This dashboard is a teaching artefact, not a record.
Northwind Outfitters is a fictional company. The twelve-month operating history below was invented to describe a plausible staged rollout, and every chart drawn from it is badged simulated. A few numbers are real — eval accuracy, per-ticket cost, cache economics and the category mix — and those are badged measured, meaning they came out of the actual system.
Internal · Support Operations
Triage programme review
Prepared for Priya Raman, Director of Support. Twelve months to December 2026, covering the staged rollout of automated triage from shadow pilot through full deployment.
Tickets triaged by hand
19%
was 100%
before triage
Safety and low-confidence tickets are still routed to a person by rule, so this floor is deliberate.
Median first response
7.4h
was 14.0h
before triage
December peak. Last year's December ran at 41h on less than half the volume.
Mis-routing rate
7%
was 23%
before triage
Every mis-route costs a re-read and a re-queue.
Safety time to queue
24m
was 25.4h
before triage
Clause 5.4 sets a one-hour ceiling with zero tolerance.
Tickets a human had to triage
simulatedThe number this programme exists to move. Shaded bands mark each rollout phase.
It does not go to zero, and it should not. Safety reports and anything below the confidence threshold route to a person by rule.
Volume, and who handled it
simulatedDecember volume more than doubled. The human-handled band barely moved.
- Routed automatically
- Handled by an agent
This is the chart to show a CFO. Peak season used to mean seasonal hires and a 41-hour response time; the load that reaches a person is now roughly flat against volume.
Median time to first response
simulatedHours. The November and December rise is peak volume, not regression.
Triage overhead was the removable part of this number. What remains is the time to actually resolve things.
Safety reports: time to reach the safety queue
simulatedHours, against the one-hour ceiling in handbook clause 5.4.
The October 2025 incident sat for three days. This is the metric that exists because of it, and the only one with zero tolerance.
From the live system
measuredEverything above this line is invented. Everything below came out of the triage service itself — the same code path the support form runs.
Category mix
measuredTwenty real tickets run through /v1/triage against claude-opus-5.
Seven categories is past the point where colour can carry identity, so this is a magnitude chart in one hue with the values labelled directly.
Unit economics
measuredPer ticket, measured against list pricing.
- Warm cache
- $0.0053
- Cold cache
- $0.0302
- Saved by caching
- 82%
- December model spend
- $259
of a $4,000 budget
Without prompt caching the same December volume would cost about $1477, which is over budget on its own.
Is the confidence score trustworthy?
measured92% on a 12-case gold set.
What matters is the gap, not the accuracy. Wrong answers score lower than right ones, which is what makes threshold routing safe to build on.
How it was rolled out
Staged rather than switched on. Each phase had a way to be wrong that did not cost a customer anything.
- 1Manual triage
Every message read and routed by an agent.
- 2Shadow pilot
Triage runs on every ticket but routes nothing. Output compared against the agent's decision.
- 3Partial rollout
Automated routing for shipping, returns and account only. Billing, defects and safety stay manual.
- 4Full rollout
All categories routed automatically. Safety and low-confidence tickets still go to a human by rule.
Agent hours returned to resolution work in December: 383 hours a week against what the same volume would have cost under manual triage. That is the number Priya reports upward, and it is the one that pays for the programme.
The system behind these numbers is documented at the Claude API triage labs. For who Priya is and why these four metrics, read the scenario.