Shipping rules are often designed for operational consistency. Choose a carrier for a zone. Choose a service for a weight band. Upgrade when a promised date is close. Review the rules periodically and repeat.
That model works - until the environment changes faster than the rules.
At high parcel volume, static routing can turn small decision errors into large annual costs. The issue is not necessarily that the contracted carriers are wrong. It may be that the same default is being applied to shipments with different destinations, package profiles, delivery windows, costs, and performance probabilities.
The supplied Luma AI Select case study offers a useful example. A global recommerce marketplace was processing more than 25,000 labels per day through EasyPost. Its primary carrier had an 80% on-time delivery rate. At that daily volume, the case study estimates that approximately 5,000 packages were arriving late each day [1].
The company did not solve the problem by adding another carrier or rebuilding fulfillment. Instead, it changed how each shipment was assigned at label creation.
After deploying Luma AI Select, the case study reports a 4-5% reduction in per-label costs, more than $2 million in annual savings, an increase in on-time delivery from 80% to 83%, and approximately 273,000 fewer late deliveries annually. Carrier mix, contracts, packaging, headcount, and fulfillment operations were not changed as part of the measured deployment. Results were evaluated across comparable U.S. shipments before and after implementation [1].
Those are customer-specific reported results, not a guarantee. But they expose a broader operational question: what is the cost of a shipping rule that is directionally right on average but wrong for a meaningful share of individual shipments?
Static Rules Solve Yesterday's Problem
A static shipping rule is a compressed version of past experience. Someone determines that a carrier or service generally performs well under certain conditions and turns that conclusion into a repeatable default.
The problem is not the existence of rules. The problem is the assumption that the conditions behind the rule remain stable.
Parcel networks are dynamic. Service performance changes by lane. Pricing and surcharges change. Package mix changes. Customer promises change. Carrier capacity changes. A service that is appropriate for one destination may be poor for another. A premium option may be unnecessary for a shipment with sufficient delivery slack. A low-cost service may look attractive until its actual on-time probability is considered.
Static rules update periodically. Shipments arrive continuously.
That timing mismatch creates decision leakage.
Why the Unit Economics Matter
Consider a high-volume operation processing 25,000 labels every day. A one-dollar decision error does not remain a one-dollar problem. Repeated across thousands of shipments, it becomes a material operating expense.
The same is true for service performance. A one-percentage-point change in on-time delivery can represent hundreds of shipments at scale.
This is why parcel optimization should be evaluated at two levels simultaneously:
Unit level: Was the right service selected for this shipment?
Portfolio level: What happens when that decision is repeated across millions of shipments?
The Luma case study illustrates the portfolio effect. A 4-5% reduction in per-label cost sounds incremental until it compounds into more than $2 million in annual savings. A three-point on-time improvement sounds modest until the operation calculates approximately 273,000 fewer late deliveries per year.
Scale changes the economics of decision quality.
The Real Optimization Problem
Many parcel programs focus on procurement: negotiate a better rate, add another carrier, or increase leverage by shifting volume. Those are important levers, and current 2026 industry commentary continues to emphasize diversification.
PARCEL's 2026 analysis recommends treating parcel sourcing more like portfolio design, with the ability to rebalance volume as conditions change. TransImpact argues that diversification can reduce costs and risk while improving negotiating leverage. DCL Logistics writes that the more useful ecommerce question is increasingly not "Which carrier is best" but "Which carrier model fits each shipment"
But procurement determines the options available. Allocation determines which option is actually used.
A shipper can negotiate excellent rates with multiple carriers and still leak value if the allocation logic is static. Conversely, improving shipment-level selection can sometimes create value without changing the underlying carrier contracts.
That is precisely what makes the Luma example interesting.
What Shipment-Level Decisioning Changes
The case study says Luma AI Select evaluates each shipment using factors including destination zone, package profile, delivery window, cost, and historical performance across more than one billion comparable shipments.
This approach changes the decision from a lookup rule into a ranking problem.
Instead of:
"If zone = X, use service Y."
