Self-Healing Supply Routing
A routing agent reroutes disrupted shipments on its own below a $25,000 cost threshold — above it, a dispatcher decides, but never from scratch.
The challenge
Six distribution centers and roughly 1,200 active lanes meant disruptions — weather, carrier delays, port congestion — arrived constantly, and every one was worked the same manual way:
- About 150 disruption signals a week, each requiring a fresh routing decision
- Dispatchers pulling rates from separate carrier portals with no unified cost comparison
- No consistent threshold for when a reroute needed a manager's sign-off versus a dispatcher's own call
- High-cost reroutes decided under the same time pressure as routine ones
- Shipments sitting idle while a dispatcher worked the phones for alternatives
At 150 disruption events a week and 47 minutes of dispatcher time each, deciding where to send affected freight consumed roughly 118 dispatcher-hours weekly, before a single truck moved.
How it works
A dollar limit the agent can't cross on its own
Rather than automate every reroute decision, the network drew one hard line and let the agent operate freely inside it:
- 01
Real-time disruption signals are ingested continuously — weather feeds, carrier EDI delay codes, port congestion indices
- 02
Every affected shipment's reroute options are scored on landed cost, transit time, and service-level risk
- 03
The agent auto-executes any reroute where the cost delta against the original plan is under $25,000
- 04
Anything at or above that threshold is not decided by the agent — it generates three priced alternative routes and escalates to a human dispatcher
- 05
The dispatcher's choice and its outcome are logged and fed back into the route-scoring model
- 06
The $25,000 threshold is reviewed monthly against the network's actual cost exposure and risk appetite
What we built
Key capabilities
A hard dollar boundary, not a suggestion
The $25,000 threshold is enforced in the routing logic itself — the agent cannot execute a reroute above it under any confidence score.
Escalation arrives priced, not blank
When a decision goes to a dispatcher, it comes with three costed alternatives already built, not a bare alert to start researching from.
Disruption signals, not just delay reports
Weather, carrier EDI codes, and port congestion indices feed the agent continuously, so rerouting starts before a shipment is already late.
Learns from every escalated decision
Dispatcher choices on the 24% of events that cross the threshold retrain the scoring model that ranks alternatives next time.
Before vs after
What changed in dispatch
- Weekly disruptions resolved without a person
- 0 → 114 (76%)
- Avg. decision time, auto-resolved reroutes
- 47 min → 4 min
- Avg. decision time, escalated reroutes
- 47 min → 12 min
- Dispatcher-hours on routing decisions weekly
- 118 → 11
Business impact
What it changed
107 dispatcher-hours reclaimed weekly
118 hours of manual routing work fell to 11 once the agent absorbed the 76% of disruptions under the $25,000 threshold.
A threshold the business actually set
$25,000 isn't a technical artifact — it's a risk boundary reviewed monthly against real cost exposure, not a number the model picked for itself.
Escalations decided in minutes, not phone calls
Even the 24% of events that still need a person are decided in 12 minutes instead of 47, because the alternatives arrive already priced.
Technology stack
“Escalation isn't a fallback here — it's a design decision. The agent earns the next dollar of authority only after the last one holds up.”
Keep reading