AI for USPS Dynamic Routing Needs a Service-Cost Baseline Before Deployment

Artificial intelligence can help postal operators adjust delivery routes as traffic, parcel volume, weather, and vehicle capacity change. But an optimized route is not automatically a better route. Before deploying AI-powered dynamic routing, the U.S. Postal Service needs a baseline that shows what each route costs and what service it actually delivers.

Without that baseline, a pilot may report fewer miles while quietly increasing overtime, failed deliveries, customer complaints, or workload on adjacent routes. The right starting point is not an algorithm. It is a measurable definition of operational improvement.

USPS has real AI momentum

The opportunity is more than theoretical. The USPS Office of Inspector General identified dynamic route optimization as a promising application as the agency tries to move from isolated AI pilots toward interconnected operating systems. According to Supply Chain Dive's summary of the OIG report, USPS already has more than 35 active AI use cases, with more applications in testing.

Those systems are producing tangible results. Since January, AI-assisted detection helped USPS identify and close 1,250 accounts generating counterfeit shipping labels. Machine learning also supports estimated delivery dates through the Advanced Expected Delivery program. Dynamic routing is a logical next step because real-time traffic and volume data could help reduce mileage and protect delivery commitments.

The financial context makes disciplined measurement essential. USPS reported a $2.5 billion net loss in its fiscal 2026 third quarter. In the same quarter, air transportation expense rose 4.7% year over year to $509 million, while highway transportation costs increased 4.1% to nearly $1.6 billion, according to Supply Chain Dive. Although those figures cover transportation between facilities rather than last-mile routes alone, they show why every claimed efficiency must survive financial scrutiny.

Build the baseline before changing routes

A useful baseline should cover at least six dimensions for every route and operating day:

1. Miles and drive time: Planned versus actual miles, travel time between stops, deadhead distance, and deviations caused by closures or traffic.
2. Stops and pieces: Delivery points served, parcels and mailpieces handled, pickups completed, and volume by size or handling requirement.
3. Service performance: On-time completion, missed delivery windows, late scans, customer complaints, and commitments carried into the next day.
4. Labor: Straight-time hours, overtime, auxiliary assistance, supervisor intervention, and time spent sorting or loading before departure.
5. Delivery exceptions: Failed attempts, inaccessible locations, incorrect addresses, signature failures, vehicle returns, and rework.
6. Operating constraints: Vehicle cube and weight, charging or fueling needs, break rules, facility dispatch windows, and carrier familiarity with the territory.

USPS should capture these measures for comparable weeks, including normal volume, peak periods, severe weather, and local disruptions. A single average day will produce a brittle model. Baseline data should also distinguish controllable route inefficiency from conditions a carrier cannot fix, such as an incorrect address or a blocked road.

The primary score should combine service and cost rather than reward one metric in isolation. A simple formulation is cost per successfully completed delivery point, paired with guardrails for on-time completion, overtime, failed attempts, and safety events. Mileage matters, but it cannot be the sole target.

Give the model the constraints it needs

Dynamic routing depends on timely, trustworthy inputs. Address quality is foundational: apartment access instructions, business receiving hours, signature requirements, and known delivery-point restrictions need consistent formats. Weather and traffic feeds must be time-stamped so planners can distinguish a forecast from an observed condition.

Capacity should be modeled at vehicle and route level. Parcel cube can become binding before weight, while mail trays, pickups, and accountable items reduce usable space. The optimizer also needs each service commitment. A shorter route that makes a priority item late is not an optimization.

Private carriers offer a useful operating lesson. Supply Chain Dive reports that UPS prioritizes clean, accessible data and uses digital-twin simulations to test routing adjustments before implementing them in its network. USPS does not need to copy another carrier's system, but it should adopt the principle: simulate a recommendation against historical operating days before asking employees and customers to absorb the consequences.

Keep people in control of deployment

The OIG noted that labor-union collaboration and agreement will be important for AI-driven route optimization. That should be treated as part of system design, not a late-stage communications task. Letter carriers understand access restrictions, unsafe turns, school-zone patterns, seasonal obstacles, and customer needs that may not appear in centralized data.

Every pilot should therefore include a carrier override with a short reason code. Overrides should not be counted automatically as noncompliance. They are valuable training data: repeated overrides may reveal a bad address attribute, an unrealistic travel-time assumption, or a missing safety constraint.

Audit controls should preserve the original plan, the AI recommendation, any human change, and the final outcome. Supervisors need to see which input or objective caused a recommendation. Routes should never change through an untraceable black box, especially when workload and contractual work rules are affected.

Run a pilot that can prove net value

Start with a small group of representative routes and a matched control group. Run the system in shadow mode first, generating recommendations without changing execution. Compare predicted mileage, completion time, and service against what actually happened. Then move to limited live use with clear stop conditions.

Evaluate the pilot over enough days to include changing volumes and operating conditions. Report net results across the whole delivery unit, not only the routes receiving recommendations. If one route saves 30 minutes by transferring stops that create 45 minutes of overtime elsewhere, the pilot lost value.

USPS should advance deployment only when savings remain after software, integration, training, supervision, and exception-handling costs are included—and when service, safety, and workload guardrails remain intact. That is the difference between a promising demonstration and a sustainable operating capability.

Turn routing data into operational decisions

Dynamic routing works best when route, shipment, vehicle, labor, and exception data share one operational view. CXTMS helps logistics teams connect those signals, measure cost and service together, and manage exceptions without losing human oversight.

Request a CXTMS demo to see how a transportation management platform can support measurable, controlled route optimization.