AI in Logistics: A Vertical Analysis

Why the world's biggest coordination problem is also the best business case for AI

Data current as of August 2026.

Part 0: The setup

Logistics is the largest industry that almost nobody has ever optimized end to end.

Depending on which analyst house you believe, the global logistics market sits somewhere between $4 trillion and $12 trillion. Grand View puts it at roughly $4.1 trillion in 2025, growing to about $8.5 trillion by 2033. Precedence Research puts it at $11.23 trillion in 2025 heading toward $24.36 trillion by 2035. IMARC lands in between at $5.88 trillion in 2025.

That spread is not a rounding error. It is a five-trillion-dollar disagreement about the size of an industry, and it is the first honest signal in this entire report. The people who count this market cannot agree on what counts. Some include freight spend, some include only logistics service revenue, some include in-house operations, some do not. If the industry cannot measure itself, imagine what its operational data looks like.

That is the opportunity.

Every one of those trillions moves through workflows that are still, in 2026, held together by email threads, PDF attachments, phone calls, spreadsheets, WhatsApp messages, and the memory of one operations person who has been there eleven years. The physical layer of logistics has been industrialized for a century. The decision layer has not.

AI in logistics is not a productivity story. It is an arbitrage on the gap between how much value moves through this industry and how little of the coordination is automated.

Part 1: Sizing the AI opportunity honestly

The market numbers, with a warning label

The "AI in logistics" market size figures floating around are, frankly, unreliable. Here is the actual spread:

• Fortune Business Insights: $8.65 billion in 2025, $12.23 billion in 2026, reaching $196.61 billion by 2034 at a 41.5% CAGR
• GMInsights (AI in logistics and supply chain): $20.1 billion in 2024 growing at 25.9% CAGR through 2034
• The Business Research Company: $26.33 billion in 2025 to $38.68 billion in 2026 at 46.9% CAGR
• Market.us: $12 billion in 2023 to $549 billion by 2033 at 46.7% CAGR

Those cannot all be true. The 2025 figure alone ranges from $8.65B to $26.33B, a 3x spread. Treat any single number you see quoted in a vendor deck with suspicion.

The number I trust most is narrower and more specific. Gartner forecasts that supply chain management software with agentic AI capabilities will grow from under $2 billion in 2025 to $53 billion in spend by 2030, with 60% of enterprises using SCM software having adopted agentic AI features by 2030, up from 5% in 2025.

That is a 26x increase in five years on a specific, definable software category. It is also a specific claim about adoption: 5% today, 60% by 2030. That is the shape of the curve.

The efficiency numbers everyone quotes, and what they actually mean

You have seen this stat: AI-enabled supply chain management delivers 15% lower logistics costs, 35% lower inventory, and 65% better service levels.

It comes from a McKinsey piece from April 2021, comparing early adopters against slower-moving competitors. It is now cited by hundreds of vendors, usually without the date and usually without the comparison group.

Use it, but understand what it is. It is a five-year-old benchmark from the pre-LLM era, measuring classical ML and optimization deployments, and it compares leaders to laggards rather than before to after. The real 2026 numbers are more granular and more interesting, and they come from operators rather than consultants. We will get to them.

The adoption reality, which is the most important number in this report

BCG surveyed more than 180 logistics providers and shippers across Europe, North America, Asia Pacific, and the Middle East in January 2026. The findings:

• About 40% of logistics service providers have deployed AI beyond pilots, but only about one in ten have embedded AI into core operations at scale. Just 13% report measurable value from AI.
• 97% call AI a strategic priority, while only 13% say it is delivering measurable results.
• More than 40% of shippers now factor a provider's AI capabilities into partner selection. Nearly 80% cite cost reduction and operational efficiency as the primary driver.
• Regionally, the gap is enormous. Asia Pacific sits at 31% scaled adoption. Europe sits at 6%.

And the broader context is worse. S&P Global Market Intelligence reported in 2025 that 42% of companies had abandoned most of their AI initiatives, up from 17% a year earlier, with almost half of all pilots scrapped before going live.

So the honest framing of this entire industry right now: near-universal intent, wide but shallow deployment, very thin value capture, and a widening gap between the 13% who have made it work and everyone else.

That gap is the whole game. Everything below is about which side of it you end up on.

Part 2: Why logistics specifically

Not every industry is equally exposed to AI. Logistics is unusually exposed, for five structural reasons.

1. The work is language, not physics

People assume logistics is about moving boxes. Operationally, the majority of white-collar logistics labor is reading and writing text. A quote request arrives as an email. A booking arrives as a PDF. A rate sheet arrives as an Excel file with merged cells. A customs declaration is a structured document derived from three unstructured ones. A delivery exception is a phone call transcribed into a note field.

This is precisely the work that language models are good at. Before 2023, this work was considered unautomatable. C.H. Robinson said so directly: generative AI in 2023 opened the door to overcoming automation challenges that had persisted for decades.

2. Margins are thin enough that percentage gains are transformative

Freight forwarding gross margins typically run 15% to 25% of gross revenue, with net margins in low single digits. Freight brokerage net revenue margins run 13% to 18%. Asset-based trucking operating ratios sit in the low 90s in good years, meaning a 6% to 9% operating margin.

In an industry where net margin is 3%, a 10% reduction in SG&A is not an efficiency gain. It is a doubling of profit.

3. The industry is extremely fragmented

The top 20 freight forwarders account for a little over 50% of the market, meaning tens of thousands of small and mid-size forwarders share the rest. The top five (DSV, Kuehne+Nagel, DHL Global Forwarding, Expeditors, Maersk) held about 52.9% of the market in 2025, with DSV leading at 15.9% (GMInsights).

In US trucking there are hundreds of thousands of active motor carriers, most operating fewer than ten trucks. Fragmentation means no shared data standard, no dominant platform, and enormous per-transaction coordination overhead. It also means AI capability becomes a genuine competitive weapon, because your competitor down the street cannot buy it off the shelf and match you next quarter.

4. Coordination failure is the dominant cost

Almost every cost in logistics that people complain about is a coordination failure rather than a physics problem.

• The American Trucking Associations estimates roughly 35% of all US truck miles are empty. Flatbed runs 40% to 45% deadhead, tanker 35% to 40%, dedicated contract carriage 10% to 20%. That is a matching problem.
• FMC data through Q1 2025 shows nine major carriers collected $15.4 billion in detention and demurrage fees between April 2020 and March 2025. That is a scheduling and visibility problem.
• Global container schedule reliability reached 64.7% in May 2026, with the average delay for late arrivals at 5.52 days (Sea-Intelligence). Which means about a third of container ships arrive late, and when they do they are nearly a week late. Maersk itself notes that importers may be carrying one to two extra weeks of inventory because of ocean network unreliability. That is billions in working capital funding a prediction failure.
• The pharmaceutical sector loses an estimated $20 to $35 billion annually to cold chain failures. Much of that is monitoring and intervention latency.

None of these are solved by better trucks or bigger ships. They are solved by better decisions made faster with more information.

5. The data exists and nobody uses it

Logistics generates enormous volumes of operational telemetry: GPS pings, ELD records, EDI messages, port call data, container tracking events, WMS transactions, telematics streams. Most of it is written to a database, used once for a status update, and never modeled.

