Executive Brief
When parcel costs rise or delivery performance slips, the standard response is often to renegotiate rates, add another carrier, change packaging, or redesign fulfillment. Those levers matter. But they can overlook a less disruptive question: are we selecting the right available service for each shipment at the moment the label is created?
The supplied Luma AI Select case study makes that question tangible. A global recommerce marketplace processing more than 25,000 labels per day through EasyPost reported more than $2 million in annual savings after deploying shipment-level AI service selection. Per-label cost fell 4-5%, on-time delivery improved from 80% to 83%, and the company calculated approximately 273,000 fewer late deliveries per year [1]. According to the case study, those results were achieved without adding carriers, renegotiating contracts, changing packaging, adding headcount, or making significant changes to fulfillment operations.
The implication for operations leaders is important: sometimes the constraint is not the carrier network. It is the logic used to allocate shipments across the options already available.
Why Static Allocation Creates Leakage
A static routing rule compresses many variables into a simple default. It may specify one carrier by zone, one service by weight, or one upgrade rule by delivery window. Simplicity is useful, but it creates a hidden assumption that shipments sharing one characteristic should receive the same decision.
In practice, parcel economics and performance are multidimensional. Destination, package profile, delivery commitment, carrier service, lane performance, historical reliability, and cost can interact differently on every shipment.
The Luma AI Select case study describes a model that evaluates shipments individually using destination zone, package profile, delivery window, cost, and historical carrier performance drawn from more than one billion comparable shipments. The objective is not simply to pick the lowest published rate. It is to identify the option that best balances expected cost and on-time delivery for that shipment.
This changes the operating model from periodic rule maintenance to continuous decisioning.
What Changed, and What Did Not
The most useful part of the case study is the evidence boundary.
What changed:
• Service selection at label creation.
• Shipment-level evaluation of cost and delivery performance.
• The amount of U.S. shipping volume flowing through EasyPost after results were observed.
What did not change during the reported measurement period:
• Carrier mix.
• Carrier contracts.
• Fulfillment operations.
• Packaging.
• Headcount.
The case study states that EasyPost compared per-label cost and on-time performance before and after Luma AI deployment across comparable U.S. shipments while the above operating conditions remained unchanged.
This matters because transformation stories often bundle multiple changes together and then attribute the entire outcome to one technology. Here, the reported intervention was narrower: improve the decision made for each label.
The Scale Effect
The case-study customer was processing more than 25,000 labels per day. Its primary carrier had an 80% on-time delivery rate, which the source translates into approximately 5,000 late packages per day at that volume.
At this scale, small percentage changes are operationally large.
A 4–5% improvement in per-label cost can compound into seven-figure annual savings. A three-percentage-point improvement in on-time delivery can remove hundreds of late shipments from the daily flow. That is why parcel optimization should not be dismissed as incremental simply because the decision occurs one label at a time.
The unit of optimization is small. The aggregate consequence is not.
A Better Executive Question
The conventional carrier-management question is often: Which carrier should receive more volume?
A more precise question is: Which eligible service should receive this shipment, given its destination, package profile, delivery promise, cost, and expected performance?
That distinction separates portfolio strategy from allocation intelligence.
Current 2026 parcel-market commentary supports this shift. PARCEL’s January/February 2026 analysis describes parcel sourcing as increasingly similar to portfolio design, with shippers needing the ability to rebalance volume as pricing and network conditions change. TransImpact’s May 2026 analysis similarly argues that high-volume parcel shippers are diversifying to manage cost, leverage, and risk. DCL Logistics wrote in July 2026 that ecommerce shippers are increasingly asking which carrier model fits each shipment rather than searching for a single universally best provider.
Diversification creates optionality. Shipment-level decisioning is what converts optionality into a repeatable operating choice.
Where AI Belongs
AI creates the most credible operational value when it is attached to a decision with clear inputs, an executable output, and measurable read-back.
For parcel shipping, label creation meets those conditions:
Input: shipment characteristics, eligible services, cost, delivery requirement, and historical performance.
Decision: select the service/carrier option.
Action: purchase the label and tender the shipment.
Read-back: actual transportation cost and delivery outcome.
That closed loop is more operationally meaningful than adding another dashboard that only explains yesterday’s performance.
The same principle is appearing elsewhere in 2026 logistics technology. HERE Technologies announced Location Reasoning in May 2026 to ground AI agents in live location and road intelligence for real-world decisions, and separately introduced AI-powered last-meter guidance for delivery drivers. The broader pattern is that operational AI is moving closer to the moment where a decision can change an outcome.
A Practical Decision-Control Model
Leaders can evaluate shipment-level AI through five controls.
1. Eligibility control
Which carriers and services can legally, contractually, and operationally handle the shipment?
2. Evidence control
Which cost and performance data are sufficiently current and granular to influence the decision?
3. Objective control
What is being optimized: transportation cost, on-time probability, customer promise, or a weighted combination?
4. Authority control
Which recommendations can be executed automatically, and which require human review or fallback logic?
5. Outcome control
Is actual cost and delivery performance read back against the recommendation so the organization can test whether the decision improved results?
