AI ROI

Why it matters

Most AI investment cases fail not because the model is weak but because the math is scoped wrong. Two patterns dominate. The first counts a pilot's demo accuracy as if it were production value — ignoring that the long tail the demo never hit is where the cost and the risk concentrate. The second books the license or platform fee as 'the cost,' then discovers the real spend is the integration, the data cleanup, the humans reviewing outputs, and an inference bill that scales with usage instead of with seats. Both produce a number that looks good on a slide and collapses in production.

• Unlike a CRM seat, an AI system's marginal cost rises with every task it does. Doubling adoption roughly doubles the inference and review cost, so ROI has to be modeled per-transaction, not per-license — and a use case that pencils out at pilot volume can go underwater at scale.
• Value is a distribution, not a point. A model that is right most of the time delivers very different ROI depending on what the wrong cases cost: a mis-tagged email is a shrug; a mis-quoted contract or a wrong dosage instruction is a liability. The expected cost of the error rate belongs in the model, not in a footnote.
• The honest denominator includes the unglamorous line items: data pipelines, evaluation harnesses, observability, prompt and model maintenance as versions deprecate, and the human-in-the-loop time that almost never appears in the original business case.

How to actually calculate it

Tie the AI to one operational metric it provably moves, then trace that metric to dollars. The discipline is isolating the AI's contribution — ideally with a holdout group or a clean before/after baseline — so you are not crediting it for trends it did not cause. A worked, illustrative example with placeholder numbers: suppose inbound leads convert at 10% today, and faster AI-driven first-response lifts that to 13%. On 1,000 leads worth $2,000 of margin each, that is 30 extra wins, or $60,000 in value. If the system costs $20,000 a year all-in to run, ROI is ($60,000 − $20,000) ÷ $20,000 = 200%. Every input there is one you can audit — the conversion lift against a holdout, the margin per win from finance, the run cost from your own bill. Swap in your real numbers and the structure holds.

• Value created = (metric lift) × (volume) × (dollar value per unit), net of cannibalization and any quality regression the AI introduces elsewhere.
• Total cost of ownership = build (one-time) + run (inference, infra, licenses) + govern (eval, monitoring, review labor) + the expected cost of the residual error rate. That last term is the one most models drop.
• Discount for attribution: only count the lift you can defend against a baseline. 'We launched it and the number went up' is a coincidence claim, not an ROI claim, until a holdout or a seasonal control rules out the alternatives.
• Track it as a live number, not a one-time slide. Inference prices fall, models get deprecated, and the task mix drifts toward the hard cases as the easy ones get automated first — so both the numerator and denominator move after launch.

Where it fits

AI ROI is the gate at the front of a project and the scoreboard for its life. Use it before committing capital to rank competing use cases — the best first AI project is usually the one with a clear metric, high volume, and a cheap cost of being wrong, because that combination lets you learn fast without betting the business on the model. Use it after launch to decide whether to scale, fix, or kill. The catch is that both the gate and the scoreboard need the same thing most organizations skip: clean, connected data and a defined baseline. Without those, the value side of the equation is a guess dressed as a forecast. This is why serious AI ROI work starts with data readiness, not with the model — and why an outcome-tied arrangement, where the vendor's fee moves with the metric, is useful beyond its commercial terms: it forces both sides to agree on what 'return' means, and how it will be measured, before anyone writes a prompt.