Executive Summary
Parcel networks have become more complex while many shipping decisions remain governed by static rules. Carrier portfolios may expand, data may improve, and dashboards may become more sophisticated, yet the final carrier/service choice can still depend on logic that changes only periodically.
This creates a decision gap.
The supplied Luma AI Select case study demonstrates the potential value of closing that gap. A global recommerce marketplace processing more than 25,000 labels per day through EasyPost deployed shipment-level AI service selection at label creation. The case study reports 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 per year [1].
The customer did not add carriers, renegotiate contracts, change packaging, add headcount, or make significant fulfillment changes as part of the measured deployment. Results were measured before and after deployment across comparable U.S. shipments while carrier mix, contracts, and fulfillment operations remained unchanged.
This whitepaper translates that proof point into a practical framework for operations, logistics, supply-chain, transportation, warehouse, IT, engineering, fulfillment, inventory, product, and consulting leaders evaluating shipment-level decision intelligence.
The objective is not to prescribe one vendor, carrier portfolio, or algorithm. It is to define the operating system required to move from static routing to governed, measurable, shipment-level decisions.
1. The Parcel Decision Problem
Every parcel label represents a choice among available services. That choice determines immediate transportation cost and influences the probability of meeting the delivery promise.
Traditional routing systems often simplify the decision using rules such as:
- Use Carrier A for Zone 2-5.
- Use Service B below a weight threshold.
- Upgrade when the delivery window is under a defined number of days.
- Route a category of orders to one default service.
These rules create consistency, but they also aggregate shipments that may differ in economically important ways.
A destination lane can perform differently from the national average. A package profile can trigger different cost economics. A premium service may not improve the expected outcome enough to justify its price. A low-cost service may carry a late-delivery probability that makes it unattractive for a particular promise.
The operating challenge is therefore not merely selecting a carrier. It is selecting the most appropriate eligible service for each shipment using the evidence available at decision time.
2. Why Static Rules Become Expensive at Scale
Static routing can work well when conditions are stable and shipment volume is modest. At high volume, however, small allocation errors compound.
The Luma case-study customer processed more than 25,000 labels per day. Its primary carrier had an 80% on-time delivery rate, which the source equates to approximately 5,000 late packages each day.
At this scale:
- A small per-label cost difference can become millions of dollars annually.
- A few percentage points of on-time improvement can represent hundreds of shipments per day.
- A rule that is wrong for a minority of shipments can still affect a very large absolute number of packages.
The economics of shipment-level intelligence are therefore driven by repetition. The decision is small; the decision volume is not.
3. The Case Study: What Changed
The Luma AI Select deployment changed the decision made at label creation.
According to the supplied case study, the model evaluated each shipment individually using factors including:
- Destination zone.
- Package profile.
- Delivery window.
- Cost.
- Historical carrier performance across more than one billion comparable shipments.
The system selected the service level that best balanced cost and on-time performance for each shipment.
Reported customer outcomes:
- $2M+ annual savings.
- 4-5% reduction in per-label costs.
- On-time delivery increased 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]
What did not change during the measured deployment:
- Carrier mix.
- Carrier contracts.
- Packaging.
- Headcount.
- Significant fulfillment operations.
The case study reports that results were measured using before-and-after comparisons across comparable U.S. shipments.
These are customer-specific results and should not be treated as guaranteed outcomes for other organizations.
4. Carrier Diversification vs. Carrier Intelligence
Diversification and decision intelligence solve different problems.
Diversification expands the option set. It can reduce concentration risk, create negotiating leverage, and give shippers alternatives when pricing or network performance changes.
Decision intelligence determines how those options are used.
2026 parcel commentary increasingly frames sourcing as a portfolio problem. PARCEL's January/February 2026 analysis argues for the ability to rebalance volume as pricing and network conditions evolve. TransImpact's May 2026 analysis describes diversification as a cost, leverage, and risk-management strategy. DCL Logistics' July 2026 guidance encourages ecommerce operators to evaluate which carrier model fits each shipment.