The system can ask:
"Among the eligible services for this specific shipment, which option best balances cost and expected on-time performance"
That is a more information-rich decision.
It also creates a stronger feedback loop. After the package is delivered, the organization can compare the expected outcome with the actual outcome and update its view of service performance.
The result is not simply more data. It is data attached to a decision and then tested against reality.
Why the Label-Creation Moment Matters
Shipping analytics often arrive after the operational decision has already been made. A weekly dashboard may show that a service underperformed, but it cannot change the labels purchased earlier in the week.
Shipment-level AI is useful when it operates close enough to the transaction to change the outcome.
At label creation, the shipment is known, the eligible options can be evaluated, and the organization is about to commit to a service. That makes it a natural decision point.
This "decision-time context" principle is becoming more visible in 2026 AI infrastructure. HERE Technologies introduced 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 applied AI to the final delivery handoff. The common idea is that AI becomes operational when context is available at the point where an action can still change the result.
For parcel shipping, that point is often the label.
How to Audit Your Static Shipping Rules
Start with one high-volume workflow and examine the logic that produces the label.
Step 1: Inventory the rules
Document which carrier and service defaults are triggered by destination, zone, weight, package type, delivery promise, customer segment, and exception.
Step 2: Identify stale assumptions
For each rule, record when it was created, when it was last validated, and what cost/performance evidence supports it today.
Step 3: Segment actual outcomes
Compare transportation cost and on-time delivery by carrier, service, lane, zone, package profile, and promised window. Avoid relying only on national averages.
Step 4: Find high-volume decision gaps
Look for segments where the default service is consistently more expensive than an eligible alternative with comparable performance, or where a low-cost default repeatedly misses the delivery promise.
Step 5: Test before automating
Run recommendations in shadow mode or on a controlled traffic segment. Compare the recommendation against the existing rule before giving the system execution authority.
Step 6: Read back actual outcomes
Measure cost, on-time delivery, exceptions, overrides, and downstream customer impact after each decision.
Step 7: Govern the exceptions
Define what happens when data are missing, a carrier becomes unavailable, an unusual shipment appears, or the model confidence is insufficient.
The objective is not to remove human control. It is to reserve human attention for decisions where judgment is actually needed.
What to Measure
A credible parcel-decision program should monitor:
• Cost per label and cost per shipment.
• On-time delivery rate.
• Late deliveries avoided versus baseline.
• Percentage of shipments where the recommended service differs from the static default.
• Savings on comparable shipments.
• Carrier/service concentration.
• Upgrade and downgrade frequency.
• Exception and human-override rates.
• Data freshness and missing-data rates.
• Actual versus expected service performance.
Avoid claiming savings simply because a cheaper rate existed. The selected service must also satisfy the operational objective.
The Leadership Takeaway
High-volume parcel operations can spend significant effort optimizing contracts while leaving the allocation layer comparatively static.
The Luma AI Select case study suggests a different sequence: before assuming the network must change, test whether the current network is being used intelligently at shipment level.
For the customer in the case study, changing the label-time decision was associated with more than $2 million in annual savings and better delivery performance without adding carriers or making significant fulfillment changes.
The broader lesson is simple: when a decision is repeated 25,000+ times per day, improving the decision itself can become a major operating lever.
Decision Quality as an Operating Metric
Most parcel scorecards emphasize spend and carrier performance after the fact. A shipment-level program adds another useful metric: decision quality. The purpose is to understand whether the information available before label purchase is being converted into the best available choice under the organization's policy.
Decision quality can be reviewed through disagreement analysis. When an alternative recommendation differs from the existing rule, classify the reason: lower cost with equivalent service expectation, stronger delivery probability within an acceptable cost boundary, changed lane performance, different package fit, or another approved factor. This creates a transparent view of what the decision system is actually changing.
The organization can then inspect whether those changed decisions perform as expected. If one category repeatedly produces value, it may deserve broader authority. If another category creates exceptions or unstable outcomes, it can remain in shadow mode while the evidence or guardrails improve.
