Part 6 of 12 · 2 min read
What the simulation measured
Cost breakdown by product, the sensitivity sweep, and four price structures against heavy-tailed customers.
How to read the cost-breakdown table: both columns face the same cost components; only the workload (support versus research) differs, which is why the model share differs so much.
| Vendor cost per task (assumed inputs) | Support | Research |
|---|---|---|
| Model tokens | $0.062 | $3.625 |
| Tools | $0.010 | $0.500 |
| Infrastructure | $0.005 | $0.100 |
| Credits for failures | $0.150 | $0.500 |
| Total | $0.227 | $4.725 |
| Model share | 27% | 77% |
| Per agent-resolved task | $0.41 | $5.91 |
What to notice: failure credits are the second-largest line for support (66 percent of the model-token line). But they are a small fraction of the research total. The lever worth pulling genuinely differs by product, not just by how much is spent overall.
How to read the price-structure table: all four rows are tuned to the identical 60 percent blended margin on the identical 200 simulated customers; only the pricing structure differs.
| Price list tuned to 60% blended margin | Customers at negative margin | p10 customer margin | Paid per resolved task (p10 / p50 / p90) |
|---|---|---|---|
| Per seat ($83.55 per seat) | 10.0% | 7% | $0.46 / $1.73 / $5.78 |
| Per task ($0.554) | 1.5% | 30% | $0.87 / $0.99 / $1.18 |
| Per resolved task ($0.994) | 5.5% | 17% | $0.99 / $0.99 / $0.99 |
| Platform $370 + $0.296 beyond 1,000 tasks | 4.5% | 21% | $0.61 / $1.59 / $6.97 |
What to notice: per-resolved-task is the only row where the p10/p50/p90 paid-per-resolved-task column doesn't move at all. It is the most uniform price to the customer. But it still leaves 5.5 percent of customers unprofitable. Hard tasks cost the vendor more to serve, regardless of what is charged.
How to read the sensitivity table: every row moves exactly one input by +20 percent against the same $0.60 task-price baseline; the sign and size of the change is what to compare across rows.
| +20% change in one input (task price $0.60, base margin 63%) | Change in gross margin |
|---|---|
| Agent resolution rate down 20% | -27.5 points |
| Failure credit up 20% | -4.7 points |
| Model price up 20% | -2.2 points |
| Tokens per task up 20% | -2.2 points |
| Cache hit share down 20% | -1.7 points |
| Escalation rate up 20% | +13.5 points (cheaper than failing, but lowers customer value) |
What to notice: the top row's effect is roughly six times the next-largest cost-side effect. An engineer might optimize model price or token count when resolution rate is the real problem. They would spend a lot of effort on a 2-point lever while a 27-point lever sits unaddressed.
{{svg:cost|Cost structure of a support task and a research task, and the effect of resolution rate.}} {{svg:pricing|Four price structures at the same average margin.}}