These perspectives support a two-layer architecture:
Layer 1: Portfolio design
Establish credible carriers and services with appropriate contracts, capacity, coverage, and risk characteristics.
Layer 2: Shipment allocation
Select the most appropriate eligible option for each shipment based on current context and evidence.
A shipper can be strong in Layer 1 and weak in Layer 2. That is the hidden decision gap.
5. The Shipment-Level Intelligence Framework
A production-ready decision system should include seven connected layers.
Layer 1 - Shipment Context
Required inputs may include destination, origin, package dimensions, weight, package type, delivery promise, customer requirement, service restrictions, and special handling.
Principle: do not ask the model to infer facts the shipping system already knows.
Control question: Is the shipment context complete and current enough to support a decision?
Layer 2 - Eligible Option Set
The system should only evaluate services that are operationally and contractually valid.
Eligibility may reflect:
- Carrier contract.
- Geographic coverage.
- Package limitations.
- Delivery-service availability.
- Customer restrictions.
- Operational cutoffs.
- Regulatory or product constraints.
Principle: optimization inside the wrong option set produces a confidently wrong answer.
Layer 3 - Cost Evidence
The decision needs shipment-specific cost evidence at the level available to the organization.
Relevant components can include:
- Transportation charge.
- Service level.
- Zone.
- Dimensional effects.
- Applicable surcharges.
- Contracted pricing.
Principle: avoid optimizing against a broad average when the transaction can be priced more precisely.
Layer 4 - Performance Evidence
Carrier performance should be evaluated at the most relevant level supported by trustworthy data.
Possible dimensions:
- Carrier.
- Service level.
- Lane.
- Zone.
- Destination region.
- Package profile.
- Time period.
Principle: aggregate performance can hide local underperformance.
Layer 5 - Objective Function
The business must define what the system is optimizing.
Examples:
- Lowest expected cost subject to a minimum on-time threshold.
- Highest expected on-time probability under a cost ceiling.
- Weighted balance of cost, delivery probability, customer promise, and risk.
Principle: "optimize shipping" is not an objective. The trade-offs must be explicit.
Layer 6 - Governance and Authority
The organization should define:
- Which shipment categories may execute automatically.
- Confidence/evidence-quality thresholds.
- Human-review triggers.
- Fallback logic.
- Override permissions.
- Incident suspension procedures.
- Policy-change approval.
Principle: human oversight should govern policy and exceptions, not manually approve every routine transaction.
Layer 7 - Outcome Read-Back
Every decision should generate evidence after execution.
Track:
- Actual cost.
- Actual delivery date/time.
- On-time outcome.
- Exceptions.
- Human override.
- Customer-impact signals where available.
Principle: a decision system that cannot compare prediction with outcome cannot demonstrate improvement.
6. Measurement Architecture
A shipment-level intelligence program should establish a baseline before claiming value.
Baseline metrics:
- Cost per label.
- On-time delivery rate.
- Late-delivery volume.
- Carrier/service allocation.
- Upgrade/downgrade frequency.
- Exception rate.
- Relevant customer-impact metrics.
Post-deployment metrics:
- Cost per label on comparable shipments.
- On-time delivery on comparable shipments.
- Savings versus prior logic.
- Late deliveries avoided versus baseline.
- Percentage of decisions changed from default.
- Actual versus predicted outcome.
- Human override rate.
Evidence discipline:
Do not attribute every business change to the model. Document concurrent changes in carrier mix, pricing, packaging, fulfillment, customer mix, geography, and demand where they could affect the result.
The Luma case study is useful because it explicitly states that carrier mix, contracts, and fulfillment operations remained unchanged during its reported comparison period.
7. Human-in-the-Loop Design
Human oversight should be designed around risk.
Routine/high-confidence decision:
System executes within approved policy.
Unusual shipment or low-quality evidence:
System routes to human review or safe fallback.
Policy change:
Human approval required.
Carrier/network disruption:
Human incident owner may suspend automation or change eligible options.