C.H. Robinson has 37 million shipments a year of proprietary data. Amazon has inventory movement datasets across more than 300 robotic sites. That data is the moat. Software can be copied. A decade of labeled outcomes on your own network cannot.

Part 3: Workflow by workflow

This is the operational core of the report. Every major logistics workflow, what it costs today, where AI actually attaches, and what the measured results look like.

3.1 Quoting and pricing

What it is: A customer asks what it costs to move something. Someone reads the request, works out the lane, mode, equipment, accessorials, and current market rate, then writes a quote back. In forwarding this can involve checking carrier contracts, tariffs, surcharges, and terminal handling charges across three continents.

What it costs today: Quoting is the single largest consumer of commercial ops labor in forwarding and brokerage. A forwarder pricing desk typically handles 30 to 60 quotes per person per day. Response time is a primary determinant of win rate, and most SME quote requests go unanswered or answered too late to matter. The economics are brutal: quoting is expensive, win rates on ad-hoc quotes run 10% to 25%, so you are paying to lose 80% of the time.

Where AI attaches: This is the most proven agentic use case in the industry, by a wide margin.

C.H. Robinson's quoting agent is the reference implementation:

• Over 1 million quotes delivered, with customer-specific price quotes returned in 32 seconds
• 2,600 emailed quotes delivered per day
• After adding LTL to the quoting agent, at least a 30% month-over-month jump in LTL quotes delivered by AI, with the email classifier able to distinguish when a shipper is asking for both truckload and LTL in one message

Why this works: Quoting is a bounded task with a clean feedback loop. The input is unstructured text. The output is a number. You know within days whether you won. That is a nearly ideal ML setup, and it is why quoting agents were first out of the gate.

What to watch: Quote automation without pricing intelligence just lets you lose faster and cheaper. The real value is in the second layer, where the model learns which quotes are worth winning at which margin. That is the difference between an email robot and a revenue system.

3.2 Order intake and tender processing

What it is: A customer sends a shipment order. Someone reads it, extracts the details, and keys them into the TMS. In a mid-size forwarder this is a full department.

What it costs: Manual order entry runs 4 to 12 minutes per order depending on complexity, with error rates of 1% to 3%. Each downstream error costs disproportionately, because a wrong address or weight is discovered at the dock rather than the desk.

Where AI attaches: Document extraction into structured tender.

C.H. Robinson: emailed load tenders converted into 5,500 shipment orders per day, in 90 seconds each. That single workflow, at 90 seconds versus a manual 6 to 8 minutes, is roughly 500 to 700 hours of daily labor displaced.

Practical note: This is the workflow where accuracy thresholds matter most. A quote that is slightly wrong gets renegotiated. An order that is slightly wrong gets shipped. Confidence-score routing (auto-process above threshold, human review below) is the standard pattern and the reason these deployments succeed.

3.3 Capacity procurement and carrier matching

What it is: Finding a truck, a container slot, an aircraft ULD position. In brokerage this is the core of the business: match a load to a carrier at a spread.

What it costs: The deadhead number tells the story. Roughly 35% of US truck miles are empty. For a truck running 120,000 miles a year that is 42,000 unpaid miles, or $25,000 to $35,000 in annual operating cost per truck at current diesel prices.

There are roughly 4 million commercial trucks in the US. Even a partial reduction in that 35% is a multi-billion dollar annual number.

Where AI attaches:

• Predictive carrier matching, scoring which carriers will accept which lanes at which price before you call
• Automated capacity outreach, which C.H. Robinson has deployed as a dedicated agent alongside quoting and appointments
• Backhaul prediction, matching outbound and return legs across a network in advance rather than reactively

The structural point: Freight matching is a market-making problem, and market makers with better information win systematically. This is where scale compounds. A broker with 37 million annual shipments sees patterns a broker with 30,000 cannot.

3.4 Route optimization and dispatch

What it is: Deciding what goes on which vehicle, in what order, on what path.

What it costs: This is where the best-documented AI ROI in logistics history sits.

UPS ORION is the canonical case (INFORMS):

• Once fully deployed, ORION was expected to deliver $300 to $400 million in annual savings from 100 million fewer miles driven, 10 million gallons of fuel not consumed, and 100,000 fewer metric tons of CO2 per year, against a $250 million total investment
• As of December 2015, ORION had already saved over $320 million
• Route reductions averaged six to eight miles per driver, with the dynamic routing upgrade adding a further two to four miles per driver

Note the leverage: six to eight miles per driver per day. That is a trivial-sounding number that compounds into hundreds of millions of dollars because it multiplies across 55,000 vehicles and 250 working days.

Amazon's DeepFleet is the modern version, applied to robots rather than trucks:

• A generative AI foundation model coordinating over 1 million robots across the fulfillment network, improving robotic fleet travel time by 10%
• Built on Amazon SageMaker and trained on Amazon's own warehouse and inventory data. The Wall Street Journal reported 75% of Amazon's global deliveries are now assisted in some way by a robot

The general pattern: route optimization gives you 5% to 15% mileage reduction in most real deployments. Vendors quote higher. Treat anything above 20% as either a very badly run baseline or a marketing number.

3.5 ETA prediction and in-transit visibility

What it is: Telling a customer when their freight will actually arrive, and knowing before they do when it will not.

What it costs: Everything downstream. Bad ETAs cause detention, demurrage, missed appointments, safety stock inflation, expedite spend, and customer churn. Maersk's own analysis links a 5.02-day industry average delay to importers carrying one to two weeks more inventory than needed. On a $500 million ocean freight program, Xeneta's analysis suggests an additional $2.7 million of working capital tied up.

Where AI attaches: Predictive ETA is one of the oldest and most mature ML applications in logistics, and it is now very good.

• C.H. Robinson reported predictive ETA accuracy of 98.2%
• Modern visibility systems can reach 95% or higher prediction accuracy for arrival times across transport modes
• C.H. Robinson also deployed an in-transit visibility agent for cases where a carrier's automated status updates are not working, which is the interesting part. The hard problem in visibility is not the 80% of shipments that report cleanly. It is the 20% that go dark.

The exception management layer is where the money is. Knowing a container will be late is table stakes. Automatically rebooking the drayage, rescheduling the warehouse appointment, notifying the consignee, and reallocating inventory is the actual value. That is agentic work, and it is the frontier right now.

3.6 Appointment and dock scheduling

What it is: Booking a time slot for a truck to pick up or deliver at a facility.

What it costs: Enormous amounts of phone and email time, and it is the root cause of detention. A driver waiting three hours at a dock is unpaid time for the driver, lost capacity for the carrier, and a detention invoice for the shipper.

Where AI attaches: C.H. Robinson claims the first touchless appointment scheduling in the industry, and for customers who email rather than use the portal, the appointment agent extracts the details needed to lock in a pickup or delivery time 3,000 times a day across more than 26,000 locations, in under 60 seconds.

DHL Supply Chain has gone the same direction with voice. Its partnership with HappyRobot targets hundreds of thousands of emails and millions of voice minutes annually, supporting appointment scheduling, transport status calls, and high-priority warehouse coordination.

Why voice matters: A large fraction of logistics coordination still happens on the phone with small carriers and small facilities who will never adopt a portal. Voice agents are the only way to automate that tier. This is currently one of the fastest-moving areas in the industry.

