AI ROI Models

Why traditional ROI models fail for AI

Traditional ROI logic works best when leaders can define a bounded investment, estimate a finite operating benefit, and compare the two with reasonable confidence. Enterprise AI makes that harder. Its cost base is layered rather than singular. Its benefits depend on workflow redesign, data quality, adoption, governance, and organisational learning. McKinsey's 2025 State of AI shows widespread AI use, but only a small share of companies reporting clear financial return. IBM's executive research points to the same measurement challenge.

This is why many AI business cases become either too narrow or too optimistic. Narrow models miss strategic and capability effects that matter. Overconfident models treat early productivity signals as realised economic proof. Both lead to weak capital allocation.

The ROI timeline reality

AI rarely behaves like a short-payback software investment. Deloitte's 2025 evidence suggests satisfactory returns often take two to four years. Only about 6% of organisations see payback in under a year, and even among high performers the share reaching payback inside twelve months is still limited. This is one reason McKinsey recommends thinking in portfolio terms: 70% of investment directed toward more predictable ROI, 20% toward strategic innovation, and 10% toward breakthrough exploration.

Benefits also tend to emerge in cycles. Optimisation benefits often appear first in the first twelve months. Reinvention benefits tend to emerge over one to two years as workflows and roles change. Disruption benefits, where they exist at all, usually arrive after two years and depend on stronger operating capability than most organisations initially expect.

A multi-dimensional ROI framework

Leading and lagging indicators

Leading indicators show whether the mechanism of value is beginning to work. Lagging indicators show whether the business outcome has actually appeared. Good governance uses leading indicators to manage delivery and lagging indicators to govern scale.

The common mistake is to collapse them. Time saved, user adoption, and model quality may all matter, but none of them is equivalent to realised margin improvement, lower loss rates, or revenue uplift. That distinction becomes even more important when AI payback runs longer than many boards expect.

High performer profile

McKinsey's research suggests AI high performers are around three times more likely to report strong senior leadership engagement and around five times more likely to invest more than 20% of digital budgets in AI. That does not mean high spend guarantees performance. It means high performers tend to commit seriously and govern seriously.

Deloitte adds an important correction. Organisations that take a purely technology-focused approach are materially more likely to underperform expectations. The strongest cases redesign workflows rather than simply adding AI tools to existing work. That matters because only a minority of organisations have meaningfully redesigned workflows around AI. The more durable return model is what Deloitte describes as Humans x Machines convergence, not a simple Humans plus Machines layering of new tools on top of old processes.

High-performer habits:

1. Define the primary return dimension before deployment.
2. Capture baseline measures before adoption scales.
3. Assign a named owner for realised value, not only implementation.
4. Review AI as a portfolio using predictable ROI, strategic innovation, and breakthrough bets.
5. Raise the proof threshold as spend, dependency, and reputational exposure increase.

The CFO's AI ROI checklist

ROI by function

• Software engineering: the strongest evidence base is currently in code generation, test automation, and incident resolution, where workflow metrics and adoption data are easier to capture.
• Customer operations: assisted resolution, knowledge surfacing, and next-best-action models have clearer links to handling time, deflection, and service quality.
• Supply chain and procurement: reported cost savings in the high twenties are among the more credible functional gains in the broader research base.
• Finance and accounting: invoicing, forecasting support, and expense auditing often show faster-cycle benefits, including materially shorter close processes.
• Sales and marketing: personalised messaging, proposal assembly, and pipeline hygiene can create value, but attribution is often weaker and must be handled carefully.

Boardroom implications

At executive level, AI ROI should be treated as a portfolio problem rather than a sequence of isolated use cases. The question is not simply whether one initiative can show positive return. It is whether the organisation has a disciplined model for comparing multiple forms of return with different time horizons, proof standards, and governance needs.