Executive Thesis
Parcel organizations already possess many of the ingredients required for better decisions: contracted carrier services, rate information, package data, delivery commitments, tracking events, and historical performance. Yet those ingredients do not create value automatically.
Value is created when evidence changes an operational decision.
In parcel shipping, one of the most consequential decision points is label creation. That is where the organization commits to a carrier/service option, a transportation cost, and an expected delivery outcome.
The supplied Luma AI Select case study demonstrates the economic potential of improving this layer. A global recommerce marketplace processing more than 25,000 labels per day through EasyPost reported more than $2 million in annual savings, a 4-5% reduction in per-label cost, an increase in on-time delivery from 80% to 83%, and approximately 273,000 fewer late deliveries annually after deploying shipment-level AI service selection [1].
The customer did not add carriers, renegotiate contracts, change packaging, add headcount, or make significant fulfillment changes during the reported measurement period. The result was associated with a change in allocation logic rather than a network rebuild.
This is the decision layer.
Layer 1: Commercial Options
Carrier contracts define what the organization can buy.
They determine available services, pricing structures, coverage, commitments, package constraints, and commercial terms. Procurement quality matters because poor options cannot be optimized into excellent outcomes.
But a strong contract is only potential value.
If the organization repeatedly chooses an unnecessarily expensive service, or chooses a low-cost service whose delivery performance is inappropriate for the shipment, negotiated rates do not guarantee efficient execution.
Commercial options are the input to the decision layer, not the decision itself.
Layer 2: Operational Evidence
The next layer is evidence.
Relevant shipment evidence can include destination, package profile, delivery window, cost, service eligibility, and historical carrier performance. The Luma case study says its model evaluates these factors and draws on historical performance across more than one billion comparable shipments.
Evidence quality matters in four dimensions:
Granularity
Can the organization see meaningful differences by service, lane, zone, package, or destination rather than relying only on aggregate averages?
Freshness
Is the evidence recent enough to reflect current operating conditions?
Completeness
Are the variables that materially affect the decision present?
Comparability
Are historical outcomes sufficiently similar to the shipment being evaluated to inform the choice?
Data that fails these tests can create false precision.
Layer 3: Decision Logic
This is where shipping AI can create differentiated value.
Decision logic converts evidence into a ranked choice among eligible services.
The business must define what "better" means. Possible objectives include:
- Lowest expected transportation cost subject to an on-time threshold.
- Highest expected on-time probability under a cost ceiling.
- A weighted balance of cost, service, customer promise, and operational risk.
The objective should be explicit because different objectives produce different routing decisions.
A system that simply selects the cheapest available service is not optimizing customer promise. A system that always selects the fastest service is not optimizing economics.
The decision layer exists to manage the trade-off deliberately.
Layer 4: Execution Authority
A recommendation creates no shipping value until it influences the label.
This is why decision-time integration matters.
The Luma case study describes AI selection occurring at label creation. That closes the gap between analysis and execution.
A governed system should define which recommendations may execute automatically and which require review. Routine, high-confidence shipments may be suitable for automated selection. Unusual package profiles, poor-quality evidence, restricted services, or high-risk customer commitments may require human intervention.
The correct goal is not maximum automation. It is the minimum human friction consistent with acceptable risk.
Layer 5: Outcome Read-Back
After the package is delivered, the organization receives the most important evidence: what actually happened.
Did the selected service cost what was expected?
Did the package arrive on time?
Was there an exception?
Did a human override improve or worsen the outcome?
Did the model systematically overestimate a service on a particular lane?
Outcome read-back converts a one-way recommendation engine into a closed decision system.
Without it, AI performance is inferred. With it, performance can be measured.
Why This Matters More in a Diversified Network
2026 parcel strategy continues to emphasize diversification. PARCEL argues that parcel sourcing should increasingly resemble portfolio design. TransImpact describes diversification as a lever for cost, negotiating power, and resilience. DCL Logistics recommends selecting carrier models based on shipment fit rather than searching for one universal winner.
As the option set expands, the allocation problem becomes more complex.
A shipper with one carrier and two services has relatively few combinations. A shipper with multiple carriers, multiple service levels, multiple package types, and varied customer promises has a much larger decision space.
This means diversification and AI decisioning are complementary:
Diversification increases optionality.
Decision intelligence manages optionality.
The value of one can increase the need for the other.
The Luma Evidence Boundary
A strong analysis must distinguish verified facts from interpretation.
Verified in the supplied case study:
- 25,000+ labels per day.
- Primary carrier on-time performance of 80% before deployment.
- Approximately 5,000 late packages per day at that baseline, as calculated in the case study.
- Shipment-level Luma AI Select decisioning at label creation.
- 4-5% lower per-label cost.
- $2M+ annual savings.
- On-time performance improving to 83%.
- Approximately 273,000 fewer late deliveries annually.
- No added carriers, renegotiated contracts, added headcount, packaging change, or significant fulfillment change in the measured deployment.
