Last-mile operations copilot

The leak was in the last mile, and invisible on the P&L

The carrier's first-attempt delivery rate had slipped to 81%. Nearly one in five first attempts was failing, against an industry average near 93% and top-performing fleets clearing 95–98%. In last-mile economics, that gap is where the money goes.

Dispatch allocated carriers by availability and cost, never by who actually completes the drop. Drivers marked stops failed that a single call could have saved. Recipients learned a parcel was "out for delivery" only after they had left the house.

The failure modes were textbook, and almost all preventable: the recipient not home, an incomplete or wrong address, building and gated-access issues, a recipient unavailable or refusing, and vehicle or route disruption. Each one cost more than it looked:

• ~$17.20 per failed attempt, once redelivery, customer-service inbound and SLA penalties were tallied.
• Climbing return-to-origin volumes, dragging on inventory turnaround and depot throughput.
• Invisible customer cost: nearly 7 in 10 recipients hit by a botched delivery never ordered again.

A copilot isn't new. Pointing one at dispatch is

The idea is simple: a system that watches the whole picture, anticipates what is coming, and recommends the next best move, while the human stays in command.

Aimed at last-mile operations, the everyday math of dispatch changes. Instead of reacting to failures after they land, the copilot reads every signal (address quality, carrier history by postcode, live ETAs, recipient behaviour) and intervenes before the window is blown. The dispatcher still decides. They just decide with a copilot that never blinks.

Dispatch before: reactive, with blind spots. Carriers picked by cost and availability, not completion rate. Misses discovered after the failed scan. The driver improvising alone at the door. No learning loop, so the same routes failed again.

Dispatch with a copilot: preventive and guided. The carrier recommended by its first-attempt record per postcode. Impending misses forecast and rerouted in flight. The driver given the next best action and a way to reach the recipient. Every outcome fed back into the model, so it gets sharper.

One decision layer, on the stack they already ran

No platform to buy, no migration. The copilot sits over the carrier's existing OMS, TMS, Salesforce, ServiceNow and Azure estate as a decision layer, reading their data and intervening at every decision point from intake to doorstep:

1. Validate at the source. Addresses verified and auto-corrected at capture.
2. Close the comms gap. Telematics-fed ETAs and two-way messaging with the recipient.
3. Recommend the carrier. Zone-level first-attempt scoring at dispatch.
4. Prevent misses in flight. Misses forecast and auto-rerouted before the window is blown.
5. Empower the driver. The next best action and recipient contact at the door.

Delivered as a service, not a product you buy

We don't drop a tool and leave. The copilot was engineered into the carrier's environment, run against live results, and handed to their team:

• Embed and discover. We sat with dispatch and operations, instrumented the data, and quantified the first-attempt gap on their real volumes.
• Co-engineer the copilot. The decision layer built on their existing stack, no rip-and-replace.
• Operate, tune and hand off. Run and tuned against live outcomes, then ownership transferred to their team, so the model kept improving rather than being handed over and left to drift.

The doorstep stopped being a cost centre

Within two quarters of go-live across all six depots, on the carrier's own systems, first-attempt delivery rose from 81% to 96%, into best-in-class territory.

• +15 points of first-attempt delivery, from 81% at go-live to 96% two quarters later.
• 30% fewer failed attempts from accurate ETA notifications alone.
• 40% less return-to-origin volume across the network.
• 18-point uplift in Net Promoter Score on the delivery experience.
• 7× return on the engagement in year one, with full payback inside seven months.
• Nothing ripped out: built entirely on their existing stack.

"Praval didn't sell us software and walk away. They embedded, built our last-mile copilot on the stack we already had, and stayed to make it stick."

— VP of Operations, regional parcel carrier

An illustrative engagement. The client metrics are representative of comparable last-mile deployments, not the audited results of a single named client. Industry benchmarks: Harvard Business Review; 2025–26 industry first-attempt delivery data.