- Who it helps
- Teams evaluating lower-cost structured classification
- Availability
- November 9 implementation report; a deployment example, not a managed ticket-classification product.
Deployment checks
The report describes twenty ticket labels plus a no-match choice, a held-out evaluation and parallel classification before switching production traffic. Decider constrains the output to the supplied options; it does not establish that the selected label is correct.
Cost and input limits
The authors estimate a 98% inference-cost reduction under their utilization assumptions, excluding Lambda. The shorter context dropped older comments, and JSON key ordering affected outcomes. These details matter when comparing a replacement on real tickets.
What you can do
- Review the shadow-mode comparison, error cases and full cost exclusions before designing a replacement. Open reference ↗
Where the evidence stops
Source report checked November 11, 2026; deployment and cost results were not reproduced. The estimate is specific to the reported hardware, workload and utilization. Review routing errors before applying a similar design.
Sources read. Public source text checked November 11, 2026 UTC. No third-party project or complete workflow executed.
Sources & further reading
Web · official
Decider ticket-classification deployment ↗
November 9 deployment report describes shadow comparisons and an inference-cost estimate. Costs exclude Lambda.
Carr Onstott and Mehul Jaiswal · page checked 2026-11-11 · Published 2026-11-09
Coverage note: Deployment, accuracy comparison and utilization assumptions were not independently reproduced.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
An Amazon team reports replacing a large-model ticket classifier with a smaller model that selects among explicit labels, after comparing both systems in shadow mode.