NEWS & ANALYSIS · AI, MACHINES & THE ECONOMY
OCTOBER 13, 2026 · THE WEEKDAY BRIEF
Good morning. A cheap model response can start an expensive chain of work. Someone checks it, asks for a correction, waits for another attempt, then fixes the result. The invoice records the machine’s activity. The business still needs a finished job.
For an AI agent, the useful unit is a correctly completed business task: a reconciled invoice, an accurate account update, or a support request resolved to a defined standard. Counting responses can reward a system for doing more work to reach the same result.
01 / THE BIG IDEA
Put the whole workflow in the numerator
Start with a simple measure: total workflow cost divided by correctly completed tasks. Include model calls, tool fees, human review, retries, error correction and an allocated share of setup and maintenance. Count a task once, even if the agent makes several attempts.
Keep spending on failed and abandoned tasks in the numerator. Excluding those costs makes an unreliable system look efficient. Define acceptance before the trial: what must be correct, who verifies it, and how long the result must remain valid before it counts as done?
02 / FOLLOW THE MONEY
A hypothetical comparison
The following numbers are invented to illustrate the calculation. They are not agent test results or vendor prices.
Suppose two workflows each receive 100 comparable tasks. Both ultimately deliver 90 accepted results. For this example, human time costs $60 an hour, or $1 a minute, including employment overhead.
Workflow A: initial machine and tool use costs $10; retry calls cost another $5. People spend 100 minutes reviewing and 60 minutes correcting errors. Add $20 for allocated setup and maintenance. Total: $10 + $5 + $100 + $60 + $20 = $195. Divide by 90 accepted tasks: $2.17 each, rounded.
Workflow B: initial machine and tool use costs $40; retry calls cost $10. People spend 40 minutes reviewing and 20 minutes correcting errors. Add the same $20 allocation. Total: $40 + $10 + $40 + $20 + $20 = $130. Divide by 90: $1.44 each, rounded.
In this scenario, the higher machine bill buys enough reduction in human work to lower total cost by $65. All effort on the ten unsuccessful tasks is included. Review and correction minutes are separate, so labor is counted once.
Change those assumptions and the ranking can reverse. A more expensive agent earns its place only if the full workflow performs better on comparable work.
03 / THE PRODUCTIVITY RECEIPT
A measurement checklist
Define done. Write the acceptance rule and deadline for one narrow task. Record failures, escalations and cases reopened later.
Establish the baseline. Measure the current process on a comparable mix of easy and difficult cases, using the same quality standard.
Log every attempt. Tie initial calls, retries and tool charges to the original task. Include work abandoned before completion.
Time the people. Separate review, correction and escalation. Use a consistent labor rate and disclose setup-cost allocation.
Read cost alongside quality and speed. Track acceptance rate, elapsed time and serious errors. A low average cost cannot excuse an unacceptable failure.
Finally, distinguish released capacity from cash savings. Ten hours returned to a team might reduce a backlog or support more customers without reducing payroll. Report that benefit honestly, then check whether the business actually uses the capacity.
OVER TO YOU
Which task would you measure first—and what would count as correctly completed?
Machines and Money Team