Introduction
Parcel optimization is often approached as a network problem: negotiate better rates, add carriers, redesign packaging, change fulfillment, or shift volume. Those levers remain important, but they can overlook a high-frequency source of value, the decision made when each label is created.
A global recommerce marketplace in the supplied Luma AI Select case study processed more than 25,000 labels per day through EasyPost. Its primary carrier had an 80% on-time delivery rate, which the source equates to approximately 5,000 late packages per day at that volume.
The company deployed Luma AI Select to evaluate shipments individually at label creation. According to the case study, per-label costs declined 4–5%, annual savings exceeded $2 million, on-time delivery increased from 80% to 83%, and approximately 273,000 fewer late deliveries occurred annually [1].
The company did not add carriers, renegotiate contracts, change packaging, add headcount, or make significant fulfillment changes during the reported measurement period.
This guide explains the operating model behind that kind of decision improvement and gives executives a practical framework for evaluating parcel AI without relying on inflated claims or generic AI language.
Chapter 1 — Start With the Decision, Not the Technology
“Use AI to optimize shipping” is not an executable objective.
A useful AI initiative begins with a specific decision.
For parcel operations, one high-value decision is:
Which eligible carrier/service should be selected for this shipment at label creation?
That decision has clear inputs, a clear action, and a measurable outcome.
Inputs can include destination, package profile, delivery window, eligible services, cost, and historical performance.
Action is the selected label/service.
Outcome is the actual transportation cost and delivery performance.
This structure makes the initiative testable.
Executive principle:
If you cannot name the operational decision AI will change, you are not yet ready to measure its value.
Chapter 2 — Understand Decision Leakage
Decision leakage occurs when the organization has better options or better evidence available but the operational rule does not use them effectively.
Examples:
- A premium service is selected even though a lower-cost option is likely to meet the promise.
- A low-cost service is selected despite weak historical performance on the destination lane.
- A routing rule remains unchanged after carrier performance shifts.
- New carrier options are added, but most volume continues to follow the old static default.
- Analytics identify underperformance, but the finding is not connected to label creation.
Each error may be small. Repeated thousands of times per day, it can become material.
The Luma case study illustrates this scale effect. A 4–5% per-label cost reduction compounded into more than $2 million in annual savings. A three-point on-time improvement translated into approximately 273,000 fewer late deliveries annually, according to the source.
Executive principle:
Measure the frequency of the decision as carefully as the size of the improvement.
Chapter 3 — Separate Portfolio Design From Shipment Allocation
Carrier diversification and shipment-level optimization are complementary, not interchangeable.
Portfolio design asks:
Which carriers and services should we have available?
Shipment allocation asks:
Which available option should handle this shipment?
2026 parcel-market commentary increasingly treats carrier sourcing as portfolio design. PARCEL recommends maintaining enough optionality to rebalance volume as pricing and network conditions change. TransImpact highlights diversification for cost, leverage, and resilience. DCL Logistics argues that ecommerce operations should evaluate which carrier model fits each shipment rather than assume one provider is best for all orders.
The more options a shipper creates, the more important the allocation decision becomes.
Executive principle:
Diversification creates choices. Decision intelligence determines whether those choices create value.
Chapter 4 — Bring Evidence to Label Time
The value of data depends on timing.
A report generated after the label is purchased can explain the past but cannot change the shipment.
A decision system needs relevant evidence before the commitment.
The Luma case study says its shipment-level evaluation includes:
Destination zone.
- Package profile.
- Delivery window.
- Cost.
- Historical carrier performance.
It also states that the historical performance foundation includes more than one billion comparable shipments.
Executives should ask whether their own systems have the evidence required at the decision point — not merely whether the evidence exists somewhere in the enterprise.
Executive principle:
Decision-time availability is a separate capability from data availability.
Chapter 5 — Optimize Cost and Service Together
The parcel decision is often framed as cheap versus fast.
A better model is cost versus the probability of meeting the required customer promise.
The cheapest service can be economically poor if it produces avoidable late deliveries. The fastest service can be economically poor if a lower-cost option would have met the promise with sufficient reliability.
Possible optimization objectives include:
- Lowest expected cost subject to a minimum on-time threshold.
- Highest expected on-time probability below a cost ceiling.
