Strategy

Problem 4: Nobody can prove the ROI

27 October 2026 · 7 min read · Hidden Node

If you've ever tried to get an AI project past a finance team, you know the conversation. "It'll save time." How much? "Well, it depends." On what? "On… adoption?" Silence.

You're in good company. In IBM's 2025 CEO Study, chief executives said only 25% of their AI initiatives had delivered the expected return over the past few years. Half of them also admitted that rapid investment had left them with disconnected technology across the organization.

Why AI value is slippery

Gartner's Rita Sallam described the core tension well when the firm warned about abandoned projects: many productivity gains from generative AI are hard to turn into direct financial returns, and many CFOs have not been comfortable with investing today for indirect value in the future.

That's the trap. "Each analyst saves 40 minutes a day" is real, but it only becomes money if those 40 minutes go somewhere: fewer overtime hours, a backlog that finally clears, growth absorbed without hiring. If the time just dissolves into slightly longer lunches, the ROI was never there.

The two mistakes we see in business cases

Counting hours as rupees. Time saved is not cash saved unless something changes: headcount plans, overtime, outsourcing spend or throughput. Be specific about which.

Forgetting half the costs. Licences and model usage are the visible part. Integration, testing, change management, monitoring and the people who review AI output are usually bigger, and they continue after launch.

A business case a CFO can live with

1. Measure the baseline before you build. Volume per month, minutes per case, error rate, turnaround time. If you skip this, you'll never be able to prove anything later.
2. Estimate conservatively. Assume only part of the work is automated, people still review outputs for the first months, and adoption ramps up gradually.
3. Convert only what's real. Say exactly how saved time becomes money, or call it a quality benefit and stop pretending.
4. Include the run costs. Model usage grows with volume. Budget for monitoring and maintenance from the start.
5. Stress-test it. What if volume halves? What if accuracy is lower than hoped? If the case only works in the best scenario, pick another use case.
6. Name an owner. One person accountable for the metric, reporting monthly against the baseline.

Don't measure only cost

McKinsey found that 80% of companies set efficiency as an AI objective, but the ones getting the most value usually set growth or innovation goals too. Faster loan decisions win customers. Quicker claim payouts reduce complaints and churn. Those are harder to measure than hours, but often larger.

Our view: an AI project that can't survive a conservative business case isn't a bad project. It's the wrong first project. Start with the one that can, prove it, and let the results fund the next.

Sources

• IBM, 2025 CEO Study
• Gartner on GenAI projects and CFO investment criteria, via Intelligent CIO
• McKinsey & Company, The state of AI in 2025