating ROI for one use case
The numbers below are illustrative, not benchmarks. They show the method.
A support team adds an AI assistant that drafts replies for agents. Before launch, the team records an average handling time of 12 minutes across 20,000 cases a month.
Monthly cost. Model usage $3,000, infrastructure and monitoring $1,500, licences $2,000, and maintenance and evaluation time $3,500. Total run cost: $10,000 a month. The build cost was $60,000.
Monthly value. After three months, handling time drops to 9 minutes. That saves 3 minutes on 20,000 cases, or 1,000 hours a month. Finance agrees to count saved hours at a loaded cost of $40 an hour, but only 60% of them, because that is the share that reduced contractor and overtime spend. Value: 1,000 × $40 × 60% = $24,000 a month.
Result. Net monthly benefit is $24,000 − $10,000 = $14,000, a monthly run ROI of 140%. The $60,000 build cost pays back in a little over four months. The unit view is just as useful: about $0.50 of AI cost per case for about $1.20 of measured value per case.
The unit view is what you track as usage grows. If the cost per case rises because prompts get longer or a larger model is used, you’ll see it long before the monthly total looks alarming.
5. How to build a monthly AI value review
KPMG found that 61% of organizations review AI costs during project approval and 59% monitor them once systems are running. The step most organizations miss is bringing cost and value together in one regular review.
Keep it simple. Once a month, for each AI use case, review:
● Cost: total run cost and cost per unit (task, case, user or conversation), compared with last month.
● Value: the primary metric against baseline, converted to money using the agreed rule.
● Adoption: active users or share of eligible tasks handled with AI.
● Decision: scale, optimize (for example, a smaller model, caching or better prompts) or stop.
Invite a business owner, a finance or FinOps lead and the technical owner. Use cases with no measurable value after two or three review cycles should be fixed or shut down, so budget moves to the ones that work.
6. How Sela helps
Sela’s FinOps services include cost analysis that links cloud spending to performance and results, establishes the cost-optimization metrics your team needs and helps you assess business value based on the actual costs your cloud resources incur. Ongoing reports support planning and forecasting, and FinOps as a Service gives your operations team continuous access to spending data.
If you are still choosing which AI use cases to fund, Sela’s AI Kickstart Program helps identify high-ROI use cases and produces an AI readiness report with ROI estimates for each opportunity. For the data and platform work behind those use cases, see Sela’s Data & AI services.
Additionally, any AI project that Sela plans and delivers contains a built-in ROI estimation and tracking component.
7. Frequently asked questions
| How do you calculate the ROI of AI?
Subtract the full cost of the use case from the measured value it creates, then divide by the cost. Measure value as a change in a business metric against a baseline, and include model, infrastructure, data, licence and people costs. How long does it take for AI investments to show ROI? It depends on the use case. Efficiency use cases with high volume can show measurable results within a few months, while revenue and transformation use cases usually take longer. Set an expected payback period before launch and review against it. Should productivity gains count as AI ROI? Only when they turn into a financial outcome, such as lower overtime or contractor spend, more output with the same team, or avoided hiring. Agree on that conversion rule with finance before launch. What is the difference between AI cost management and AI value measurement? Cost management tracks and controls what you spend on AI. Value measurement links that spend to business outcomes, so you can decide which use cases to scale or stop. Who should own AI ROI measurement? A business owner should own each use case’s outcome, with finance or FinOps owning the cost and value method and the technical team owning usage and cost data. |
8. Key takeaways
| ● Almost every organization manages AI spend now, but only 12% consistently assess AI value against costs, according to KPMG’s Q3 2026 AI Pulse.
● Count the full cost of each use case: models, infrastructure, data work, licences and people. ● Record a baseline before launch and agree with finance how value converts to money. ● Track cost and value per unit and review them monthly so you can scale, optimize or stop each use case. |
Want a clear view of what your AI spend returns? Talk to Sela about FinOps services.
| For general information only. Statistics are from the FinOps Foundation’s State of FinOps 2026 survey (press release, February 19, 2026), KPMG’s Global AI Pulse Q3 2026 and McKinsey’s The State of AI in 2026. The worked example uses illustrative figures. Results depend on your use cases, data and costs. |