Without these controls, “AI optimization” can become a black box. With them, it becomes an auditable operating system.
Five Tests for Your Current Shipping Logic
1. Can your team explain why a specific service was selected for a specific shipment?
2. Does the decision use current shipment-level cost and lane/service performance, or mostly static defaults?
3. How quickly can routing change when service performance deteriorates?
4. Can you quantify the cost and late-delivery impact of your current default rules?
5. Do you compare recommended versus actual outcomes after shipment delivery?
If those questions are difficult to answer, the next parcel initiative may be less about adding options and more about using existing options intelligently.
Intent Amplify Perspective
The Luma AI Select story is compelling because it reframes parcel optimization as a decision problem rather than solely a procurement problem.
The customer’s reported outcomes- $2M+ annual savings, 4–5% lower per-label costs, an on-time improvement from 80% to 83%, and approximately 273,000 fewer late deliveries per year, came from changing shipment-level selection rather than rebuilding the physical operation.
Those results should not be generalized as guaranteed outcomes. They should be used as evidence that, at sufficient scale, improving a repeatable operational decision can materially affect both cost and service.
The Finance and Operations Handshake
Shipment-level optimization crosses functional boundaries, so the value definition needs to be shared. Transportation may focus on rate and carrier performance. Operations may focus on workflow stability and exceptions. Finance needs a method for distinguishing realized savings from modeled opportunity. Customer-facing teams care about whether the delivery promise is protected.
A common scorecard prevents these functions from optimizing against different definitions of success. For a controlled shipment population, report realized transportation cost, on-time delivery, late-delivery rate, recommendation-change rate, exceptions, overrides, and relevant concentration. Where a recommendation was not executed, keep its expected value in a separate opportunity view rather than blending it with achieved results.
The same discipline applies to service improvements. A predicted on-time probability is not an achieved outcome. Read back the actual delivery event and compare it with the expectation. Over time, this creates the evidence required to determine whether the decision logic is calibrated to current network behavior.
Governance should also define who can change the objective or guardrails. A technical team should not silently alter the trade-off between cost and service. Material policy changes belong with the business owners who are accountable for transportation economics and customer commitments.
Finally, review whether the program is changing the carrier portfolio in unexpected ways. Shipment-level optimization can shift volume even when no explicit sourcing decision has been made. That may be beneficial, but procurement and transportation leadership should understand the resulting concentration, contractual implications, and resilience profile.
This cross-functional handshake turns a promising recommendation engine into a controlled operating capability. The technology makes the decision; the organization remains accountable for the objective, evidence, authority, and measured outcome.
Operating Playbook: From Audit to Controlled Deployment
A practical program can begin without a network redesign. The first phase is observation. Select a parcel workflow with enough volume to reveal patterns and capture the decision inputs that exist at label creation: destination, package profile, delivery promise, eligible services, shipment-specific cost, and recent performance evidence. Map which of those inputs actually affect the current routing rule and which are visible but ignored.
The second phase is baseline construction. Establish the cost and service performance of the current logic using comparable shipments. Avoid mixing changes in carrier contracts, fulfillment configuration, packaging, or customer promise into the same test when possible. The goal is to create a baseline that makes the effect of a decision change easier to interpret rather than attributing every movement in cost or service to the new system.
The third phase is shadow decisioning. Generate an alternative service recommendation without changing the purchased label. Record when the recommendation differs from the default, why it differs, the expected cost difference, and the expected delivery implication. This gives operations, finance, and transportation teams a reviewable set of decisions rather than asking them to approve an abstract AI initiative.
The fourth phase is controlled execution. Once the recommendation logic has demonstrated sufficient quality, introduce it on a bounded shipment segment with explicit eligibility rules and fallback behavior. Define which conditions force the workflow back to the approved default: missing data, unavailable services, unusual package characteristics, contractual restrictions, or low-confidence evidence. Automation should expand only when the exception model is understood.
The fifth phase is read-back. Every executed recommendation should ultimately be connected to actual transportation cost and delivery outcome. This closes the loop between prediction and reality. It also prevents a common failure mode in operational AI: celebrating the recommendation while never testing whether the downstream result improved.
For executive governance, review four questions on a fixed cadence. Did the system improve the combined cost-and-service objective on comparable shipments? Are exceptions and overrides concentrated in identifiable scenarios? Has the underlying carrier or service performance changed enough to alter the decision logic? And can the team explain why the system selected an option when the recommendation differs from the historical default?
This playbook keeps the scope deliberately narrow. Procurement can continue to manage carrier economics and network strategy. Fulfillment can continue to manage throughput and physical execution. Shipment-level decisioning sits between them, using the options procurement has created to make a better transaction-level choice before the label is purchased.
The Luma AI Select case study provides evidence that this layer can matter at scale, but the correct internal standard is still local validation. Use the reported customer outcome as a benchmark for what to measure, not as a forecast. A production decision should depend on your own comparable-shipment evidence, operational constraints, exception rates, and observed cost and delivery results.