- Weighted balance of cost, service, customer promise, and risk.
The objective should be explicit and approved.
The Luma customer’s reported results are notable because both cost and on-time performance improved. This should not be interpreted to mean the cost-service trade-off disappears. It suggests that inefficient baseline allocation can sometimes be improved on both dimensions.
Executive principle:
Do not optimize a parcel program on one metric while allowing another critical metric to deteriorate invisibly.
Chapter 6 — Design Human Control Correctly
Human-in-the-loop does not require a person to approve every label.
A scalable governance model separates policy from routine execution.
Humans should control:
- Eligible carriers and services.
- Contractual and regulatory constraints.
- Optimization objective.
- Customer-promise thresholds.
- Automation authority.
- Exception categories.
- Fallback logic.
- Incident suspension.
- Policy changes.
AI can handle:
- High-frequency comparison of eligible options.
- Consistent application of the approved objective.
- Routine selection within defined thresholds.
- Detection of unusual or low-confidence conditions for escalation.
Executive principle:
Put people around the policy and exceptions, not unnecessarily inside every routine transaction.
Chapter 7 — Build the Closed Loop
A recommendation is not an outcome.
After execution, the system should read back:
- Actual transportation cost.
- Actual delivery timing.
- On-time/late status.
- Exception events.
- Human overrides.
- Customer-impact signals where available.
This enables the organization to compare expected versus actual results and identify systematic errors.
The Luma case study states that EasyPost measured per-label cost and on-time performance before and after deployment across comparable U.S. shipments while carrier mix, contracts, and fulfillment operations remained unchanged.
That evidence boundary is important. If multiple operational variables change simultaneously, attribution becomes weaker.
Executive principle:
Never confuse a model recommendation with verified business impact.
Chapter 8 — Establish the Right KPI System
Primary economics
- Cost per label.
- Total transportation spend on comparable shipment populations.
- Savings versus prior decision logic.
Primary service
- On-time delivery rate.
- Late-delivery volume.
- Delivery performance by carrier/service/lane.
Decision quality
- Percentage of shipments where the recommended service differs from the default.
- Upgrade/downgrade frequency.
- Predicted versus actual delivery performance.
Governance
- Human override rate.
- Exception rate.
- Data-quality failure rate.
- Fallback activation rate.
Portfolio
- Carrier/service concentration.
- Allocation changes by lane and package profile.
Executive principle:
Use a balanced scorecard that makes cost, service, and risk visible at the same time.
Chapter 9 — Pilot Without Overclaiming
A credible pilot should produce evidence before it produces a transformation narrative.
Step 1 — Select one workflow
Choose a high-volume parcel flow with reliable cost and delivery data.
Step 2 — Capture the baseline
Measure current cost, on-time delivery, allocation mix, and exceptions.
Step 3 — Run shadow recommendations
Compare the new recommendation with the current rule without executing it.
Step 4 — Analyze decision differences
Identify where the new logic would upgrade, downgrade, or change service/carrier selection and why.
Step 5 — Deploy to a bounded segment
Use explicit authority and fallback rules.
Step 6 — Read back outcomes
Measure actual cost and delivery performance.
Step 7 — Compare comparable populations
Document carrier, pricing, fulfillment, packaging, customer-mix, and other changes that could affect the result.
Step 8 — Scale only on evidence
Do not extrapolate a narrow pilot beyond the population actually tested without additional validation.
Executive principle:
A small well-measured pilot is more valuable than a large deployment with ambiguous attribution.
Chapter 10 — 2026 Market Context
Parcel sourcing is becoming more dynamic.
PARCEL’s 2026 analysis recommends portfolio-style sourcing and the ability to rebalance carrier volume as conditions change.
TransImpact’s May 2026 guidance emphasizes diversification as a way to manage cost, negotiating leverage, and risk.
DCL Logistics’ July 2026 analysis encourages ecommerce shippers to evaluate carrier fit by shipment rather than search for a single best provider.
Operational AI is also moving closer to the action point. HERE Technologies announced Location Reasoning in May 2026 to ground AI agents in live location and road-network context for real-world decisions. Its separate last-meter guidance announcement applies AI to delivery execution.
These developments support a broader operating model: more optionality, more contextual evidence, more decision-time intelligence, and stronger outcome read-back.
