A US CFO’s Guide to AI Investment in Last-Mile Delivery Efficiency: Proving the Savings Arrived

The month-twelve conversation nobody prepared for

A US retailer approved a logistics AI investment on four numbers: cost of inaction, investment required, payback period, risk-adjusted return. The case projected roughly $6M in annual savings against $30M of transportation spend. It was well built, and it was approved.

Twelve months later the CFO asks whether the $6M arrived.

Operations reports cost per successful drop down 11%, which is real and measured. Finance reports total transportation spend up 8% against last year, which is also real and measured. Volume grew 19% over the same period, two distribution centers changed, and the carrier rate card was renegotiated in Q3.

Everyone in the room is telling the truth and nobody can answer the question. The investment will now be judged by whichever number the most senior person in the room finds most intuitive, which is usually total spend, which is the one number guaranteed to be wrong.

This is the failure mode worth designing against, and it is symmetric. A deployment that genuinely delivered cannot demonstrate it, so the next phase does not get funded. A deployment that underdelivered cannot be identified, so nobody corrects the course. Both outcomes cost money, and both come from the same source: nobody fixed the verification method before the results started arriving.

Also Read: How to Measure and Maximize the ROI of Logistics Technology Investments

Why the obvious metrics fail

Total transportation spend is unusable for attribution because it is dominated by volume. An operation that grows 19% and holds unit cost flat has improved efficiency materially and will show spend rising. A CFO reading total spend on a growing business is reading a demand signal, not a performance signal.

Cost per successful drop is the correct family of metric and is still not sufficient on its own, because unit cost moves for reasons unrelated to the platform. Drop density changes when demand concentrates or disperses. Average order size changes with promotional activity. The urban-to-rural mix shifts with channel growth. Each of those moves cost per drop by amounts comparable to the savings being claimed, in either direction.

So the verification question is not which metric to use. It is what to normalize that metric against, and the honest answer is that this requires deciding in advance what would have happened otherwise.

Also Read: Last-Mile Delivery Efficiency: Cost Reduction Guide 2026

The four confounders that invalidate your baseline

Each of these will have moved by the time the review arrives. Each has a normalization, and none of them is difficult once named.

Volume change. Higher volume usually improves unit economics by increasing density, independent of any software. Normalize by comparing unit cost at comparable volume bands rather than period against period, or by modeling the density effect explicitly and removing it.

Network change. A new distribution center, a closed depot, or a changed store-replenishment pattern alters stem distance for every route touching it. This is the confounder most likely to be large and least likely to be mentioned in the review, because the network change was a separate project with its own business case, frequently claiming some of the same savings. Exclude affected lanes from the comparison or treat the change as a baseline reset with a documented step.

Input price movement. US diesel prices, driver wage rates, and carrier rate card renegotiations all move the cost line without any change in operational efficiency. Hold input prices constant at baseline levels and report the price effect as a separate, clearly labeled line. A CFO who sees fuel savings folded into a software business case will discount the whole document, correctly.

Mix shift. Service level mix, channel mix, order size, and geography. A shift toward same-day or toward rural delivery raises unit cost regardless of platform performance. Compare like-for-like cohorts rather than the whole book.

The pattern across all four is that a savings claim is a comparison, and a comparison needs a counterfactual. The business case built one implicitly when it forecast, and nobody wrote it down, so at month twelve there is nothing to compare against except a year-old number that described a different business.

Also Read: How to Build a Business Case for Logistics Transformation

What defensible attribution looks like

Three approaches, in ascending order of rigor and descending order of convenience.

Like-for-like cohort comparison. Restrict the comparison to lanes, depots, and service levels present in both periods, with input prices held at baseline and volume bands matched. This is achievable with data every operation already holds and is sufficient for most internal purposes. It is also the minimum that survives a serious question from an audit committee.

Phased rollout comparison. Where deployment reached sites in waves, sites not yet live during a given window are a concurrent control operating under the same volume, weather, fuel prices, and network conditions. This is the strongest evidence available at no incremental cost, and it expires: once the last site is live the comparison is no longer possible. If the rollout is running now, the decision to capture this has to be made now.

Documented modeled counterfactual. Where neither of the above is available, model what would have happened, state every assumption, and choose conservative values rather than likely ones. A conservative attribution with a visible method is worth more to a board than an optimistic one without, and it is considerably more durable when someone challenges it.

Whichever is used, one discipline matters more than the choice: fix the method, the metric definitions, and the normalizations in writing before go-live, alongside the business case. A methodology selected after the results are visible is not evidence, whatever it concludes, and everyone in the review knows it.

Three ways CFOs can verify logistics AI returns

The last two rows are the argument. The cheap metrics fail in both directions, which means the choice is not between rigor and convenience. It is between having an answer and not having one.

Also Read: 3PL CFO ROI Framework: Quantifying Dispatch Automation

When the number misses, diagnose before you escalate

A miss has four possible causes and they call for entirely different responses. Establishing which one you have is the first task, and skipping it is how organizations replace a working platform.

Adoption. The capability is deployed and not being used as modeled, most often visible as a low plan execution rate where dispatchers or sites override the system. This is the most common cause of a projected-to-realized gap and it is an internal change management problem, not a product problem.

