AI at work
Who Checked the Supplier’s Installation Charge?
Counting every AI draft as completed work hides checking and correction time. Count a supplier comparison as finished when someone can safely use its conclusions.

Suppose a purchasing team uses AI to summarise quotes. One draft says a supplier’s price includes installation, even though the quote lists it as an optional extra. If the weekly tally counts every draft sent for approval as complete, the work of catching that error disappears from the numbers. Approve the wrong quote and the installation bill could arrive later.
Some AI tests reward a guess more than an admission of uncertainty
In a 2026 Nature paper, Adam Tauman Kalai and colleagues examined ten evaluations drawn from four prominent AI leaderboards. Only one offered even minimal credit for acknowledging uncertainty. On a test where a wrong answer and a blank both get zero, guessing can earn a point while leaving the answer blank cannot.
Put the evidence in the completion rule
For a quote comparison, decide which terms the buyer will rely on, such as the total price and what it includes. Ask the preparer to record the quote version and the line that establishes each term. On the team scorecard, move the comparison from submitted to complete only after the approver confirms those lines support the comparison.
Keep returned work in the weekly figures. Record whether a claim was wrong or the quote did not settle it, and include the time spent checking or correcting the draft. A missing installation term may require asking the supplier. A misstated term can be corrected from the quote. Record the supplier follow-up separately from the correction so the team can see where its time went.
Take a sample of drafts from a recent reporting period and apply the proposed rule before changing targets. Compare the number recorded as finished with the number a buyer could act on at the time. Use the gap and the checking hours to set a target the team can meet with review time included.
Sources and further reading
Sources checked 8 November 2026. Edited and reviewed by AI Boss.