This approach avoids a false choice between static simplicity and opaque automation. The system can remain explainable because each changed decision is connected to a business reason, an evidence set, and an observed outcome.
It also creates a more useful management conversation. Instead of asking whether the AI is accurate in the abstract, leaders can ask whether specific categories of decisions improve cost and service, whether the evidence remains current, and whether the operating controls are sufficient for the authority being granted.
Where Disconnected Decisions Show Up in the P&L
Decision leakage rarely appears as one clean line item. It is distributed across transportation spend, premium-service usage, late-delivery handling, support activity, refunds or concessions where applicable, and the management time required to investigate recurring exceptions. That fragmentation is one reason static allocation can survive for years: each individual decision looks small, while the aggregate consequence is difficult to isolate.
A useful audit therefore connects operational decisions to economic outcomes. Begin with transportation spend, but do not stop there. Identify shipments where the selected service materially exceeded the speed required by the customer promise. Separately identify shipments where the lowest-cost default repeatedly failed to meet the required delivery window. These two populations represent different forms of leakage: unnecessary service spend and under-purchased reliability.
The analysis should also distinguish controllable allocation decisions from structural constraints. A shipment may have only one eligible service because of destination, dimensions, contractual limitations, hazardous-material rules, pickup schedules, or other operational requirements. Those shipments are not decision opportunities. The useful population is the set where two or more viable options existed and the selected option can be compared with credible alternatives.
That distinction improves the business case. Instead of claiming that every parcel is optimizable, leaders can quantify the addressable decision pool, measure how often the current default appears suboptimal, and estimate the value of improving only those decisions. The resulting opportunity is more defensible because it respects eligibility and operating constraints.
Building a Better Feedback Loop
Static rules become less risky when they are continuously challenged by outcome evidence. For each service decision, capture what was expected at label creation and what actually happened after delivery. Over time, this creates a lane- and service-specific evidence base that can reveal where historical assumptions remain valid and where they are degrading.
The feedback loop should include more than carrier performance. It should track the freshness of cost inputs, changes in service eligibility, exception frequency, and human overrides. Overrides are especially useful evidence. If operators repeatedly reject recommendations for the same reason, the system may be missing an operational constraint. If overrides are rare but concentrated in a small number of shipment profiles, those profiles can be governed explicitly instead of weakening the entire automation model.
Finance also has a role. Savings claims should be tied to comparable shipments and measured against an agreed baseline. A lower selected rate is not automatically a realized saving if it causes a service failure that triggers downstream cost. Likewise, a more reliable service is not automatically an improvement if it buys performance the customer promise did not require. The economic objective needs to reflect the combined operating outcome.
What a Controlled Pilot Should Prove
A controlled pilot should answer a limited set of questions before expansion. First, can the recommendation logic identify eligible alternatives consistently? Second, does it improve the cost-and-service objective on comparable shipments? Third, are the reasons for changed selections understandable to operators? Fourth, are exceptions predictable enough to govern? Fifth, can actual delivery outcomes be connected back to the original recommendation?
A strong pilot also defines stopping conditions in advance. If data quality falls below an agreed threshold, a required service becomes unavailable, or observed performance diverges materially from the expected pattern, the workflow should fall back to approved logic. This is not a weakness in automation. It is a control that protects the operation while evidence develops.
The case-study evidence makes the scale opportunity visible, but the internal decision to automate should still be based on local proof. The reported Luma results show what happened for one high-volume customer under the documented measurement conditions [1]. Your own program should reproduce the discipline of the measurement approach rather than assuming the magnitude of the outcome.
A final operating discipline is to assign ownership for the decision loop. Transportation should own approved carrier and service policy; operations should own exceptions and workflow fit; finance should validate realized economics; and data or technology teams should maintain input quality and monitoring. That shared ownership keeps shipment-level optimization connected to business outcomes rather than allowing it to become an isolated model exercise.
Read the complete case study
If your operation still routes high-volume parcels with fixed defaults, use this case study as a benchmark for the questions to test-not as a guaranteed outcome.
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
