American Banker argues that banks leave the cost of AI mistakes out of return calculations. Errors, false positives and control failures create remediation and investigation expenses. A fraud system saving $5 million but adding $1 million in investigation costs has not created $5 million. Net value should subtract full lifecycle cost and expected costs of errors. Lifecycle cost includes data, compute, integration staff and shared team costs, not just token spend. Model choice is framed as an economic decision, not only a technical one. Costs and benefits should also be measured over matching time periods. Payback period and net present value suit longer projects; simple return figures suit short ones. Figures come from an American Banker Market Intelligence risk survey and EY's AI Pulse Survey.
What changed
AI return estimates often count productivity gains without the cost of errors.
What it unlocks
Valuing an AI project net of error costs, shared costs and timing.
- 81% call AI model risk at least moderate
- 60% see deepfake fraud loss likely
- 82% concerned about token costs
- 64% monitor token usage
Sources