Cost per accepted output: the AI metric that counts review time

Somebody finally named the number that AI vendors would rather you not calculate. At Reuters Momentum AI Austin, KPMG's Swami Chandrasekaran, Head of AI and Data Labs, proposed cost per accepted output — what a business actually spends to produce one piece of usable AI work, including the human who reviewed it, the systems that catch the errors, and the fallback process for when the model fails. It is a better metric than anything on your invoice, and it is usually the one that kills a bad automation project before you scale it.

What actually happened

Per Reuters Events' day-one summary, the framing at the event was that enterprise AI has entered its "prove-it" era — executives moving past pilots to ask whether any of it delivers measurable value. Chandrasekaran's metric looks past the cost of the technology itself to capture what businesses really spend producing usable output: human review, error-catching systems, and the fallback path when AI fails.

The same session put real numbers next to it. FedEx cited more than 200 data and AI use cases contributing to over $3 billion in cost reductions. Mars reported roughly 20% top-line growth upside in measured Amazon digital-commerce activity. Indeed said AI-generated recommendations account for around 70% of matches. Note what those three have in common: each is a specific, instrumented workflow with a number attached, not a company-wide "AI transformation." That is the actual lesson of the event, and it is free.

Why cost per accepted output matters for your business

The denominator is "accepted," and that word is doing all the work. A model that produces ten drafts where seven get thrown out is not 10 units of output at your token price. It is 3 units at more than triple the token price, plus the salary cost of the person who read all ten to find the three. Track acceptance rate per workflow from day one. If you cannot say what fraction of your AI output ships without human rework, you do not know what the automation costs.

Review time is the cost line nobody budgets. We have written before about measuring cost per task instead of per token. Cost per accepted output goes one step further and puts a human hourly rate in the formula. Run it honestly: token spend, plus reviewer minutes times loaded hourly cost, plus the cost of the guardrail and eval tooling, divided by outputs that shipped. Compare that against the fully loaded cost of the manual process. Sometimes the answer is that the automation wins by 5x. Sometimes it is that you moved the work from "doing" to "checking" and made it slower.

A low acceptance rate is an engineering signal, not a reason to quit. Acceptance rates climb when you narrow the task, feed the model your actual data instead of a generic prompt, and add a deterministic check before a human ever looks at it. A validator that rejects malformed output automatically is cheaper than a person who catches it on read three. The biggest wins we see come from shrinking scope until the output is verifiable by code.

Pick one workflow with a countable unit. FedEx, Mars, and Indeed all reported against a measurable thing: a cost line, a revenue line, a match. Do the same. A resolved support ticket, a published product description, a reconciled invoice, a qualified lead. Instrument it, get a baseline before you automate, then measure the same number after. Without the before, there is no after — just a vendor bill and a feeling.

Sources: Reuters Events via GlobeNewswire, Momentum AI Austin 2026.