Replace AI Seat Counts With Cost per Accepted Task

Primary source: https://openai.com/index/a-scorecard-for-the-ai-age/.

OpenAI's July 17 scorecard argues that AI value should be measured through useful work, successful-task cost, dependability, and return on compute—not adoption alone.

Use one workflow ledger:

workflow,attempts,accepted,retries,review_hours,model_cost,platform_cost
invoice-match,120,91,24,8.5,180,70

complete_cost = model + platform + review_hours * loaded_rate
cost_per_accepted = complete_cost / accepted
acceptance_rate = accepted / attempts

Report attempted, accepted, rejected, escalated, and silently abandoned work. A cheaper token price can lose if it causes more retries or review.

Gate	Example policy
quality	zero irreversible false actions
evidence	95% of runs traceable
economics	downside case below baseline
ownership	named workflow and incident owner

The numbers above are illustrative. Compare the same outcome and quality threshold against the existing process, then run sensitivity on review time and acceptance rate.

This scorecard does not prove causality or capture every strategic benefit. Which denominator would change your current AI ROI story most: accepted work, retries, or reviewer time?