3.7 Warehouse operations

Warehousing deserves its own subsection because it is the most capital-intensive and most measurable part of the chain.

The labor economics:

• Labor accounts for 50% to 70% of total warehousing budgets, and wages climbed 7% to 9% year over year in 2024
• The BLS turnover rate for warehouse workers is 36%, with the cost of filling a vacant position running 25% to 150% of salary
• Nearly 500,000 warehouse and logistics jobs remain open in the United States, and workforce instability drives operating costs 15% to 25% above industry averages
• US warehousing and storage employment reached about 1.851 million workers in June 2026, with average hourly earnings of $26.76 in May 2026

The automation market:

• Warehouse automation grows from $29.98 billion in 2025 to $34.17 billion in 2026 and $65.74 billion by 2031, at roughly 14% CAGR (Mordor)
• 80% of warehouses still operate without automation
• Gartner projects that by 2030, 50% of new warehouses in developed markets will be designed as robot-centric facilities. Interact Analysis projects that by 2030 only 13% of warehouses will have deployed even one fulfillment AMR

Those last two are worth sitting with. New builds go robotic. The installed base largely does not. That bifurcation is the defining structural fact of warehousing this decade.

Sub-workflows and where AI attaches:

Receiving and putaway. Vision models for damage detection, label reading, and count verification. One documented deployment (SLK) took error rates from 1% to 2% down below 0.01% by combining goods-to-person handling with data-driven controls, using source confirmation and weight checks rather than final inspection.

Slotting. Deciding where each SKU lives. This is a classic optimization problem that ML improves because demand correlation between SKUs shifts constantly. Good slotting cuts pick travel distance by 15% to 30%, and pick travel is typically 50% to 60% of picker time.

Picking. The single largest labor line in a DC. Drakes Supermarkets reports more than 700 units per labor hour at goods-to-person stations, versus 60 to 120 units per hour for conventional cart picking. The caution in that same source is the right one: measure sustained shipped output, because a faster workstation can simply move the bottleneck to packing.

Robotic coordination. DeepFleet's 10% reduction in robot travel time across a million-robot fleet is the clearest example of AI as a coordination layer on top of existing automation. The robots were already there. The model made them 10% better without new hardware.

Labor planning. Only about 25% of DCs currently use advanced labor planning tools. The majority still run on spreadsheets. This is probably the highest-ROI, lowest-capex AI application in warehousing and it is badly under-adopted.

Expected results range: 25% to 30% reductions in labor costs, order fulfillment speed improvements, and accuracy approaching 99%, with AMRs showing payback under 24 months and 250%+ ROI where infrastructure fully supports them.

The overlooked prerequisite: Network infrastructure upgrades cost $30,000 to $150,000 per facility and are rarely included in automation vendor quotes. Less than 10% of mobile device issues on the warehouse floor are ever reported to IT. I include this because it is the single most common reason warehouse AI projects underperform their business case, and no vendor deck mentions it.

3.8 Demand forecasting and inventory positioning

What it is: Predicting what will be needed where, and pre-positioning it.

What it costs: Forecast error is the tax that funds all safety stock. Every point of forecast accuracy improvement releases working capital.

Where AI attaches: This is the most mature AI application in the wider supply chain and the one where the classical numbers hold up.

McKinsey's figures: AI-powered forecasting reduces errors by 30% to 50%, leading to a 65% reduction in lost sales from stockouts, with warehousing costs down 10% to 40%. Transformer-based demand models show 20% to 40% higher prediction accuracy than traditional time series methods like ARIMA and exponential smoothing, with the largest gains in high-volatility and new-product categories.

Why transformers specifically help: Classical forecasting struggles with sparse, intermittent, and cold-start demand. That is exactly where the money is lost. The gains are concentrated in the hard tail rather than the easy head of the SKU distribution.

For logistics providers specifically: The forecast that matters is not product demand, it is volume demand. How many containers, how many pallets, how many pickups, on which lanes, next week. That drives labor rostering, equipment positioning, and space commitments. It is under-modeled relative to retail demand forecasting and it is where 3PLs have a data advantage they mostly waste.

3.9 Customs, trade compliance, and classification

What it is: Assigning HS codes, filing declarations, screening restricted parties, calculating duty, managing free trade agreement eligibility.

What it costs: Classification is skilled, slow, and legally consequential. Manual classification of a novel product takes 10 to 45 minutes. Misclassification produces delays, penalties, and retroactive duty liability.

The tariff environment has made this dramatically more expensive. The World Customs Organization's 2025 Smart Customs Report names AI and ML as the top technology of interest to customs administrations worldwide, and CBP has deployed its own Cargo Classification Tool using text analysis to suggest HTS codes.

Where AI attaches and how well it actually works:

Be careful here, because vendor claims and independent testing diverge.

• Independent testing in December 2025 found top-performing AI tools reaching up to 88% accuracy at full 10-digit classification, with mid-tier tools at 70% to 80% for routine products
• Hybrid systems with dual-threshold decision policies can achieve 85% to 90% automation on routine items, cutting classification time from hours to seconds and reducing error rates 30% to 50% in targeted categories

88% accuracy at 10 digits sounds impressive until you remember the liability structure. When AI misclassifies, the importer of record is still responsible. CBP does not care which tool made the mistake. And there is a data problem that no model solves: manufacturers write product descriptions for marketing rather than for tariff classification.

The correct architecture is therefore not full automation. It is confidence-routed triage: auto-classify the high-confidence routine items, escalate the ambiguous ones, and maintain an audit trail with rationale for every decision. The value is in reallocating expert time to the 10% of items that actually need expertise, rather than eliminating the expert.

C.H. Robinson built exactly this for the July 2025 NMFC overhaul: a dedicated AI agent to automate freight classification ahead of the national LTL classification change.

3.10 Documentation

What it is: Bills of lading, commercial invoices, packing lists, certificates of origin, arrival notices, delivery orders, proofs of delivery.

What it costs: This is the most quantified inefficiency in global trade.

McKinsey estimates 100% electronic bill of lading adoption could unlock around $18 billion in gains for the trade ecosystem through faster document handling and reduced human error, plus $30 to $40 billion in global trade growth from reduced friction. A different McKinsey framing puts direct cost savings at $6.5 billion with $40 billion in increased trade.

The adoption reality is embarrassing:

• In 2021 about 1% of bills of lading were issued electronically. By August 2025 that figure was 11%. Nearly half of industry players use eBLs in some form, up from a third in 2022
• DCSA member carriers, covering roughly 70% of global container trade, have committed to 50% eBL issuance within five years and 100% by 2030
• In May 2025 DCSA completed the first standards-based interoperable eBL transaction, addressing the platform fragmentation that had blocked adoption
• In January 2026 a live cross-platform eBL transaction was completed across IQAX and ICE CargoDocs, spanning issuance, multiple transfers, bank presentation, and surrender
• DCSA's own analysis attributes the slow adoption to soft barriers: organizational inertia rather than technical limitation. Teams that have processed paper for decades resist workflow change, especially when the current process appears to work

Where AI attaches: Two places, and they are opposite in spirit.

The first is the optimistic path: help the industry go paperless faster.

The second is the realistic path: assume paper and PDF persist for another decade, and build extraction that reads them perfectly. That second path is where nearly all the deployed value is today. Document AI that reads a scanned Arabic-language commercial invoice, a handwritten delivery note, and a badly faxed bill of lading, and outputs clean structured data, is worth more right now than any standards initiative.

