Exception brain
LLM-powered classifier routes cases by root cause, geography, and customer tier. It checks policy repositories and suggests next actions with confidence scores.
$12M annual savings
"Exception handling moved from 48 hours to under four."
Operational backdrop
10,000+ shipment exceptions hit ops every day—missing customs docs, incorrect weights, damaged pallets, weather events. 200 analysts triaged cases via spreadsheets and email, and leadership had little visibility into trends.
-
Symptoms
48-hour average resolution time, duplicated effort between regional control towers, and unhappy customers.
-
Hidden cost
Over $12M annually in write-offs, penalties, and emergency freight due to preventable delays.
Automation blueprint
Playbook automation
For the top 40 exception types, we automated data gathering (tracking APIs, warehouse scans, customs systems) and triggered templated responses or escalations.
Human-in-the-loop console
When confidence is low, the console surfaces the AI’s reasoning, recommended actions, and missing data so analysts can decide in seconds.
Observability
Dashboards show backlog, SLA risk, and savings by lane. Executives finally saw which trade lanes or customers created the most drag.
Six-week delivery
-
Week 1
Mapped top exception types, data sources, and the true decision tree (not the PowerPoint version).
-
Week 2–3
Built classifiers, connectors, and a shared timeline view so everyone saw the same state.
-
Week 4–6
Automated the top 40 playbooks, rolled out the console, and trained 180 analysts. Savings clocked $12M annualised.
Results
80%
Exceptions handled without human touch
4 hrs
Average resolution time (down from 48)
$12M
Annual savings in freight + penalties
Tech stack
- Event streaming via Kafka + Snowflake for historical context
- LLM orchestration on AWS Lambda with Bedrock + open-source backups
- Workflow/state machine built on Temporal with full audit logs
- Analyst console in React + tRPC embedded inside the existing ops portal
Want to automate your operations at scale?
Let's talk efficiency