How Allocation Leakage Appears in Daily Operations
Allocation leakage is rarely visible as one obvious failure. It usually appears as a collection of small inconsistencies: a premium service used where a standard option would likely have met the promise, a low-cost service retained on a lane where delivery reliability has weakened, or a static rule that ignores a package or destination characteristic that materially changes the economics. Because each shipment still moves, the process can look operationally healthy even while value is leaking at scale.
For operations leaders, the diagnostic is to look for repeated disagreement between policy and evidence. Start with high-volume shipment segments and compare the default service with credible alternatives that were actually eligible at label time. Then examine where the default repeatedly produces higher cost without a clear service benefit, or where it produces weaker on-time performance without a meaningful cost advantage. The objective is not to prove that every static rule is wrong. It is to identify the subset of decisions where the rule no longer reflects current operating conditions.
This distinction matters because not every shipment represents an optimization opportunity. Some shipments have only one viable service because of destination, dimensions, contractual limitations, operating cutoffs, or customer commitments. Those should remain outside the addressable decision pool. A credible program focuses on shipments where multiple legitimate options exist and where better allocation can plausibly change cost or delivery performance.
The Economics of a Better Decision
The financial case for shipment-level allocation should be built from realized operating outcomes, not from rate-card comparisons alone. A lower quoted rate is only an opportunity until the shipment is executed and the downstream service outcome is known. If the cheaper option creates a missed delivery commitment that triggers support activity, credits, reshipment, or customer dissatisfaction, the apparent transportation saving may not represent a better business decision.
Likewise, a higher-cost service can be economically rational when the additional spend materially increases the probability of meeting an important delivery promise. The key is to make that trade-off explicit. Finance and operations should agree in advance on the objective being optimized, the service boundary that cannot be violated, and the method for distinguishing theoretical, predicted, and realized savings.
A useful review therefore separates three categories. First, realized efficiency: a lower-cost executed service that met the approved delivery objective on a comparable shipment. Second, justified protection: a higher-cost executed service selected because the evidence supported a material reduction in delivery risk. Third, unresolved opportunity: recommendations that were not executed, or where the outcome evidence is incomplete. Keeping those categories separate prevents modeled opportunity from being reported as achieved value.
What IT and Engineering Must Make Reliable
For technology teams, the challenge is not simply connecting an AI model to a shipping API. The production requirement is a reliable decision service inside the label-generation workflow. The system needs to receive the relevant shipment context, determine the eligible option set, evaluate current cost and performance evidence, return a recommendation quickly enough for the transaction, and fall back safely when the evidence or service is unavailable.
That creates several engineering requirements. Inputs need clear authoritative sources. Data freshness needs thresholds. API latency needs monitoring because a recommendation that arrives after the label is purchased has no operational value. Errors must be observable, not silently converted into arbitrary selections. Fallback logic needs to be deterministic so operators know what happens when a required input is missing, a carrier service is unavailable, or the decision engine cannot respond within the allowed time.
Decision logging is equally important. For every automated selection, the organization should be able to reconstruct the shipment context, eligible alternatives, selected option, reason for the recommendation, evidence available at the time, and any fallback or override. That record turns the system from a black box into an auditable operating capability and gives operations a practical way to investigate exceptions.
A Measurement Model That Survives Executive Scrutiny
An executive review should be able to answer whether the new decision logic improved the agreed objective on comparable shipments. That requires more than comparing one month with another. Parcel economics can change because of carrier contracts, package mix, geography, customer promise, seasonality, fulfillment changes, or shifts in the services available to the operation.
Before execution, define the comparison population and document the operating conditions that could affect the result. Where practical, retain a control or comparable baseline. Track actual transportation cost, on-time delivery, late-delivery volume, recommendation-change rate, exceptions, overrides, and carrier/service concentration. If a material operating variable changes during the test, disclose it rather than allowing the change to disappear into the reported result.
The Luma AI Select case study is useful precisely because the source identifies major factors that were reported as unchanged during the measured deployment [1]. Another shipper should not assume the same result, but it can adopt the same measurement discipline: narrow the intervention, compare like with like, and separate observed outcomes from modeled opportunity.
When to Expand, Hold, or Stop
A pilot should end with an operating decision, not simply a presentation of results. For each tested shipment segment, leadership should choose whether to expand automated authority, hold the current scope, constrain the logic further, or return to the approved fallback.
Expansion is justified when comparable outcomes support the cost-and-service objective, exceptions are understood, data quality is stable, and the reasons for changed selections are operationally explainable. Holding is appropriate when early evidence is positive but the sample or operating conditions are not yet representative. Constraining the system is appropriate when a narrow class of shipments creates recurring exceptions or weak evidence. Stopping or reverting is appropriate when service performance deteriorates, input quality becomes unreliable, or the system cannot operate inside the approved risk boundary.
This creates a practical governance principle: automation authority should follow evidence. The goal is not to maximize the number of labels selected by AI. It is to increase the share of routine shipment decisions that can be made consistently, quickly, and accountably while preserving clear human ownership of policy, exceptions, and risk.
Operations and logistics leaders: compare the reported before/after evidence with your own label-time allocation logic.
See the complete Luma AI Select case study and measurement approach
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 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
[6] 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