Both are true simultaneously. The standards will win eventually. The extraction will pay the bills until they do.

3.11 Freight audit, billing, and settlement

What it is: Checking that the invoice matches the contract, matching charges to shipments, and paying.

What it costs: Freight invoices contain errors at rates of 5% to 12% depending on mode. Ocean and air, with their dozens of accessorial charges across multiple parties and currencies, are worst. Most shippers under-audit because the labor cost of a full audit exceeds the recovery on small invoices.

Where AI attaches: Automated three-way matching between rate agreement, shipment record, and invoice; anomaly detection on charge patterns; and automated dispute generation.

This is a quietly excellent AI use case because it is pure arithmetic on messy documents, has a hard ground truth, and produces cash. It is also less glamorous than autonomous trucks and therefore under-invested.

3.12 Claims and cargo damage

What it is: A shipment arrives damaged. Someone files, someone assesses, someone pays.

Where AI attaches: Vision models on dock and delivery photos to detect damage at the point of transfer rather than at destination, which reframes the entire liability question. If you can prove the pallet was intact at the outbound dock and damaged at the inbound one, the claim resolves itself.

Automated claims triage, precedent matching against historical settlements, and predictive claim likelihood scoring (which loads and which lanes generate claims) are all live applications.

3.13 Fleet maintenance

What it is: Keeping trucks, trailers, reefers, forklifts, cranes, and vessels running.

What it costs: Unplanned roadside breakdown of a Class 8 truck costs $400 to $600 per hour in direct and consequential cost, before service failure penalties. Over 69% of fleets report running trucks older than their ideal replacement cycle, with parts and maintenance costs rising year over year.

Where AI attaches: Telematics-driven failure prediction. Telematics systems now predict brake wear, oil change intervals, and tire replacement needs with reported 94% accuracy.

In maritime, the honest view is more nuanced. A Maersk second engineer writing in 2026 described the gap directly: predictive maintenance dashboards were on almost every stand at Posidonia 2026, but the message from the engine room is more complicated than the press releases. AI flags the deviation pattern. The engineer still makes the call. That is the correct division of labor and it is where most industrial predictive maintenance actually lands.

3.14 Fraud, carrier vetting, and cargo security

This has become one of the most urgent AI applications in North American freight, and it is under-covered.

The scale of the problem:

• Estimated losses from supply chain crime reached nearly $725 million in 2025, a 60% increase over 2024, with average value per theft rising to $273,990, up 36% from $202,364 (Verisk CargoNet)
• A 2025 ATRI report found cargo theft results in a $520,000 average annual loss per carrier, with 75% of stolen motor carrier cargo never recovered
• The FBI warned that cyber-enabled strategic cargo theft is surging, with attackers gaining access to logistics systems through phishing, then flooding load boards with fraudulent listings, bidding on legitimate shipments with hijacked identities, and altering bills of lading and delivery destinations
• Highway's Q3 2025 Fraud Index recorded fraud attempts up 219% year over year, with over 48,700 fake carrier identities flagged in a single quarter
• Deceptive pickup schemes involving forged credentials and fake identities jumped 31% year over year in Q1 2026, with an average of 6.4 theft incidents reported per day

Where AI attaches: Identity and behavioral verification at scale.

• Highway blocked nearly 2 million fraudulent email attempts and 8.5 million spoofed phone numbers in 2025, cross-referencing DOT and MC data, insurance records, and behavioral signals to flag double brokering before a load is tendered
• Systems evaluating 150 or more signals per load, spanning identity, lane history, insurance verification, and behavior patterns
• Uber Freight Network reported a 20% decrease in fraud incidents through tightened carrier onboarding, monitoring, and proactive identification

Why this matters strategically: Fraud detection is an adversarial ML problem, which means it is one of the few logistics AI applications where the defender genuinely needs to keep spending. It is also a network-effect business. Highway's value comes from proprietary data disbursed across 1,500 broker customers who report stolen loads and double-brokering events back into the network. Solo deployment is much weaker than networked deployment.

3.15 Reverse logistics and returns

What it is: Getting things back, and deciding what to do with them.

What it costs (NRF / Happy Returns, 2025 Retail Returns Landscape):

• Retailers estimate 15.8% of annual sales will be returned in 2025, totaling $849.9 billion, versus 16.9% and $890 billion in 2024
• An estimated 19.3% of online sales go back. Gen Z consumers averaged 7.7 online returns in the past 12 months, more than any other generation
• Processing a single return costs an estimated $20 to $30 per item, within a $10 to $40 range. Apparel and fashion lead at roughly 24% to 30% return rates, shoes near 27%
• 9% of all returns are fraudulent, and 85% of retailers are deploying AI to detect and prevent return fraud

That last stat is remarkable. Returns fraud detection may be the single most widely adopted AI application in all of retail logistics, at 85% deployment, and almost nobody talks about it.

Where AI attaches:

• Return fraud scoring at the point of authorization
• Automated disposition decisions (restock, refurbish, liquidate, destroy), which is a margin decision made millions of times per year, mostly badly
• Predictive returns, forecasting return volume by SKU and cohort to pre-position reverse capacity
• Vision-based condition grading at the returns dock

3.16 Emissions accounting and sustainability

What it is: Measuring and reporting the carbon intensity of freight movements.

Why it now matters commercially: Regulatory reporting requirements and shipper procurement criteria have made emissions a contractual matter rather than a marketing one. Last mile produces approximately 54% of transport sector emissions and around 13% of overall city emissions, and without intervention delivery vehicle numbers in cities are projected to rise 61% by 2030 with delivery emissions up around 60% globally.

Cities are acting. Dutch municipalities began designating inner-city zones closed to polluting vans and trucks, with rollout planned across 28 cities and Schiphol Airport by 2030.

Where AI attaches: Activity-based emissions calculation from telematics rather than spend-based estimation, which is the difference between a defensible number and a guess. Also route optimization with a carbon objective alongside cost, which produces different routes than pure cost optimization.

3.17 Sales, customer service, and account management

Under-discussed and highly automatable.

Logistics customer service is dominated by "where is my shipment" queries. Gartner projects agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029. In logistics that threshold arrives sooner, because the query set is narrower and the answer lives in a database.

The interesting second-order effect: when status queries are fully automated, the account management role changes from information relay to actual commercial strategy. That is a talent problem more than a technology problem, and most firms are not preparing for it.

Part 4: Segment by segment

Different logistics businesses have completely different economics, different data, and therefore different AI leverage. Here is each major type.

4.1 Freight forwarders

Market: Revenue moving from $170.1 billion in 2026 to $267.8 billion by 2035 at 5.2% CAGR, with air freight forwarding at $36.3 billion in 2025 (21.8% share) heading to $61.8 billion by 2035 (GMInsights). Note the definitional issue again: Mordor puts the market at $602.58 billion in 2026 growing to $776.04 billion by 2031, because it counts gross rather than net revenue.

Structure: Top five hold 52.9%, DSV leads at 15.9%. Below that, thousands of regional players. Consolidation is accelerating: DSV acquired DB Schenker for approximately EUR 14.3 billion, creating the world's largest freight forwarder.

