An AI project can show 300% return on investment in a business case and just 33% once finance accounts for the full cost.
Same project. Same business benefit. Different cost calculation.
300%
Consider an illustrative AI initiative delivering $400,000 in measurable benefits against $100,000 in model and token charges. That produces a 300% ROI.
33%
Now include another $200,000 for infrastructure, software, implementation, governance, and supporting labor. Total investment reaches $300,000, reducing ROI to approximately 33%.
This example exposes an increasingly important issue for enterprise finance. AI business cases can look convincing while leaving substantial costs outside the calculation.
EY puts a number on the problem
In its September 2026 report, Agentic AI ROI: Can AI pay for itself?, EY estimates that tokens account for roughly one-third of an AI agent’s total enterprise cost.
The remaining costs accumulate across software subscriptions, infrastructure, governance, organizational change, expected failures, and regulatory requirements.
Under EY’s model, full enterprise cost can reach roughly three times the token invoice.
This is an estimate based on EY’s assumptions, rather than a fixed multiplier for every deployment. Nevertheless, it exposes a serious weakness in how organizations assess AI investment.
The cost that is easiest to measure may represent only a fraction of the cost that matters.
Productivity gains need financial evidence
Measuring returns creates another challenge.
An employee saving five hours each week through AI might represent a substantial productivity improvement. But those hours only translate into a financial return when the organization captures their economic value.
That could mean processing more transactions with existing resources, avoiding additional hiring, increasing revenue, reducing errors, or shortening delivery cycles.
Simply multiplying hours saved by employee salaries risks overstating realized savings.
Deloitte’s 2026 State of AI in the Enterprise illustrates the distinction. While 66% of surveyed organizations reported productivity and efficiency gains, only 40% reported cost reductions.
Meanwhile, PwC’s 2026 AI performance research found that companies with the strongest AI capabilities generated 7.2 times greater AI-driven revenue and efficiency gains than less advanced organizations.
The opportunity extends beyond reducing labor costs. Enterprises also need to understand where AI creates new capacity, improves business performance, and supports growth.
Finance needs the complete investment picture
A credible AI ROI calculation needs costs and benefits measured against the same initiative, over the same period, using an agreed financial baseline.
That includes ongoing model consumption, cloud infrastructure, licences, data, vendors, internal labor, and the cost of managing agentic workflows.
It also requires business teams to distinguish projected benefits from realized outcomes.
For CIOs and CFOs, this changes the investment conversation. A low-cost AI initiative might deliver little measurable value. A more expensive deployment could justify further investment through stronger unit economics and demonstrable business results.
Neither conclusion is possible without reliable cost attribution and outcome measurement.
This is where Yarken’s approach to AI Economics becomes relevant.
Yarken connects AI consumption, cloud, software, vendor, and labor costs within one governed financial model. Finance and technology leaders can assign expenditure to accountable teams and initiatives, then assess it alongside usage and business outcomes.
The objective is a defensible view of AI investment that supports budgeting, forecasting, governance, and funding decisions.
The next AI investment decision
EY’s findings should prompt finance leaders to examine how they calculate returns before approving the next wave of agentic AI investment.
Complete costs, measurable outcomes, and clear ownership belong in every serious AI business case.
The next AI budget deserves the same scrutiny on cost as it receives on promised value.