- Before/after comparison across comparable U.S. shipments.
- 10× more U.S. shipping volume shifted to EasyPost after the customer observed the results. [1]
Unknown from the supplied case study:
- Customer identity, which was withheld at the customer's request.
- Exact carrier names.
- Exact per-label baseline dollar cost.
- Exact implementation duration.
- Exact statistical model architecture.
- Whether identical results would occur in another shipper's network.
These unknowns should remain unknown. They are not reasons to dismiss the case; they define the boundary of what can responsibly be claimed.
A Decision-Layer Architecture
Input plane
Shipment, customer promise, package, destination, eligible services, commercial constraints.
Evidence plane
Current cost and historical performance at the most relevant supported granularity.
Decision plane
Explicit objective and ranking logic.
Control plane
Automation thresholds, human review, exceptions, fallback, incident suspension.
Execution plane
Label purchase and shipment tender.
Outcome plane
Actual cost, delivery performance, exceptions, customer impact.
Learning plane
Performance monitoring, calibration, policy review, and evidence refresh.
Each plane has a distinct owner and failure mode. Treating the entire system as "the AI model" obscures operational accountability.
Who Owns What
Operations/Logistics
Own operational objective, exceptions, and service-performance review.
Transportation/Supply Chain
Own carrier portfolio strategy, allocation policy, and network constraints.
IT/Engineering
Own API and label-workflow integration, decision-time data reliability, latency, observability, fallback behavior, and technical controls.
Fulfillment/Warehouse
Own operational compatibility with pick-pack-label-tender workflows, with the goal of improving selection without forcing unnecessary process redesign.
Product
Own customer-promise implications and platform/customer experience where relevant.
Finance/Procurement
Validate cost methodology and carrier economics.
Leadership
Approve material policy, risk, and authority boundaries.
A 30-Day Diagnostic
Week 1
Map the current label decision and static rules.
Week 2
Build baseline cost and on-time views by service and relevant shipment segments.
Week 3
Generate alternative recommendations in shadow mode and quantify decision differences.
Week 4
Review economics, service impact, exceptions, data quality, and governance readiness before deciding whether to pilot.
The diagnostic should end with evidence, not a predetermined technology purchase.
2026 AI Context
HERE Technologies' May 2026 Location Reasoning announcement is useful context for the evolution of operational AI. HERE describes grounding AI agents in live location and road-network intelligence to improve real-world decision reliability. Its separate AI-powered last-meter guidance announcement applies AI directly to delivery execution.
These technologies solve different problems from Luma AI Select, but the design pattern is consistent: operational AI becomes valuable when context is attached to a bounded decision close to the action point.
The parcel decision layer follows the same principle.
Operational Maturity Model
Organizations can assess shipment-level decisioning across four maturity stages. In the first stage, routing is predominantly static and performance is reviewed retrospectively. The operation may have strong carrier contracts and reporting, but the decision made for an individual shipment changes infrequently.
In the second stage, teams introduce segmented rules and more granular performance analysis. Decisions may vary by zone, weight, service, or customer requirement, but updates remain periodic and manual. This stage improves fit while retaining much of the timing gap between observed performance and changed routing logic.
In the third stage, recommendations are generated dynamically using shipment context, cost, and performance evidence. Human operators or controlled workflows review the recommendations, creating a bridge between analytics and execution. Shadow mode and bounded pilots belong here.
In the fourth stage, defined shipment populations can be executed automatically within explicit policy and evidence boundaries. Actual outcomes feed back into monitoring, exceptions are governed, and authority can be reduced when data quality or network conditions deteriorate.
The maturity model is not a race toward maximum automation. An organization may intentionally keep certain shipment classes at an earlier stage because their risk, volume, or evidence quality does not justify automated authority. The objective is appropriate control for each decision domain.
This framework also creates a more precise investment roadmap. If an operation lacks reliable outcome data, the next investment may be measurement rather than modeling. If evidence is strong but exceptions are poorly understood, the next investment may be workflow governance. If both are mature, controlled automation becomes a more defensible step.
Implementation and Governance Considerations
The analytical case for shipment-level decisioning becomes operational only when the organization defines how recommendations enter production. The first requirement is a bounded decision domain. Teams should identify the shipment classes where multiple eligible options exist, the data needed to compare those options, and the conditions that make automated selection inappropriate.
A useful deployment pattern begins with shadow recommendations. The system evaluates shipments and records an alternative service choice while the existing routing logic continues to purchase the label. This produces direct evidence about where the proposed logic disagrees with current policy and gives operators an opportunity to identify undocumented constraints before any execution risk is introduced.
The shadow period should not be judged by recommendation volume. A high rate of changed decisions is not inherently good. The relevant question is whether the changed decisions are supported by better cost-and-service evidence. Review should therefore examine expected cost, expected delivery performance, eligibility, and the reason for the recommendation difference.