Scope. Phase 1 went live narrower than the case assumed, covering fewer sites, regions, or decision types. The savings are proportional to scope, so a 60% scope delivering 60% of the benefit is performing exactly as designed. Compare against a scope-adjusted forecast rather than the original.

Baseline error. The original forecast was wrong, usually because it applied a vendor percentage to a baseline that was already better than average. This is uncomfortable to surface and cheaper to surface early.

Genuine underperformance. The capability is in use, at scope, against a sound baseline, and is not producing the modeled result. This is the only cause that is a vendor conversation, and in practice it is the least frequent of the four.

The order matters. An organization that escalates to the vendor before checking adoption and scope will spend a quarter on the wrong conversation, and will usually find that the answer was visible in its own execution data the whole time.

What to fix at approval, not at review

The CFO’s real leverage sits at the approval gate, before anyone has an incentive to prefer one measurement over another.

Agree the verification method and write it down. Name the four normalizations you will apply. Define each metric precisely, including what counts as a successful drop. Name one owner for the number, sitting in finance rather than operations, because operations should not be scoring its own investment. Set the substantive review at twelve months rather than six, since a six-month reading captures a fraction of the eventual return and reliably understates it. And if the rollout is phased, mandate that the control comparison is captured while the phasing still exists.

None of that is expensive. All of it is close to impossible to introduce later.

Also Read: Transportation Management System TCO: The Renewal Trap

What to measure

Cost per successful drop, normalized. Input prices held at baseline, volume bands matched, like-for-like cohorts. The headline number, and only meaningful with the normalizations stated alongside it.

Plan execution rate. The share of planned decisions executed as planned. This is the adoption metric and the single best early predictor of whether the forecast will be met.

Scope coverage against the modeled scope. Sites, regions, and decision types live as a percentage of what the case assumed. Without this, every variance conversation is confounded.

Price effect, reported separately. Fuel, wage, and rate card movement isolated as its own line. Reporting it separately is what makes the rest of the number credible.

Forecast-to-realized variance, with the cause attributed. The variance split across adoption, scope, baseline error, and performance. A variance without an attributed cause is not a finding, it is an argument waiting to happen.

How Locus makes the realized number auditable

Locus, the world’s first Decision-Intelligent, Agentic TMS, treats the record of what was decided and why as part of the product rather than a reporting afterthought, which is what allows a realized-savings claim to be traced rather than asserted. Its Dispatch, Capacity, Carrier, and Settlement agents, coordinated by an Orchestrator, run a continuous Sense-Decide-Execute-Learn loop against a model of more than 250 real-world constraints.

Three properties matter to a CFO specifically. Traceability means each decision retains its inputs and the plan version it produced, so plan execution rate is measurable rather than estimated, which addresses the most common cause of a forecast miss. Explainability means a decision arrives with the constraints it honored, so a variance can be investigated at the decision level instead of debated at the summary level. And because deployment is typically staged across sites, the rollout sequence is retained, which preserves the concurrent control comparison that produces the strongest available attribution.

Locus is recognized by Gartner for seven consecutive years, featured in the 2026 Hype Cycle for Supply Chain Execution and Logistics Technologies, named a Leader in TMS by QKS Group (SPARK Matrix), and ranked #1 in Route Planning on G2’s 2026 Best Software Awards. In October 2025, Ingka Investments, the investment arm of Ingka Group, the world’s largest IKEA retailer, acquired Locus. Locus continues to operate independently. Further analyst recognition is published in full.

Two deployments illustrate what a verifiable result looks like.

A Fortune 50 parcel and logistics provider centralized dispatch across 51 sites in a 120-country network, running more than a million freight shipments a year against a 4,500-strong driver pool. Two features of this deployment make its numbers unusually defensible. The rollout was staged site by site, so live and not-yet-live sites operated concurrently under identical conditions. And weekly execution rate rose from 75% to 92%, which is the adoption metric moving, meaning the capability was demonstrably in use rather than merely installed. Against that, more than $14 million in previously unused capacity was surfaced, including $565,000 at a single site. Capacity that was already paid for and not being used is the cleanest category of saving to attribute, because it does not depend on a price assumption.

A leading North American retailer replaced six legacy systems with a single orchestration layer across multi-hundred stores and ocean, rail, and road movements, reaching break-even inside year one alongside $1M+ in savings, sub-two-hour exception resolution, route compliance above 95%, and more than 80% less manual dispatch effort. Break-even in the first year is the claim a CFO should interrogate hardest, and the supporting numbers are the ones that make it checkable: compliance and manual-effort reduction are adoption evidence, which is what separates a realized return from a modeled one.

Request a Locus ROI verification and attribution review to fix your metric definitions and normalizations before go-live, and to establish whether your rollout can still support a concurrent control comparison.

Write the verification method into the approval

If an investment in AI for last-mile delivery efficiency is on your desk this quarter, the highest-value edit you can make to the paper is not to the savings forecast.

It is a page specifying how the savings will be verified: the metric definitions, the four normalizations, the owner of the number, the review date, and the control comparison you will capture during rollout.

That page costs an afternoon. Without it, the most likely outcome twelve months from now is not a bad result. It is a result nobody can interpret, defended by whoever speaks last.