Economics: Asset-light. Gross revenue is mostly pass-through carrier cost. Net revenue (gross margin) is the real business, typically 15% to 25% of gross. SG&A eats most of it. Headcount is the cost base.

The AI thesis for forwarders is the strongest in the industry. A forwarder is, functionally, a knowledge-work firm that happens to book freight. Its cost structure is people reading documents, sending emails, and making phone calls. That is the exact profile that agentic AI attacks.

Highest-value workflows in order:

1. Quoting (highest volume, clearest ROI, proven) 2. Document extraction and order entry 3. Customs classification and declaration prep 4. Exception management and proactive customer notification 5. Freight audit and billing 6. Carrier and rate procurement

What incumbents are doing: Kuehne+Nagel's CEO framed Q2 2026 explicitly around this, citing accelerating AI deployment across the organization, from optimizing operational processes to integrating AI agents, as the foundation for measurable efficiency gains. K+N reported Air Logistics EBIT up 35% year over year in Q2 2026.

What this means for small forwarders: The margin compression is coming from both directions. Large forwarders are automating their cost base. Digital-native competitors are pricing against an automated cost base. Flexport moved from 42nd to 33rd in the Transport Topics rankings at $2.1 billion, reflecting continued expansion in digital freight forwarding.

A 40-person forwarder that automates quoting, order entry, and document handling can operate at the cost structure of a 25-person forwarder while responding faster. That is not an efficiency improvement. That is a different company.

4.2 Customs brokers

Structure: Highly regulated, licence-gated, relationship-driven, and extremely labor-intensive. Revenue per entry is small. Volume is everything.

Why AI hits here hardest of all: The entire value proposition of a customs broker is expert interpretation of documents against a rulebook. That is a language task with a defined ground truth.

But the liability structure protects the incumbent. As noted above, the importer of record remains responsible for correct classification regardless of the tool used. The broker's licence and professional liability is the product being sold, not the keystroke.

So the model is augmentation, not replacement. The broker who processes 300 entries a day with AI assistance instead of 80 manually, at the same accuracy and same liability posture, wins on price and takes share. The broker who does not, does not.

Specific applications: HS classification with confidence routing, restricted party screening, FTA eligibility determination, duty optimization and tariff engineering analysis, post-entry audit and correction, and regulatory change monitoring. That last one has become genuinely valuable given the pace of tariff change since 2025.

4.3 Ocean carriers

Economics: Extremely capital intensive, cyclical, and consolidated into alliances. The operational KPIs are vessel utilization, schedule reliability, bunker consumption, and asset turn.

The reliability problem:

• Global schedule reliability at 64.7% in May 2026, average delay for late arrivals at 5.52 days
• By July 2026 reliability had fallen to a 2026 low, down 8.8% year on year, with average delay rising to 6.06 days, the highest since January 2024. Maersk was the most reliable of the top 13 carriers at 73.7%, the only carrier above 70%
• The Gemini Cooperation between Maersk and Hapag-Lloyd hit 85.0% schedule reliability on all arrivals in March and April 2026, while other alliances sat around 55%

That last comparison is the single most interesting number in ocean shipping right now. An 85% versus 55% reliability gap between alliances on the same trades is a network design and execution difference, not a market condition.

Where AI attaches:

• Stowage planning. A container ship stow plan is a brutal combinatorial problem across weight distribution, port rotation, reefer plugs, hazmat segregation, and restow avoidance. Optimization here directly reduces port time.
• Voyage optimization. Speed and routing against weather, current, and arrival windows. Just-in-time arrival alone eliminates both fuel burn and anchorage waiting.
• Berth and port call prediction. The main input to everything downstream.
• Equipment repositioning. Empty container imbalance is ocean's version of deadhead, and it is a forecasting problem.
• Predictive maintenance on main engines and reefer units.
• Commercial: dynamic pricing, allocation management, and no-show forecasting. Carriers systematically overbook because shippers systematically no-show, and both behaviors are predictable.

Maersk has been building AI agents for autonomous quoting, booking, and compliance.

4.4 Air cargo

Market:

• Full-year 2025 demand measured in cargo tonne-kilometres rose 3.4% over 2024, with capacity up 3.7% and yields down 1.5% (IATA)
• IATA initially forecast 71.6 million tonnes and $158 billion in cargo revenue for 2026
• IATA later revised 2026 volumes to 71.7 million tonnes, up just 0.2% year on year, following disruptions from the Middle East conflict, while raising the cargo revenue forecast to $162 billion, up 7.2% from $151 billion in 2025
• Fuel costs are projected to surge nearly 40%, from $252 billion in 2025 to $350 billion in 2026, following the closure of the Strait of Hormuz in early March

That revision is a useful lesson about forecasting in this industry. A December forecast of 2.4% growth became a June forecast of 0.2% growth because of a geopolitical event. Any AI system that models freight demand needs to handle regime change, not just seasonality.

Where AI attaches:

• Capacity and revenue management. Air cargo is a perishable-inventory yield problem, structurally identical to airline passenger revenue management but far less sophisticated. This is the largest untapped AI opportunity in air freight.
• ULD build and load planning. Volumetric optimization against weight and balance constraints.
• Booking no-show prediction. Air cargo no-show rates are notoriously high and carriers overbook blindly.
• Special cargo handling. Pharma, perishables, live animals, and dangerous goods each carry compliance workflows that are document-heavy and automatable. DHL Global Forwarding's GDP-certified pharmaceutical air product and Kuehne+Nagel's KN PharmaChain are the established examples of end-to-end cold-chain documentation and excursion management.

4.5 Truckload carriers (asset-based)

Economics: Operating ratio is everything. A 93 OR means 7% operating margin. Driver cost, fuel, equipment, and insurance are the four big lines.

The labor question, handled honestly. The driver shortage number is contested and you should know that before quoting it.

• ATA cited an 82,000-driver gap for 2026
• Other analysis puts it at 60,000 to 80,000 unfilled positions, roughly 2% of the approximately 3.5 million US truck drivers
• ATRI figures cited elsewhere put the 2026 shortage at 175,000 positions
• And a substantial part of the industry rejects the framing entirely, arguing the issue is pay, parking, home time, and respect rather than a lack of licence holders

The regulatory picture compounds it. The Drug and Alcohol Clearinghouse has removed over 200,000 drivers since 2020, the fully enforced ELD mandate reduced effective capacity by an estimated 3% to 5%, and federal rules implemented in March 2026 bar asylum seekers, refugees, and DACA recipients from obtaining or renewing CDLs.

Whatever you believe about the shortage framing, the operational conclusion is the same: driver capacity is constrained, expensive, and getting more so.

Where AI attaches:

• Load planning and network balancing to attack the 35% empty mile figure
• Driver assignment optimization against home-time preferences, which is a retention lever disguised as an ops problem
• Predictive maintenance
• Safety: in-cab vision for distraction and following distance, plus predictive risk scoring that identifies drivers likely to have an incident before they do
• Fuel optimization and route-level fuel purchasing
• Detention prediction and appointment management

Retention is the AI application people miss. At 36% to 90%+ annual turnover depending on segment, and $8,000 to $12,000 replacement cost per driver, a model that predicts which drivers are about to quit and what would keep them is worth more than a routing optimizer at most carriers.

4.6 LTL carriers

Economics: Network businesses with terminal density as the moat. Freight class, dimensions, and density determine profitability, and historically carriers guessed at all three.