Once the evidence is credible, a limited production segment can be activated. The operating design should specify fallback behavior in advance. Missing data, service unavailability, unusual package characteristics, contractual exceptions, or rapidly changing network conditions may require the workflow to revert to an approved default or route the shipment for review.
Outcome read-back is the final control. Actual transportation cost and delivery performance should be attached to the original recommendation. This makes it possible to distinguish a recommendation that looked attractive at label time from one that actually improved the business outcome. It also creates the evidence needed to update performance assumptions when carrier behavior changes.
Economic Attribution
A rigorous economic analysis separates modeled opportunity from realized value. A lower quoted or expected rate is not realized savings until the shipment is executed under comparable conditions. Likewise, improved predicted service is not an achieved delivery outcome until the package arrives.
The baseline should be defined before the test. Teams should document the current routing logic, the shipment population, and material concurrent changes such as new contracts, packaging changes, fulfillment modifications, or shifts in delivery promise. Where these factors cannot be controlled, they should be disclosed in the analysis rather than absorbed into the claimed effect.
The addressable population also needs precision. Shipments with only one viable service option should not be included in estimates of allocation opportunity. The business case should focus on decisions where the organization could realistically choose among alternatives. This reduces headline opportunity but increases credibility.
The Luma AI Select case study is notable because the source explicitly describes major operating factors that were not changed during the reported measurement period [1]. That does not eliminate every attribution question, but it provides a clearer evidence boundary than a transformation where carrier mix, contracts, fulfillment, and decision logic all change simultaneously.
Decision Rights and Accountability
Transportation leadership should define approved carriers, services, and policy constraints. Operations should validate that the decision logic fits label-time workflows and exception handling. Finance should define the method for validating realized economic impact. Technology and data teams should maintain the reliability, freshness, and observability of the inputs used by the decision system.
Automation authority should be proportional to evidence. Common shipment profiles with stable data and repeatable outcomes can receive broader automated authority. Unusual or high-risk shipments can remain under human review. This tiered model avoids forcing the organization to choose between fully manual routing and unrestricted automation.
Override data should be retained. Repeated human corrections can reveal missing constraints or weak evidence. Rather than treating overrides only as resistance to automation, teams can use them as diagnostic signals that improve the decision model.
Executive Evaluation Criteria
Before expanding a shipment-level decision system, leadership should be able to verify four conditions. First, comparable shipments show improvement on the agreed cost-and-service objective. Second, exception and override patterns are understood. Third, actual outcomes are being read back against the recommendation. Fourth, fallback behavior is tested and available when data quality or network conditions deteriorate.
If those conditions are not yet met, the appropriate conclusion is that production readiness remains unproven. That is an evidence gap, not evidence that the approach cannot work. The next action is to extend the controlled evaluation until the missing proof is available.
This standard also provides a disciplined way to use external case studies. The reported Luma customer outcome can establish that shipment-level allocation produced material results in one documented environment [1]. It cannot establish the result another shipper will achieve. Local production decisions should be based on local baseline data, eligible alternatives, controlled comparisons, and realized outcomes.
Conclusion
Shipping AI creates value in the space between data and commitment.
Rates, contracts, and performance data define what is possible. The label-time decision determines what the business actually buys. Outcome read-back determines whether that choice was good.
For the customer in the supplied case study, improving this layer was associated with more than $2 million in annual savings and better on-time delivery without a carrier or fulfillment overhaul.
That does not make shipment-level AI universally successful. It makes the decision layer worthy of executive scrutiny.
Read the complete Luma AI Select case study
References
[1] EasyPost. Luma AI Case Study: $2M+ in Savings. 273,000 Fewer Late Deliveries. Without Changing Carriers. Primary campaign evidence. https://intenttechpub.com/POC/supply-chain-now/luma-ai-case-study.html
[2] PARCEL. Rethinking Parcel Diversification. January/February 2026 issue; published April 9, 2026. https://parcelindustry.com/article-6631-Rethinking-Parcel-Diversification.html
[3] TransImpact. Beyond the Big Two: When Parcel Carrier Diversification Makes Sense. May 11, 2026. https://transimpact.com/blog/beyond-the-big-two-when-parcel-carrier-diversification-makes-sense
[4] DCL Logistics. Understanding the Parcel Carrier Landscape and How to Diversify for Ecommerce Growth. July 20, 2026. https://dclcorp.com/blog/shipping/understanding-the-parcel-carrier-landscape-and-how-to-diversify-for-ecommerce-growth/
[5] HERE Technologies. HERE Technologies unveils Location Reasoning, redefining geospatial grounding for real-world AI decisions. May 19, 2026. https://www.here.com/about/press-releases/here-technologies-unveils-location-reasoning-redefining-geospatial-grounding-for-real-world-ai-decisions
[6] HERE Technologies. HERE unveils AI-powered last meter guidance solution to help delivery drivers complete the final handoff. May 14, 2026. https://www.here.com/about/press-releases/here-unveils-ai-powered-last-meter-guidance-solution-to-help-delivery-drivers-complete-the-final-handoff