Guide
Is your AI agent saving time? A small-business pilot worksheet
Count setup, checking, repairs and software costs before calling an AI pilot a success. Includes a blank ledger and a worked example.
By Agent Report Editorial
Editorial review · · 4 min read
Written with Codex assistance. Sources checked October 10, 2026. Example numbers are invented to explain the calculation. Worksheet arithmetic is checked; no business pilot or product comparison was run.
Measure the human work required to finish an acceptable task. Include the time spent checking the agent, correcting it and finishing the cases it could not handle. Then put the software bill beside that result.
Bluevine's 2026 small-business survey asks owners about returns from AI. It covers broad AI use, not this kind of agent trial. Its unweighted survey of 942 US owners at firms with 2–249 employees is useful context for the question, not evidence that your workflow will save money.
The method below is an original proposal. Two weeks can expose a recurring problem or justify a longer trial. It cannot establish a reliable annual return for a seasonal business.
Choose one task and define acceptable work
Pick something you can count: preparing a callback record, drafting a standard customer reply, or assembling an invoice draft. Define “accepted” before the trial. For an invoice draft, acceptance might mean the correct customer, approved job lines and no duplicate charge, all confirmed by a reviewer.
Keep harder cases in the log. Excluding every exception would measure the easy portion of your business while hiding the work that still needs a person.
Download the pilot ledger and cost worksheet. Use anonymous case IDs and one currency throughout. Blank cells mean unmeasured, not zero.
Observe the existing process first
During the first week, record comparable tasks completed through the current process. Log active human minutes from starting work through final acceptance. Include correction and follow-up. Record elapsed time separately if waiting matters to the customer.
During the second week, record the agent-assisted process against the same acceptance rule. Track human minutes in distinct, non-overlapping categories: preparing the case, checking, repairing, and manually finishing a failed attempt. Log setup and training separately in the cost worksheet.
A before-and-after comparison can be distorted by easier work or a quieter week. Compare like task categories, preserve the individual rows and say when the mix differs. If you can allocate comparable cases between the two processes during the same period, that can reduce calendar effects, although staff learning and selection can still influence the result.
Calculate net human time first
For an equal number and comparable mix of completed cases:
net human minutes freed = baseline human minutes
- assisted human minutes
Assisted human minutes include every attempted case and any manual fallback needed to complete the work. Do not compare 50 accepted baseline cases with 40 accepted agent cases and ten unfinished ones as if the output were equal. Either complete and count the fallback, or report the unfinished work separately without claiming equivalent time savings.
Report acceptance counts, errors and time together. A shorter queue is not enough if customers are receiving incorrect answers.
Keep capacity value separate from cash
Here is an invented example using USD. Capacity value estimates the value of freed time; it is not cash saved:
| Item | Illustrative amount |
|---|---|
| Comparable accepted cases in each process | 50 |
| Baseline human time | 400 minutes |
| Assisted human time, including review and repairs | 250 minutes |
| Human time freed | 150 minutes, or 2.5 hours |
| Chosen value of that time | 30 USD per hour |
| Capacity value, not cash savings | 75 USD |
| Software and usage cost for the same period | 40 USD |
| Capacity value after software cost, not cash savings | 35 USD |
If wages and paid hours stay the same, the 75 is an estimate of available capacity. It is not 75 leaving your bank account less often. Cash savings need evidence of an expense that actually fell. Extra revenue needs evidence of additional completed business; do not count every captured lead as a sale.
Suppose setup also took four hours. At the same chosen rate, that is another 120 of time value. Including setup makes the example's first-period capacity balance negative 85: 75 minus 40 minus 120. This still is not a cash-profit calculation. Neither figure predicts the next month.
Make the decision before expanding the job
Write one of four outcomes: continue, narrow, stop, or gather more evidence. Include the case count, accepted outcomes, human minutes, charges and unresolved errors.
Continue when the result meets your prewritten requirements and the work of maintaining it fits your team. Narrow the scope when one category works and another creates repairs. Stop when a core requirement fails. If the sample is too small or the work changed, say so and extend the observation instead of presenting a percentage return.
Keep the ledger for the next review. A process that works this month still needs a named owner and a maintenance routine.
Apply this worksheet to a real shortlist
Use our receptionist vendor comparison, shared-inbox comparison or accounting-assistant comparison to choose a trial. Copy the actual quote into the cost sheet, including usage, required seats and integrations. Keep the reviewer's time and vendor-support time in the calculation. Public reviews can suggest costs to investigate; they cannot supply your measured savings.