The 2025 classification change was a forcing event. The NMFC overhaul in July 2025 shifted the system toward density-based classification. C.H. Robinson responded with a dedicated freight classification AI agent, noting it moves more LTL freight than any other 3PL in North America.

Where AI attaches:

• Dimensioning and density. Vision-based dimensioners at the dock, feeding automatic reclassification and billing correction. LTL carriers historically lost enormous revenue to under-declared dimensions.
• Linehaul optimization. Which terminals connect to which, at what frequency, is a network design problem that changes with demand and is usually re-optimized annually rather than continuously.
• Pickup and delivery route optimization at the terminal level.
• Dock load planning. Cube utilization on a trailer is a packing problem worth real money at scale.

4.7 Freight brokers and non-asset 3PLs

Market: The 3PL market was valued at $1.6 trillion in 2025, growing to $1.8 trillion in 2026 and a projected $4.3 trillion by 2035 at 10.1% CAGR, with DHL leading at 7.3% share and the top five holding 17% (GMInsights). Other estimates are more conservative: $1,261.0 billion in 2025 growing to $2,502.2 billion by 2033 (Grand View).

Economics: Pure spread business. Net revenue margin of 13% to 18%. Headcount per load is the entire operating leverage question.

This is the segment where AI has already visibly changed the P&L. C.H. Robinson is the proof:

• Over 30 AI agents deployed across the quote-to-cash lifecycle, with productivity improved by more than 35% since the end of 2022
• More than 3 million shipping tasks automated, contributing to a 30% productivity increase since 2023
• Over 10,000 transactions per day automated
• Operating across 75,000 customers and 450,000 contract carriers, managing 37 million shipments annually representing $23 billion in freight

A 35% productivity improvement in a business whose cost base is people is not incremental. It is the difference between shrinking and growing during a freight recession.

The strategic implication for smaller brokers: Brokerage has historically been an easy business to enter and a hard one to scale, because scale required linear headcount. AI breaks that link. The consequence is that mid-size brokers get squeezed from both ends, by large brokers with automated cost structures and by small brokers running lean with off-the-shelf AI tooling. The uncomfortable middle is 50 to 300 people.

4.8 Contract logistics and warehousing 3PLs

Economics: Cost-plus or gain-share contracts, thin margins, labor pass-through, and multi-year commitments. Productivity improvement is the only real lever.

The scale example: Kuehne+Nagel is rolling out an agentic-AI warehouse management platform across more than 1,000 sites in close to 100 countries, built on Blue Yonder's Warehouse Management Solution, with phased deployment beginning April 2026 and first customer go-live in Asia in July 2026.

That is one of the largest single AI-adjacent deployments in logistics history, and it tells you what the incumbent playbook looks like: standardize the platform globally, then layer intelligence on a consistent data model. The standardization is the hard part and it is why it takes years.

DHL's parallel path: over 8,000 collaborative robots deployed globally, with the SVT Robotics SOFTBOT platform live across 30 sites in March 2026 and expanding beyond 100, making robotics integrations 12 times faster. Notably, DHL still hired 40,000 new employees alongside robot deployment.

Where AI attaches: Everything in section 3.7, plus contract-level intelligence: which accounts are profitable at which volumes, which SLAs are systematically breached, and which sites are drifting from their cost model. Most 3PLs discover a bad contract at renewal. AI finds it in month three.

4.9 Parcel and express integrators

Volume: US parcel volume reached 23.1 billion shipments in 2025, up 3.3% year over year, with projections reaching 31 billion by 2031. "Other" carriers more than doubled their revenue share from 3.4% in 2024 to 7.2% in 2025 (Pitney Bowes Parcel Shipping Index).

That carrier-share shift is a signal. Regional and alternative carriers are taking share from the big three, and they are doing it partly on cost structure.

Where AI attaches: Integrators are the most AI-mature logistics segment because they solved the hard problems first out of necessity. UPS ORION and Amazon DeepFleet both live here. The current frontier is:

• Dynamic delivery density optimization, deciding not just the route but which packages go today versus tomorrow
• Address quality and geocoding correction, which is a surprisingly large source of failed delivery
• Delivery window prediction and customer-facing precision
• Automated sortation with vision-based induct
• Network flow optimization across hubs

4.10 Last mile and courier

Cost structure:

• Between 2018 and 2023 the last-mile share of total shipping costs rose from 41% to 53%. US delivery costs increased an average of 12% from 2024 to 2025
• Labor alone accounts for roughly 50% of last-mile expenses, with fuel adding another 10% to 25%, and surcharges inflating invoices by 30% to 40%
• The last-mile delivery transportation market was $186 billion in 2025, forecast to reach $487 billion by 2035 at 10.12% CAGR

The profitability problem is real and under-acknowledged: 85% of retail executives surveyed in 2024 said reducing total cost per order was their number one last-mile priority. Three out of four said home delivery does not add to profitability under current cost structures.

Failed deliveries: Approximately 5% of last-mile deliveries fail, at an average cost of $17.78 each. Urban deliveries typically cost around $10 per package while rural deliveries can reach $50.

Where AI attaches:

• Route optimization with real-time re-sequencing (the ORION dynamic model)
• Failed delivery prediction, so you know before dispatch which stops will fail and can pre-empt them with a different delivery method
• Delivery time window prediction and proactive customer communication
• Crowdsourced driver supply forecasting and dynamic incentive pricing
• Address resolution

The autonomous layer is arriving faster than most people expect.

• Zipline crossed 2 million commercial deliveries in January 2026 at a $7.6 billion valuation, with Platform 2 hovering at roughly 300 feet and lowering a tethered droid that places orders within about 3 feet
• Wing has completed more than one million commercial deliveries and is expanding its Walmart network toward nearly twenty US markets
• Amazon Prime Air plans to expand to nearly 500 US cities and towns by the end of 2026, having already delivered hundreds of thousands of packages this year

The economics are the interesting part. Zipline's CEO argues a flight can cost about $1 at scale, against McKinsey's $9 to $11 estimate for ground delivery. If that number holds at scale for even a subset of parcels, the entire last-mile cost curve changes.

4.11 Drayage and intermodal

The most fragmented segment in the industry and the one with the least technology penetration. Thousands of small carriers, port-adjacent, running on phone calls.

The cost of failure here is measured in D&D. Nine major carriers collected $15.4 billion in detention and demurrage between April 2020 and March 2025, with levels still hovering 85% above pre-2020 norms even after Q1 2025 declines. 2025 averages ran $150 to $300 per day for demurrage and $100 to $250 for detention globally.

Where AI attaches: Container availability prediction (when will it actually be free at the terminal), appointment automation at terminals, chassis positioning, street-turn matching (using an import container for an export instead of returning it empty), and drayage capacity matching.

Street-turn optimization alone is a large unexploited value pool. It requires knowing simultaneously which importers have empties and which exporters need boxes, in the same geography, within the same time window. That is a matching problem no individual party can see.

4.12 Cold chain and pharma logistics

Market: The global cold chain logistics market was valued at $313.3 billion in 2024, projected to reach $410.7 billion by 2028.

