Measuring return on AI investment

Why AI return is hard to measure

AI value spreads across efficiency, decision quality, error reduction, and risk prevention at the same time, which single-metric models do not capture. It also compounds slowly. Organizations now expect satisfactory return in two to four years, against the seven to twelve month payback typical of conventional IT investment.

A large share of the value is prevention. Anomaly detection creates value by stopping events from occurring, which is harder to quantify than incremental revenue.

The investment ladder

I size spend against the problem using six rungs: use features already in your existing software, call a frontier model by API, run RAG over proprietary data, self-host a smaller model, fine-tune, and build a foundation model. Returns concentrate in the first three rungs. Custom foundation models are rarely justified outside a small number of organizations.

Most enterprises are still in buy-and-integrate territory, and the discipline is staying there until a specific requirement justifies moving up.

What I insist on

Separate hard return from soft return and track both, while recognizing that only hard return appears in the P&L. Budget monitoring, retraining, governance, and human review as ongoing cost. Unbudgeted run cost is one of the most common reasons finance leaders cannot demonstrate AI return.