The Cost Model

Cost per resolved ticket is the number that decides whether an agent is a product or a demo. Sundry's support agent started at $0.41 and ends this book at $0.06, and not one cent of that came from a cheaper model or a better deal with the provider.

Read the chain before the arithmetic, because the headline hides where the work happened. Prefix caching took 41 cents to 14 in Chapter 5. Routing by task type took 14 to 11 in Chapter 7. The triage split in Chapter 10 then put 11 back up to 12 on purpose, buying handle time with money and saying so. So this page starts at twelve cents and takes six off — not thirty-five — and those six are the hard ones, because everything obvious has already been done.

The Unit Is a Resolved Ticket

A ticket that ran twelve turns, hammered track_parcel nine times, hit the loop limit and escalated cost $0.34 and settled nothing. Chapter 8 tells that story as a control-flow failure; on a cost report it is worse than that, because per-call and per-ticket averages both absorb it silently. Only the per-resolution figure treats it as what it is: money spent, work not done, and a person still needed.

So the denominator is resolutions, not tickets, with escalations left in the numerator where they belong. At 4,200 tickets a week and 88% resolution, six cents a ticket is $252 of model spend against 3,696 resolutions — 6.8 cents each. On the twelve-cent figure this chapter starts from, the same arithmetic gives 13.6 cents. That is the number to put next to what a person costs, and it is the only one that survives contact with somebody who asks what happens to the other 12%.

Where the Money Went

Nobody can optimize a number they cannot decompose, so this topic follows tracing rather than preceding it. Once every model-call span carries its token counts, tagging each one with what the tokens were spent on — schemas, history, tool results, retrieved passages, deliberation, the final reply — turns the invoice into a breakdown. Here is Sundry's, at twelve cents a ticket, blended across the whole queue.

The top row is the one that surprises teams and should not. A tool result is not a return value that disappears — Chapter 3 established it as context, re-sent on every remaining turn of the run, so a fat result arriving on turn 3 of a nine-turn ticket is billed six more times. The bottom row is where the system prompt lives. It is 7% of the bill, and it is where most cost-reduction effort in this field goes, which is how a fortnight of careful prompt editing ends up worth 4%.

The Levers, Ranked by What They Actually Returned

Every lever below already exists in this book. What Chapter 13 added was not a technique but the breakdown, which said where to point the techniques next. Each row was measured on its own against the same replayed week of production traffic, so the numbers are comparable to each other rather than to a plan.

Two things about that table are worth more than the ordering. It sums to 6.4 cents and the measured result of shipping all six together was 6.0, because the levers overlap — tokens removed by shaping are tokens the cache no longer has to hold, and a run that ends four turns earlier stops paying for everything at once. Report the joint figure, not the sum. And the largest row is a technique Chapter 3 taught nine chapters ago, applied to two tools it never got around to: get_order was shaped early because it was obvious, while the carrier's full scan history — around 1,100 tokens for a parcel with a long journey — went straight into context untouched until the breakdown pointed at it. Trimmed to the last three scans plus the delivery state, it is 150 tokens.

Model Routing

Route by task, not by agent. Sundry's triage agent already ran on a small fast model in Chapter 10; what this chapter added was splitting the specialists' own turns — classification, field extraction and composing the reply for the status class go to the small model, while policy judgement and any turn that can end in money stay on the large one. On the 120-ticket eval set resolution held at 106 of 120, the trajectory score did not move, and the blended saving was 1.5 cents.

The version before that one was greedier and got rejected, which is the more useful half of the story. Sending the whole returns class to the small model saved another 0.9 cents and dropped resolution to 97 of 120 — seven points. Seven points on 4,200 tickets a week is 294 more tickets landing on a person at roughly $2.40 each, about $700 a week, to save $38. Route a class to a cheaper model without running the eval set and that is the trade you are making blind; run it, and the answer takes eleven minutes and costs less than lunch.

Budgets and Alerts

Two ceilings live in code, next to the turn limit from Chapter 1. A per-run ceiling of 50 cents ends the run stuck and escalates with a summary, the same way the turn limit does, because a run that has spent eight times the mean has already failed at something. A per-day budget trips a circuit breaker that routes new tickets to the human queue rather than silently degrading — an agent that quietly stops trying is worse than one that visibly stops.

The alert is on the ratio, never the total. Sundry's total model spend doubled on a Monday afternoon when a seller's consignment went missing and 340 near-identical tickets arrived in three hours; cost per resolution never moved, because nothing was wrong. A regression looks completely different: volume flat, cost per resolution up 25% over a six-hour window against the seven-day median, which is the condition that pages somebody. On total spend those two events are the same graph, and a team that alerts on total spend learns within a month to ignore the alert.

Comparing Honestly

Sundry's support team handles a ticket in about six minutes at a loaded cost near $24 an hour, so the human baseline is roughly $2.40. Six cents against $2.40 is the number that gets put on a slide, and it is wrong by a factor of nearly three, because the agent does not handle the queue on its own and never will.

Sixteen and a half cents against $2.40 is still a large number and it is a defensible one. It survives the first question anybody senior asks, which is what happens to the tickets the agent cannot finish, and the second, which is what it costs when the agent is confidently wrong. State one more thing while the ledger is open: the human baseline carries its own error rate, and at Sundry nobody had ever measured it — so the fair reading is that the agent's mistakes are counted and the team's are not, and the comparison is generous to the agent in exactly one place.