The failure cost:

• The pharmaceutical sector loses $20 to $35 billion annually to cold chain failures, and even minor temperature deviations of 1 to 2 degrees Celsius can degrade biologics, vaccines, or insulin
• About 20% of temperature-sensitive products are damaged in transit
• The WHO estimated that before COVID-19 up to 50% of vaccines were wasted globally each year due to lack of proper temperature control and logistics

Where AI attaches:

• Excursion prediction rather than excursion detection, which is the whole game. Detecting that a shipment went out of range tells you it is already ruined. Predicting it 6 hours out lets you intervene.
• Route selection weighted by thermal risk, including ambient conditions and dwell exposure at transfer points
• Packaging configuration optimization against route thermal profile
• Automated GDP compliance documentation
• Stability-model-informed disposition: a 40-minute excursion at 9 degrees C does not always mean destroy the product, and AI-assisted stability budgeting recovers product that would otherwise be scrapped

This last one is worth emphasizing. Most temperature excursions result in conservative destruction because nobody has time to run the stability calculation. Automating that decision recovers real inventory.

4.13 Bulk, tanker, and project cargo

Bulk and tanker: Voyage economics dominate. AI attaches at weather routing, hull and propeller performance monitoring, bunker optimization, laytime and demurrage management, and chartering decision support. The commodity trading overlay makes freight rate prediction genuinely valuable here in a way it is not in other segments.

eBL adoption is ahead of container here: BHP, Rio Tinto, Vale, and Anglo American reached 25.1% eBL usage by mid-2024, ahead of schedule, with some commodities hitting 60%.

Project and heavy lift: Low volume, high value, enormously document and engineering intensive. Every shipment is bespoke: route surveys, permits, lift plans, escorts, bridge clearances. AI attaches at route feasibility analysis, permit workflow automation, and historical precedent retrieval. The dataset is small, so the application is retrieval and reasoning over prior projects rather than statistical prediction.

4.14 Ports and terminals

Where AI attaches:

• Berth allocation and quay crane scheduling
• Yard planning and container stacking to minimize rehandles, which is one of the largest hidden costs in terminal operations
• Gate automation with OCR and vision
• Truck appointment systems with predictive slotting
• Equipment predictive maintenance

Results: McKinsey Global Institute research indicates automation can increase port productivity by up to 30%, though this requires genuine executive commitment and sustained change management. Academic work has measured a 12.7% improvement in operational throughput from partial automation including automatic scanning and handling systems.

Scale of the data operation: The Port of Antwerp-Bruges runs 3,000+ sensors feeding its APICA system. Rotterdam handles 1.2 million data points daily. A single port's digital twin implementation can exceed $2 million in initial sensor and software investment alone.

4.15 Shippers' own logistics organizations

Easy to forget, but a very large share of logistics activity never touches a 3PL. In-house transportation and distribution teams at manufacturers and retailers manage enormous freight spend with worse tooling than the providers they hire.

Where AI attaches: Freight procurement and bid optimization, mode selection, carrier scorecarding, network design, landed cost modeling, and inventory positioning against service targets. The single highest-value application is usually the simplest: knowing what you actually spend, on what, with whom, at what service level. Most large shippers cannot answer that question cleanly today.

Part 5: The reality check

I have quoted a lot of impressive numbers. Here is the counterweight, because a research piece that only cites wins is marketing.

Most of this fails

Only 13% of logistics providers report measurable value from AI. 42% of companies had abandoned most of their AI initiatives by 2025, up from 17% a year earlier. Gartner found only 28% of infrastructure and operations AI use cases fully succeed and meet ROI expectations, with 20% failing outright, from a survey of 782 leaders. Gartner also predicts over 40% of agentic AI projects will fail by 2027 without proper controls.

The reasons are consistent and boring

Data. McKinsey's own warning alongside its efficiency numbers is that fewer than 20% of enterprises successfully scale from AI pilots to full supply chain deployment, and the biggest obstacles are data silos, organizational inertia, and lack of cross-departmental collaboration rather than technology.

People. BCG's diagnosis is blunt: companies are underinvesting in the human side of AI deployment, lacking the talent and change management practices to sustain adoption. And in logistics specifically, the leading gap is people readiness, meaning workforce digital literacy and change management capacity, rather than process discipline, which tends to be strong.

Tools instead of systems. BCG's sharpest line: companies deploying standalone tools are behind because their investment is not yet a system, rather than because they failed to invest.

That is the most useful sentence in the entire body of 2026 logistics AI research. Thirty disconnected point solutions do not compound. One connected system does.

The timeline is longer than vendors say

85% of organizations increased AI investment over the past 12 months, yet only 6% saw ROI in under a year. Most achieve satisfactory returns within two to four years.

Two to four years. Plan for that. Budget for that. Do not promise your board twelve months.

One genuinely encouraging signal

72% of logistics employees adopted AI tools in 2024, the highest rate across all industries, and workplace analytics show they were already using them in some form before their organizations rolled out formal programs.

The workforce is ahead of the leadership. That is unusual and it is an asset. The risk is that this grassroots adoption fragments into unsupported shadow tooling instead of being channeled into the operations stack.

Part 6: The 10 to 20 year picture

Now the speculative part, stated specifically rather than vaguely, because vague futurism is worthless.

The near horizon: 2026 to 2030

Agentic operations become the default. Gartner's forecast of 60% agentic adoption in SCM software by 2030, up from 5% in 2025, and $53 billion in spend is the single best-anchored prediction available. Combine that with Gartner's projection that by 2030, 50% of cross-functional SCM solutions will use intelligent agents to autonomously execute decisions in the ecosystem.

The practical meaning: by 2030, a quote, a booking, a carrier assignment, an appointment, a customs entry, and an invoice will move through a normal shipment without a human touching any of them, and a human will be involved only when something is unusual.

The org chart changes before the headcount does. Gartner surveyed 509 supply chain leaders globally between July and October 2025 and found 55% expect agentic AI to reduce entry-level hiring needs.

Note the specific mechanism. Not layoffs. Reduced entry-level hiring. That is how this actually plays out in logistics: attrition plus a hiring freeze at the bottom of the pyramid, over five years, quietly. The consequence is a broken talent pipeline, because the entry-level job is how people learned the business. Nobody has solved that yet and almost nobody is talking about it.

Autonomous trucking crosses from demo to line-haul infrastructure. The 2026 state of play:

• Aurora had accumulated over 5.3 million cumulative commercial miles through April 2026, validating driverless operations on Dallas-Laredo routes within six weeks of starting supervised runs
• Aurora surpassed 250,000 driverless miles in January 2026, nearly triple the cumulative total from early October, while maintaining 100% on-time performance and zero Aurora Driver-attributed collisions
• A software release validated driverless operation in rain, fog, and heavy wind. During 2025, inclement weather had constrained Aurora's Texas driverless operations roughly 40% of the time
• Aurora expects 200 self-driving trucks in operation by the end of 2026, with second-generation technology produced at half the cost
• Kodiak doubled its fleet to 20 fully driverless trucks with Atlas Energy Solutions and introduced triple-trailer capability, logging over 10,700 hours of paid driverless operations by the end of 2025, a 106% quarter-over-quarter increase
• Volvo Autonomous Solutions plans fully driverless US highway operation in Q1 2027, targeting more than 300 trucks by year-end. PepsiCo and Gatik have launched the largest commercial driverless deployment to date across Texas, Arizona and Arkansas. Hirschbach plans to own 500 Aurora Driver-powered trucks
• California lifted its ban on driverless vehicles over 10,000 pounds in April 2026, and both Aurora and Kodiak received testing permits

But keep the scale honest. Fewer than 50 driver-out trucks run nationally as of mid-2026, against roughly 3.5 million drivers. The correct reading is that autonomy is real, commercially operating, and scaling fast in percentage terms from a tiny base.

