In this note
One option costs $20 in model calls. Another costs $200. I’d ask how much work each leaves for the support team before choosing either. A person opening every ticket twice can erase the difference rather quickly. Here, a route means the whole candidate setup: its model, prompts, tools, checking, and recovery, not just the model's price.
Give both options the same job and judge their answers by the same rules. Then count checking, corrections, failed work, and service costs alongside model calls. Otherwise a lower model bill may mean someone else is doing more work off the dashboard.
The monthly support example below is hypothetical. Its USD amounts and review times are assumptions, not provider prices or observed results. Use the calculation to see what you would need to measure.
Give both routes the same job
A support team wants help classifying 1,000 tickets. Each draft needs an allowed issue category, urgency, supporting facts from the ticket, and any missing information. A reviewer can correct it, reject it, or pass the problem to someone else. The automation cannot contact customers, change access, or close tickets.
Count a draft as accepted only after the reviewer checks it against the ticket and the team's category rules and records the decision. A corrected draft can count, but the correction takes time. Assume both options process all 1,000 tickets. Each finishes the month with 900 accepted drafts, 60 rejected drafts, and 40 cases passed on for help but still unresolved.
Use the same tickets, available evidence, category rules, reviewers, and month-end cutoff. Keep each ticket's identity across retries so attempts don't become extra cases. An option that skips difficult tickets or gets better source material has an unfair advantage. In a real trial, include the current manual process too; we have not established that either option beats it.
Price all the work that reaches the service
Assume option A needs 2,500 minutes of checking and correction. At $45 an hour including employment overhead, that costs $1,875. Option B needs 1,200 minutes, costing $900. Count time spent on all 1,000 tickets, including those rejected or still unresolved at month-end. The review times are teaching assumptions, not a timed study.
| Cost item | Route A / route B |
|---|---|
| Model calls, including initial calls and retries | $20 / $200 |
| Tools, compute, and storage | $80 / $80 |
| Review and correction labor at $45/hour | $1,875 for 2,500 minutes / $900 for 1,200 minutes |
| Separate technical recovery work | $120 / $60 |
| Routine support allocation | $200 / $200 |
| Assurance allocation | $100 / $100 |
| Delivery allocation | $200 / $200 |
| Total allocated service cost | $2,595 / $1,740 |
| Independently accepted drafts | 900 / 900 |
| Cost per accepted draft | About $2.88 / about $1.93 |
The $200 delivery allowance spreads $2,400 of setup over 12 months. The actual payment may happen upfront; keep that cash timing separately. Here, review means checking tickets, recovery means technical repair, and support is routine service help. Assurance covers the work of checking that the service continues to meet its requirements. These are cost categories, not reports of completed work. Don't charge the same hour to two rows. Before funding a real service, add any missing integration, evaluation, license, delay, or incident costs.
Keep the work spent on the 100 unaccepted cases in the bill. Their remaining costs are unknown, so the table shows costs through month-end, not the cost of finishing every ticket. Carry the unfinished work into the next month's record. Passing a difficult case to a person doesn't make its cost disappear.
Find how much review would change the choice
B spends $180 more on the model, but $975 less on review and $60 less on recovery. That makes it $855 cheaper for the same 900 accepted drafts. The calculation follows from our assumptions. A real comparison still has to show whether either option achieves those review times; a higher model price does not guarantee less checking.
- A's cost outside review
- $20 + $80 + $120 + $200 + $100 + $200$720
- Review rate per minute
- $45 / 60$0.75
- Review cost that matches B
- $1,740 − $720$1,020
- Corresponding review effort
- $1,020 / $0.751,360 minutes
A ties B at 1,360 review minutes. Get A down to 1,000 minutes and it costs $1,470, below B's $1,740, provided the accepted count and other costs stay the same. A clearer review screen or better source material might help. Test that revised setup on the same work. Removing a necessary check would change what the team is buying.
The accepted count can reverse the choice too. If B still costs $1,740 but only 600 drafts are accepted, each costs $2.90, above A's original $2.88. That changes one assumption to show its effect; it does not predict how failures would change labor costs. With no accepted results, report the bill and the zero count separately. You cannot calculate a cost per accepted result by dividing by zero.
Check what the lower cost leaves behind
Two options can produce 900 accepted drafts and still make different mistakes. Before testing, the support owner needs to set limits for missed urgent tickets, wrong urgency, unauthorized disclosure, unresolved work, and delay. Check important ticket types separately. A low average price cannot compensate for a dangerous miss or a data leak.
Find out when review work arrives. A manageable monthly total can still overwhelm the team just before a shift ends. Record waiting, available reviewer time, overflow, and who takes unfinished cases. Our calculation includes none of those observations or actual harmful-error results, so it supports a comparison to run, not a production choice.
For that comparison, save each option's model and prompt versions, tools, input rules, limits, and test setup. Use the same permitted cases and independently checked expected answers. Repeat runs if varying model answers could change the decision. Time the reviewer's active work separately from an open browser tab, and watch for reviewers memorizing repeated cases.
Choose what to measure next
If the model bill is the immediate concern, check whether it rose because more work arrived, requests grew, searches repeated, or prices changed. Reconcile different models and input, output, and cached-token prices separately. The worksheet includes a more detailed spend check, but the choice here still depends on the whole service cost per accepted draft.
Test a smaller context or batched requests separately from a model replacement. Keep the model and workload constant for that test, and check errors as well as cost. A lower bill caused by dropping needed evidence would not be a cheaper way to deliver the same result.
For these two options, measure reviewer time and errors on matched tickets first, then follow the remaining 100 tickets through resolution. Those results could reverse the apparent saving. Take the worksheet and that measurement request to the support owner before buying the option with the $20 model bill.