Meridian Logistics — Operations AI Agent
An AI agent that reads shipment exceptions, drafts the customer update and books the recovery — inside the tools the team already uses.
The client challenge
Meridian's exception desk handled roughly 1,900 delayed or damaged shipments a month. Median time to the first customer update was 4 hours 20 minutes, most of it spent reconstructing context across four systems. Two experienced coordinators had resigned citing the workload.
Research
Before proposing a model we mapped how the best coordinator on the floor actually resolves an exception.
- Task analysis with 6 coordinators across 40 recorded resolutions
- Classification of 12 months of tickets into 9 exception archetypes
- Risk review with legal and insurance on what an agent may never decide alone
- Baseline measurement of handling time, accuracy and escalation rate
Strategy
Autonomy was granted per archetype and per freight value, not switched on globally.
- Tiered autonomy: draft-only, approve-to-send, or fully autonomous by risk band
- Grounded retrieval over historical resolutions, carrier terms and SLA documents
- Deterministic tool calls for anything that moves money or freight
- Rollout by archetype, each gated on a passing evaluation run
Experience Design
The interface is built for a coordinator reviewing 30 decisions an hour, not for demoing a chatbot.
- Queue view ranked by SLA risk, with the agent's proposal already drafted
- Every claim traceable to its source document in one click
- Single-key approve, edit or escalate
- Confidence and reasoning surfaced without burying the action
Engineering
Durable workflows, hard guardrails and continuous evaluation in production.
- Temporal workflows for retries, timeouts and long-running recoveries
- pgvector retrieval with recency and carrier-specific weighting
- 480-scenario eval suite blocking any deploy that regresses accuracy
- Full audit log of prompts, tools, outputs and human overrides
Final Showcase
The agent runs inside Zendesk and the TMS — no new tool for the team to learn.
- Console: exception queue, reasoning trace, one-key actions
- Zendesk: drafted replies attached to the existing ticket
- Ops dashboard: autonomy rate, accuracy and savings by archetype
What we shipped.
Everything handed over at the end of the engagement.
Measured outcomes.
Reported by the client after launch.
68%
Exceptions resolved without a human
-81%
Time to first customer update
+29 pts
Customer satisfaction score
0
Unreviewed actions on high-value freight
“It behaves like a careful new hire who read nine years of our tickets. The team trusts it because they can see exactly why it said what it said.”
Project gallery.
Selected screens and assets from the engagement.
Other case studies.
More recent engagements from DEZMOLAB.
