Trucking and last-mile delivery: AI in the cab and at the dock

Trucking and last-mile delivery is one of the largest and most fragmented operational categories in the economy. The work involves millions of trucks, drivers, and routes across many segments — long-haul, regional, local delivery, last-mile parcel, white-glove delivery, food and grocery delivery, freight brokerage, intermodal coordination.

AI has been integrated across these segments in the past few years. The economics, the deployment reality, and the regulatory framework all favor local inference for the bulk of the operational work.

The work

Trucking and delivery AI workloads include:

Dispatch and load matching: matching drivers to loads based on equipment, location, hours of service availability, and customer requirements. The work happens at the dispatch center but increasingly on driver-facing apps.

Driver support: real-time AI assistance for drivers — route guidance, hours-of-service questions, equipment troubleshooting, customer location details, delivery instructions.

Route optimization and documentation: drafting route plans, generating optimization narratives, documenting deviations and exceptions, producing the documentation that customers and dispatch require.

Customer communications: drafting delivery confirmations, generating exception communications, producing the proof-of-delivery documentation that customers expect.

Compliance documentation: drafting hours-of-service logs (ELD-driven now but still requiring documentation), drug and alcohol testing records, IFTA fuel tax documentation, hazmat documentation when applicable.

Driver retention and management: drafting communications to drivers, generating performance feedback, producing the documentation that DOT inspections and internal reviews require.

Maintenance and equipment: drafting maintenance work orders, generating equipment performance reports, producing the documentation that fleet management requires.

Broker and shipper communications: for freight brokers, drafting carrier communications, generating shipper updates, producing the documentation that brokerage operations require.

The math

Trucking has an unusual cost structure for AI workloads. Volume per truck is moderate, but truck count is enormous in the industry as a whole.

A representative midsize trucking company — say, two hundred to five hundred trucks — generates a substantial aggregate AI workload across dispatch, driver support, customer communications, and documentation. The per-truck per-day volume is in the tens of operations; across the fleet and the day, the aggregate is in the thousands.

At a representative cost, the bill is in the low to mid five figures per year for a midsize operator.

For larger operators — major LTL carriers, large TL fleets, national parcel operators — the numbers scale to the high six figures or low seven figures per year. For the largest trucking operations and the largest last-mile platforms (Amazon Logistics, UPS, FedEx, USPS), the figures climb to seven and eight figures per year.

For freight brokers, the cost structure is different — brokers handle large volumes of communications without operating equipment. Major freight brokers have similar AI bills to large carriers, but the work shape differs.

Why trucking is a strong local-SLM workload

The deployment reality and the cost economics both favor local inference.

The connectivity reality is mixed. Highway driving generally has good cellular coverage; rural areas and certain terrain have poor coverage; loading docks and warehouses can have spotty coverage. Cloud-LLM workflows in driver-facing applications fail when coverage drops; local-SLM workflows on driver devices work regardless.

The latency story matters at the dock. Drivers waiting for documentation, customer interaction, or dispatch guidance need response in seconds, not minutes. At a busy dock, cloud round trips in poor coverage are operationally disruptive.

The work is narrow within the operator. Each trucking company has its specific lane mix, customer base, equipment types, and operational style. A model fine-tuned on the operator's corpus outperforms a general model.

The work is enormously repetitive. The same kinds of pickups, deliveries, dispatch decisions, and exceptions, repeated across millions of loads per year at any major operator. Specialization compounds aggressively.

The privacy and competitive intelligence story is real in freight brokerage especially. Broker-carrier relationships, shipper-broker pricing, lane-specific economics — all competitively sensitive information that defines the broker's market position. Sending through third-party cloud LLMs creates exposure that the brokerage's commercial team takes seriously.

The brand voice story matters in customer-facing communications. Trucking and last-mile delivery is one of the customer touchpoints where brand experience increasingly matters. Communications that sound generic undermine the brand; communications that sound like the specific operator reinforce it.

What changes with local inference

A trucking and delivery AI workflow on a local SLM looks like this.

A model is fine-tuned on the operator's corpus — historical dispatch decisions, driver communications, customer interactions, compliance documentation. The fine-tune captures the operator's specific patterns.

The model deploys across the operator's environment — at the dispatch center for dispatch operations, on driver devices for in-cab and at-dock support, at customer-facing systems for shipper and consignee interactions. The on-device deployment is critical for driver-facing workflows.

Operations flow through the inference pipeline. Dispatch decisions, driver support, customer communications, compliance documentation — all produced locally, all working regardless of connectivity status.

The cost flips from per-operation to fixed. Fleet growth doesn't scale the AI bill.

The competitive intelligence stays inside.

The fleet-scale economics

A specific dynamic in trucking: the fleet-scale economics make the local-SLM case clean at scale.

At the largest trucking operators — major LTL carriers, large parcel operators, national TL fleets — the per-truck cost of cloud LLM at scale becomes a significant operational expense. The cost compounds with fleet growth.

Moving to local inference at the fleet level produces savings across the entire fleet simultaneously. The infrastructure investment pays back quickly at scale. The operational benefits of consistent AI assistance across the fleet are real.

For the largest last-mile platforms in particular — Amazon Logistics with hundreds of thousands of contractor drivers, UPS with hundreds of thousands of employee drivers, FedEx with its mixed model — the architectural conversation is increasingly active. The platforms that move to local inference first will operate at structural cost advantage.

Where the cloud LLM is still acceptable

A few cases.

For very small operators — owner-operator drivers, small fleets — where the infrastructure investment doesn't pay back at small scale.

For research and analytics workflows operating on aggregated industry data.

For internal training and continuing education content.

For most trucking and last-mile operations of meaningful scale, the local-SLM case is strong on cost, on deployment reality, and on competitive intelligence considerations.

The pattern, on the road

Avery NXR is not a trucking tool. It scaffolds Next.js applications. The architectural pattern repeats, with the fleet-scale economics and the deployment reality giving it specific shape.

Trucking and last-mile AI is a narrow, enormously-repetitive, moderate-per-truck-volume, fleet-scale-cost-meaningful, connectivity-constrained workload. The cost case is meaningful at fleet scale. The deployment case is real on driver-facing devices. The competitive intelligence case applies particularly to freight brokerage.

The trucking and last-mile technology vendors that build on local infrastructure — with appropriate fine-tuning, integration with the major TMS platforms and ELD systems, and pricing models that fit different operator scales — will own the institutional segment. The cloud-LLM-default products will hold pockets at the long tail but face structural friction at fleet scale.

The pattern continues. Trucking is one of the workflows where the local-SLM case is supported by fleet-scale economics, deployment reality, and competitive intelligence — and where the largest operators have the most to gain from moving first.