Repeated override pattern:
Operations reviews whether the model, data, or policy requires adjustment.
The objective is governed automation, not maximum automation.
8. Implementation Roadmap
Phase 1 - Diagnose
- Select one high-volume parcel workflow.
- Inventory current routing rules.
- Capture baseline cost and on-time performance.
- Identify the variables that materially change the decision.
- Document data gaps.
Phase 2 - Shadow
- Generate shipment-level recommendations without executing them.
- Compare recommendations with current rules.
- Estimate where and why decisions differ.
- Validate cost and performance evidence.
Phase 3 - Controlled Pilot
- Apply the new decision logic to a bounded shipment segment.
- Preserve a comparable baseline or control where practical.
- Monitor exceptions and overrides daily.
- Do not generalize early results beyond the tested population.
Phase 4 - Governed Scale
- Expand only after evidence meets predefined thresholds.
- Automate routine decisions inside approved authority.
- Keep human review for exceptions.
- Monitor carrier/service concentration and unintended effects.
Phase 5 - Continuous Read-Back
- Recalculate service performance.
- Investigate persistent prediction errors.
- Review overrides.
- Refresh policies as business objectives change.
9. Executive Readiness Assessment
Score each category from 1 to 3.
Decision context
1: Static defaults dominate.
2: Shipment context exists but is not fully connected to selection.
3: Relevant context is available at label creation.
Cost evidence
1: Broad averages or rate cards.
2: Shipment costs available retrospectively.
3: Decision-time cost evidence is available for eligible options.
Performance evidence
1: Aggregate carrier scorecards.
2: Service/lane views exist for selected areas.
3: Granular performance is continuously read back into decisioning.
Objective
1: Cheapest or default service.
2: Multiple metrics reviewed manually.
3: Explicit cost/service objective drives the decision.
Governance
1: Informal overrides.
2: Some thresholds and approvals documented.
3: Automated authority, human-review triggers, fallback, and incident controls are explicit.
Measurement
1: Savings inferred.
2: Before/after metrics exist but concurrent changes are unclear.
3: Comparable populations and relevant operational changes are documented.
A low score does not mean AI is inappropriate. It identifies the operating controls that should be built before scaling automation.
10. 2026 Market Context
Three current developments strengthen the case for a more dynamic parcel decision layer.
First, carrier diversification remains a strategic theme. PARCEL and TransImpact both emphasize reducing overdependence and creating the ability to shift volume as conditions change.
Second, ecommerce parcel strategy is becoming more shipment-specific. DCL Logistics' July 2026 analysis explicitly frames carrier selection around fit by shipment rather than a universal winner.
Third, operational AI is moving toward context-grounded decisions. HERE Technologies' May 2026 Location Reasoning announcement focuses on grounding AI in live location and road-network intelligence for real-world decisions.
Together, these trends point toward a parcel operating model where options are diversified, decisions are contextual, execution is governed, and outcomes are continuously measured.
11. Executive Action Plan
For the next 30 days:
- Identify the highest-volume parcel decision workflow.
- Quantify its current cost and on-time baseline.
- Map the static rules currently making the decision.
- Identify where current cost/performance evidence could change those rules.
- Define one explicit optimization objective.
- Establish automation and human-review boundaries.
- Design a shadow test before production execution.
For the next 60-90 days:
- Run a bounded pilot.
- Compare comparable shipments.
- Measure cost and service together.
- Investigate overrides and exceptions.
- Scale only when the evidence supports the expansion.
Conclusion
The next stage of parcel optimization is not simply more carrier data, more dashboards, or more carriers. It is a better decision at the point where the business commits to a shipment.
The Luma AI Select case study shows why that matters. For one high-volume recommerce marketplace, changing shipment-level service selection was associated with $2M+ annual savings and improved on-time delivery without a major network or fulfillment redesign.
The transferable lesson is not the exact result. It is the operating model: define the decision, bring relevant evidence to the decision moment, govern execution, and read back the outcome.
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
