Shipsy’s predictive ETA engine replaces static, slot-based delivery windows with a rolling machine-learning estimate that updates every 90 seconds based on the vehicle’s live telemetry, the driver’s remaining task stack, and historical completion patterns for the lane. When the ETA drifts beyond a tolerance, Shipsy’s Clara agent sends a proactive message to the consignee — before they start asking “where is my order?” That single chain, predictive model plus autonomous outreach, typically cuts WISMO ticket volume by 40%+ within 60 days.
Why we built this
Static ETAs are a promise the network can’t keep. A 4-hour delivery window becomes meaningless once a driver is stuck in traffic, a hub sortation runs 30 minutes late, or an earlier customer in the sequence takes twice as long as usual. The operational pain lands in two places: CX agents flooded with “where is my package?” tickets, and customer trust eroding with every silent slip.
Enterprises were patching this with basic regression ETAs and after-the-fact SMS blasts. That wasn’t enough. They needed an ETA that learned from the actual execution of their own network, and a communications layer that would act on drift without a human in the loop.
How it works
The predictive ETA engine runs as a real-time pipeline with four feature families feeding a gradient-boosted model per network segment:
Feature family 1 — Live telemetry. Vehicle GPS position, speed, heading, and stopped-dwell state polled at 15–30 second intervals. The engine detects parking-search behavior (low-speed hovering near the address polygon) and factors it into the remaining time.
Feature family 2 — Task-stack residuals. For every pending task on the route, the model computes a remaining-service-time estimate — which varies by delivery type (B2C apartment vs B2B loading dock), time of day, and consignee history (has this location historically taken 3 minutes or 15?).
Feature family 3 — Lane and slot history. The model uses rolling 14-day averages of actual completion times on the same lane, same slot, same day-of-week, so a Wednesday 2pm delivery to a business district is priced differently from a Saturday 2pm to a residential cluster.
Feature family 4 — Real-time exceptions. Hub delays, reassignments, rerouting events, and local incident data all feed the ETA. When a trip is reassigned mid-route via dynamic rerouting, the ETA recomputes across the full affected task set within one polling cycle.
The engine publishes a rolling ETA to the consignee-facing tracking page, the merchant API, and the Clara CX agent. Clara watches ETA drift against two thresholds: a soft threshold (typical 15–20% slip) that triggers a proactive notification, and a hard threshold (30%+ slip or SLA breach) that escalates to a human CX lead with a pre-drafted explanation.
Typical notification cadence: one predictive ETA at dispatch, one “driver is 2 stops away” message, one “driver has arrived” message, and proactive interventions when drift crosses soft threshold. For B2B consignees with booked slots, the cadence extends to include dock-ready checks.
Here’s the flow at a glance:
Early results
Enterprises deploying predictive ETA plus Clara-driven outreach typically report, within 60 days:
- 40%+ reduction in WISMO (where-is-my-order) ticket volume as consignees get the answer before asking.
- ETA mean absolute error below 8 minutes on urban routes and below 15 minutes on long-haul and rural, versus 25–40 minute errors common with static ETA systems.
- Customer satisfaction lifts of 5–10 points on post-delivery surveys for the same physical network.
- Substantially higher FADR because consignees are present when the driver arrives.
A global big-and-bulky retailer leading in furniture and home goods uses this engine to drive its 95% first-attempt delivery rate — proactive narrow slot reminders dramatically reduce failed attempts.
What’s next
Three upgrades in flight: multi-language conversational confirmations via Clara (consignees can reschedule conversationally instead of tapping through a form), ML-priced SLA risk scores exposed to operations leads before the trip starts, and supplier/carrier ETA aggregation for multi-leg shipments so the first-mile delay propagates automatically to the last-mile consignee view.