Executive principle:
The competitive advantage is not owning more data. It is turning relevant data into better governed decisions faster.
Chapter 11 — Executive Readiness Scorecard
Score 1–3 for one defined parcel workflow.
Decision definition
1: Objective is broad or unclear.
2: Decision is known but inputs/authority are incomplete.
3: Decision, inputs, output, and outcome are explicit.
Data readiness
1: Aggregate or stale data dominates.
2: Some shipment-level evidence is available.
3: Relevant cost and performance evidence is available at decision time.
Carrier optionality
1: Highly concentrated with limited alternatives.
2: Multiple services/options exist but allocation is mostly static.
3: Eligible options are deliberately managed and can be rebalanced.
Decision logic
1: Static default.
2: Periodically updated rules.
3: Shipment-level evaluation against an explicit objective.
Governance
1: Informal.
2: Partial thresholds and review processes.
3: Authority, exception, fallback, override, and incident controls are explicit.
Measurement
1: Business impact inferred.
2: Before/after metrics exist with attribution gaps.
3: Comparable populations, concurrent changes, and actual outcomes are documented.
Read-back
1: Limited outcome feedback.
2: Delivery outcomes are reviewed retrospectively.
3: Actual outcomes continuously inform monitoring and decision quality.
Interpretation
7–10: Foundation stage. Focus on data, baseline, and decision definition.
11–16: Pilot-ready with targeted control gaps to close.
17–21: Strong candidate for governed shipment-level automation, subject to workflow-specific risk review.
This score is a diagnostic recommendation, not an externally validated industry benchmark.
Chapter 12 — The Luma AI Select Proof Point
Verified customer context from the supplied case study:
- Global recommerce marketplace.
- Customer name withheld at its request.
- 25,000+ labels per day through EasyPost.
- Primary carrier baseline on-time performance: 80%.
- Approximate late volume at baseline: 5,000 packages/day, as calculated in the source.
Verified reported deployment outcomes:
- 4–5% reduction in per-label cost.
- $2M+ annual savings.
- On-time delivery improved from 80% to 83%.
- Approximately 273,000 fewer late deliveries annually.
- 10× more U.S. shipping volume shifted to EasyPost after the customer observed the results. [1]
Verified operating conditions during the measured deployment:
- No new carriers.
- No contract renegotiation.
- No additional headcount.
- No packaging change.
- No significant fulfillment change.
Unknown / not supplied:
- Customer identity.
- Exact carrier names.
- Exact baseline dollar cost per label.
- Exact implementation duration.
- Exact model architecture.
- Expected result for another shipper.
The unknowns should remain unknown unless additional authoritative evidence is provided.
Chapter 13 — Executive Action Plan
Next 7 days
- Select the parcel workflow.
- Define the baseline.
- Map the label-time decision.
- Inventory current routing rules.
- Confirm cost and delivery data availability.
Next 30 days
- Build shipment-level performance views.
- Define the optimization objective.
- Establish automation and human-review boundaries.
- Run shadow recommendations.
Next 60–90 days
- Launch a bounded pilot.
- Read back actual outcomes.
- Compare comparable populations.
- Review exceptions and overrides.
- Scale only when evidence supports it.
Implementation Playbook
The move from static routing to shipment-level decisioning should be treated as an operating transformation, even when the physical network remains unchanged. The technology may enter at label creation, but successful adoption depends on policy, data, measurement, and human decision rights working together.
Start with one bounded workflow. Choose a parcel segment with sufficient volume, reliable cost and delivery data, and multiple eligible service options. Document the current rules and every known exception. The purpose is to understand the real decision system, including the informal workarounds operators use when the written rule does not fit the shipment.
Next, establish a baseline. Measure actual transportation cost and delivery performance for comparable shipments under the current logic. Separate shipments with meaningful alternatives from those where only one service is viable. This defines the addressable decision pool and prevents the program from overstating its potential reach.
Then build the evidence contract. Define which source is authoritative for service eligibility, cost, delivery outcome, package characteristics, and customer promise. Establish freshness expectations and fallback behavior when an input is unavailable. A recommendation should not execute simply because the system can produce one; the evidence required for execution should be explicit.
