Ticket intake
omnichannel
- feeds
- Classifier
All case studies
NLPAutomationOps2 months
Support Ticket Triage + Auto-Routing
A third of support tickets went to the wrong team. Now it is closer to one in thirteen.
Routing accuracy
92.4%
Calibrated multi-label classifier
Was
65.3%
Keyword rule engine
In plain terms
What this actually was.
No jargon in this section. The technical write-up is further down.
What was going wrong
Tickets were sorted by keyword rules that had been patched for years. A third went to the wrong team, and every mistake cost a full handoff before anyone started helping — which was the single biggest reason customers rated support poorly.
What we built
We replaced the rules with a classifier trained on how tickets were actually resolved, and retired the old rule engine rather than adding one more rule to it.
What changed
Correct routing went from 65% to 92%, first replies got faster, and satisfaction rose as a direct consequence. Agents stopped sorting tickets and started resolving them.
Architecture
How it fits together.
Scroll through the stages. Anything marked as added is a component that did not exist before this project.
Classifier
multi-label
- receives from
- Ticket intake
- feeds
- Calibration
Calibration
cost-aware
- receives from
- Classifier
- feeds
- Router · Reply draft
Router
queue assign
- receives from
- Calibration
- feeds
- Agent
Reply draft
from resolved
- receives from
- Calibration
- feeds
- Agent
Agent
resolves
- receives from
- Router · Reply draft
Ticket intake
omnichannel
- feeds
- Classifier
Classifier
multi-label
We added this- receives from
- Ticket intake
- feeds
- Calibration
Calibration
cost-aware
We added this- receives from
- Classifier
- feeds
- Router · Reply draft
Router
queue assign
- receives from
- Calibration
- feeds
- Agent
Reply draft
from resolved
We added this- receives from
- Calibration
- feeds
- Agent
Agent
resolves
- receives from
- Router · Reply draft
The write-up
How it was actually built.
Dataset, approach, measured results and the stack. Written for whoever has to review it.
NLPAutomationOps2 months
Support Ticket Triage + Auto-Routing
Classification, tagging, and routing that keeps queues clean and responses consistent.
Problem
Tickets were routed by keyword rules that had accreted for years. Misroutes cost a full handoff cycle, and first-response time was the top driver of poor CSAT.
Dataset
- rows
- 2.4M tickets
- features
- 128
- sources
- Helpdesk, chat, email
- window
- 36 months
Approach
- 01Trained a multi-label classifier over historical resolved tickets
- 02Calibrated probabilities so routing thresholds map to business cost
- 03Added intent-based reply suggestions drafted from resolved cases
- 04Shipped behind a shadow-mode rollout compared against the rule engine
Metrics
- 92.4%
- routing accuracy +27.1pp
- 4.2 min
- first response -38 min
- 7.6%
- misroute rate -27pp
- 19%
- deflection self-serve
Business impact
- CSAT improvement driven directly by faster first response
- Agents stopped triaging and started resolving
- Rule engine retired rather than extended again
Tech stack
- Python
- FastAPI
- XGBoost
- Elasticsearch
- PostgreSQL
- Celery
- Docker
Next
Need a system like this?
Answer six questions and we will tell you whether this shape fits your problem — including when it does not.