Measure AI By Accepted Work, Not Activity.

OpenAI published a July 17 scorecard arguing that AI value should be measured by useful work completed, the full cost of each successful task, dependability, and whether value grows faster than cost. Its practical point is stronger than the usual model comparison: cheap tokens do not guarantee cheap outcomes when retries, review, and rework are included.

That is the right correction for remodelers, builders, designers, showrooms, suppliers, distributors, and trades. A busy team can accumulate licenses, prompts, generated reports, and impressive demonstrations while completing no more reliable work. The scorecard has to begin where the business can accept or reject an outcome.

Start with an accepted unit of work

Choose one recurring result that already has a finish line: a bid comparison ready for estimator review, a selection list with every required field, a weekly purchasing report reconciled to current purchase orders, or a project update approved for the client. Define the source records, required fields, quality bar, reviewer, and deadline.

Then count only work that meets that bar. A generated draft is activity. An accepted artifact is an outcome. This distinction prevents a team from calling an experiment successful because the model produced something quickly, even though an operator had to rebuild it.

Price the whole workflow

Model fees are only one line in the cost. Add setup, integrations, employee review time, retries, corrections, exception handling, and the cost of mistakes. Divide that total by the number of accepted outcomes. The result is a much more useful cost per completed task than cost per token or cost per seat.

A more capable model may be cheaper for a difficult scope comparison if it reaches an acceptable result with fewer attempts. A faster, lower-cost model may be the better choice for classifying clean daily records. Route models by tested workflow economics, not by a single leaderboard or vendor claim.

Track corrections and escalations separately

OpenAI proposes three operational outcomes: ready to use, needs correction, and needs escalation. A building business can make those states even more concrete. Ready to use means the output passed the named checks. Needs correction means the operator repaired it before use. Needs escalation means missing evidence, conflicting records, permissions, or business judgment prevented completion.

Do not hide corrections inside a general success rate. Record what changed, why it changed, and how long the review took. Those corrections become an eval set for the next instruction, integration, or model version. Escalations reveal where the workflow needs better source data or a deliberate human boundary—not a more confident guess.

Use a small operator scorecard

• Accepted outcomes: how many tasks met the quality bar on time.
• Correction rate: how often a person had to edit or rerun the result.
• Escalation rate: how often the workflow stopped for evidence, access, or judgment.
• Review minutes: how much human attention each accepted outcome required.
• Full cost per accepted outcome: tools, model usage, labor, retries, and rework.
• Business effect: cycle time, avoided rework, recovered capacity, or a better customer decision.

Review the scorecard by workflow and over time. If accepted volume rises while cost and correction effort fall, the system is improving. If activity rises but accepted work does not, stop expanding it and inspect the sources, instructions, tools, and quality bar.

This evidence is also the part competitors and AI summaries cannot easily replace. A real case note can show the task definition, source boundary, acceptance test, correction rate, cost, and limitation. Google says AI Overviews and AI Mode need no special AI-only schema; helpful original content, crawlable pages, accurate structured data, and visible sources remain the foundation. Google has also begun testing dedicated generative-AI visibility reports in Search Console, but impressions there still do not replace qualified business outcomes.

• Related: A Scheduled AI Task Still Needs An Exception Queue.
• Related: Coverage Evals Stop Agents From Missing What Matters.

Next step, if this note maps to a problem on your desk: Discovery Call — a 1-on-1 leverage assessment for your business ($1,500 · 90 min).