Shadow Mode Before Automation
Run the new decision logic beside the current production process before allowing it to purchase labels. For every disagreement, capture the recommended option, the existing default, the expected cost difference, the expected service difference, and the evidence behind the recommendation.
Review these disagreements with transportation and operations teams. Some will represent clear opportunities. Others will expose missing contractual rules, package constraints, service limitations, or operational realities that were not visible in the data. Shadow mode is therefore both a validation stage and a requirements-discovery stage.
The team should also monitor how often the proposed logic changes a decision. The objective is not to maximize change. If the current default is already appropriate for a shipment, retaining it is a valid outcome. Value comes from improving the decisions that are genuinely misallocated.
Controlled Execution
When shadow evidence is strong, activate the system on a limited production population. Define monitoring and rollback conditions before launch. Missing data, carrier disruption, unexpected service behavior, or a material increase in exceptions should be capable of constraining or suspending automated execution.
Track actual outcomes at shipment level. The key distinction is between predicted improvement and realized improvement. A recommendation that selects a cheaper service creates modeled opportunity. It becomes realized economic value only when the shipment executes and the business objective remains satisfied. The same discipline applies to predicted delivery performance.
A controlled rollout should also retain enough comparison evidence to detect whether external conditions changed during the test. New carrier rates, altered fulfillment processes, changes in package mix, or different customer promises can all affect results. These changes should be documented rather than silently attributed to the decision system.
Scaling the Decision Service
Expansion should occur by shipment population, not by executive enthusiasm. Each new population should have defined eligibility, sufficient data, understood exceptions, and a measurable baseline. A model that performs well for common domestic parcels may require different constraints for oversized, specialized, international, or otherwise unusual shipments.
As coverage grows, the operating review should monitor concentration as well as cost and service. A decision system may discover that one service is frequently optimal under current evidence. That can be economically rational, but leadership should understand the resulting dependency and how quickly the system can respond if performance changes.
Human overrides remain important at scale. Capture why operators reject a recommendation. Repeated patterns can reveal missing features, stale data, or a mismatch between the optimization objective and actual policy. Treating overrides as structured evidence makes the system more governable and improves future decisions.
Executive Governance Model
A mature program has clear accountability. Transportation leadership owns carrier and service policy. Operations owns workflow execution and exception handling. Finance owns the validation method for realized savings and economic impact. Technology and data teams own system reliability, evidence quality, and monitoring. Customer-experience stakeholders should inform the delivery promise and service-risk boundaries where appropriate.
Executives should review outcomes, not model activity. Useful questions include: Is the cost-and-service objective improving on comparable shipments? Are exception and override rates stable? Are actual outcomes matching expected performance? Have changes in carrier behavior altered the recommendation pattern? Is the fallback process working when data or services are unavailable?
The Luma AI Select case study provides a strong illustration of why this operating model deserves attention. The reported customer achieved material cost and delivery improvements through a focused change in label-time selection while major surrounding operating factors were reported as unchanged [1]. The exact results are customer-specific. The broader lesson is that a repeated operational decision can become a strategic lever when it is measurable, governed, and connected to execution.
A Reader’s Decision Checklist
Before approving a shipment-level optimization initiative, confirm that the organization can identify the addressable shipment population, establish a comparable baseline, define authoritative inputs, articulate the cost-and-service objective, specify automation guardrails, run shadow recommendations, measure actual outcomes, capture overrides, and execute a tested fallback path.
If any of those elements is missing, the next step is not to assume the initiative will fail. It is to close the specific evidence or control gap. That approach keeps implementation grounded in operational proof and prevents a compelling technology story from outrunning production readiness.
Conclusion
Parcel optimization is entering a decision-centric era.
Carrier contracts determine what can be purchased. Data explains what has happened. Shipment-level intelligence determines what the organization buys for the next package.
The supplied Luma AI Select case study shows the potential leverage of that decision at scale: more than $2 million in reported annual savings and improved on-time performance without adding carriers or significantly changing fulfillment.
The executive mandate is not “deploy AI.” It is more disciplined:
Choose a high-value decision.
Bring trusted evidence to the decision moment.
Define the objective.
Govern the authority.
Execute.
Read back the outcome.
Scale only on evidence.
Start with the proof, then use the guide to diagnose your own decision leakage.
Read the full 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