The structural change that matters most is not labor cost. It is the roughly 1,000-mile Fort Worth to Phoenix lane, the first autonomous freight corridor exceeding federal hours-of-service limits for human drivers. A truck that runs 22 hours a day instead of 11 does not halve the cost per mile. It roughly halves the asset base needed to move the same freight, and it collapses transit times on long lanes. That reshapes network design, warehouse siting, and inventory positioning.

Warehouses bifurcate permanently. 50% of new warehouses in developed markets designed as robot-centric by 2030, while only 13% of the total warehouse base will have deployed even one fulfillment AMR. Two industries, one label.

The far horizon: 2030 to 2045

1. The company sells outcomes, not capacity.

Today you buy a container slot, a truckload, a pallet position. Tomorrow you buy a delivery guarantee at a price, and the provider's system decides mode, route, carrier, and timing dynamically. The provider absorbs the variance because it can predict the variance better than you can.

This is already visible in the reliability data. An 85% versus 55% schedule reliability gap between alliances is a market inefficiency that will eventually be arbitraged by whoever can price it. The first forwarder to sell a genuine delivery-date guarantee at scale, priced off its own prediction model, takes the market.

2. Pricing becomes continuous and personalized.

Freight pricing today is a negotiated annual contract with a spot overlay. In an AI-native industry it becomes a continuously quoted market where every shipment is priced against real-time network state, the specific customer's history, and the marginal cost of that specific movement given everything else in the network.

The precursor exists: customer-specific price quotes in 32 seconds, over a million of them. Extend that to full dynamic pricing and freight starts behaving like airline seats.

3. The unit of competition becomes the model, and the input is proprietary outcome data.

Software is commoditized. Any forwarder can buy a TMS. Any broker can buy a load board. What cannot be bought is ten years of labeled outcomes on your own network: which carriers actually deliver on time in which conditions, which lanes go wrong in which weather, which customers actually pay, which quotes convert at which margin.

This is why C.H. Robinson's 37 million annual shipments and Amazon's inventory movement datasets matter more than either company's algorithms. The algorithms will be published. The data will not.

Corollary: mid-size operators who do not start capturing structured outcome data now will find in 2032 that they cannot buy their way back in. Not because the software is expensive, but because they have no history to train on.

4. Physical AI closes the loop.

Gartner's 2026 Supply Chain Technology Trends report groups the shifts into autonomy and agency (agentic AI, collaborative multiagent systems, physical AI, polyfunctional robots), specialization and intelligence (intelligent simulation, domain-specific language models), and trust and governance (product provenance, decision governance), describing these as transformational rather than incremental upgrades to existing systems.

The phrase that matters is polyfunctional robots. Today's warehouse robot does one thing. The next generation does many things, which changes the economics of automation for facilities that handle varied product. That is what unlocks the 80% of warehouses currently without automation.

5. Trade documentation finally dies.

DCSA carriers covering roughly 70% of container trade committed to 100% eBL by 2030. Combined with MLETR-aligned legislation across Singapore, the UK, the UAE, Bahrain, and India's Bills of Lading Act 2025 replacing a statute last updated in 1856, the legal foundation is being laid now.

By 2035 the paper bill of lading is a curiosity. The interesting consequence is not the saved courier costs. It is that trade finance becomes programmable. When title transfer is a verifiable digital event, working capital can be released at the moment of transfer rather than weeks later. That unlocks the $30 to $40 billion in trade growth McKinsey attributes to reduced friction, and it disproportionately helps small exporters currently locked out by documentation complexity.

6. The exception becomes the job.

If 95% of shipments move without human touch, the human role becomes exclusively exception handling, relationship management, and network design. That is a smaller, more senior, better-paid workforce. It is also a workforce that needs to be trained differently, because you cannot learn to handle exceptions if you never handled the routine cases.

Every logistics company that automates entry-level work without redesigning how expertise is built will have a competence crisis around 2033. Watch for it.

7. Consolidation accelerates, then reverses.

Near term, AI favors scale, because data compounds and fixed technology costs amortize. Expect the DSV-Schenker style megadeals to continue.

Longer term, I think it reverses. Once AI capability is available as infrastructure, the minimum viable size of a competitive logistics operation collapses. A ten-person forwarder with excellent tooling and deep niche expertise can serve a specialized trade lane better than a 100,000-person global. The industry may end up barbelled: a handful of infrastructure-scale networks, and thousands of small specialists, with the middle hollowed out.

8. Physical throughput stops being the constraint.

The deepest change is the one hardest to see. Logistics has always been constrained by physical capacity: ships, trucks, docks, warehouses. Once the decision layer is fully automated, the binding constraint moves to the physical layer, and capital flows there.

That means the winning logistics companies of 2040 might look less like software companies and more like infrastructure companies with software-company margins. They will own the terminals, the trucks, and the buildings, and they will run them at utilization rates that are currently impossible because nobody can coordinate that well.

Part 7: What to actually do

If you run or work in a logistics business, here is the sequence I would follow.

Start with document extraction and quoting. They have the clearest ROI, the fastest payback, and the most proven implementations. Quoting in particular, because 32-second quote turnaround is a commercial weapon, not just a cost save.

Instrument everything you do before you try to optimize it. The reason most of these projects fail is data, and data is a two-year problem you should have started two years ago. Start logging outcomes now, structured, with timestamps, even if you have no model to feed. Especially if you have no model to feed.

Do not buy thirty tools. BCG's finding is the one to internalize: the leaders of the next decade will be the companies that build AI as one system rather than as a drawer of tools. Pick a spine (usually your TMS or WMS data model) and hang everything off it.

Budget two to four years and staff for change management. Only 6% see ROI within twelve months. If your business case requires payback in nine months you will kill the project at month eight, one month before it starts working.

Assume your competitors are further along than they say and less far along than you fear. With only about one in ten having scaled AI into core operations, the window is still open. With Asia Pacific at 31% versus Europe at 6%, it is closing at different speeds in different places.

Fix the boring infrastructure. Network upgrades at $30,000 to $150,000 per facility that vendors do not quote will kill your warehouse AI project quietly, and you will blame the AI.

The execution gap

The technology question in logistics AI was settled around 2024. Everything since has been an execution question.

The numbers in this report are not evenly reliable. Market sizing is guesswork. Vendor case studies are marketing. The McKinsey 15/35/65 figure is a five-year-old comparison of leaders to laggards that has been laundered into a universal promise.

But some numbers are solid, and they point the same direction. Three million shipping tasks automated. 35% productivity improvement since 2022. A million robots coordinated by one model, 10% faster. A hundred million miles removed from a delivery network. Five million commercial autonomous miles. Those are audited operational facts from companies that had to report them.

And against that: 13% of the industry reporting measurable value.

The distance between those two sets of numbers is the entire strategic opportunity in logistics right now. It is not a technology gap. It is an execution gap, and execution gaps close for the people who start early and stay boring about it.

Part 8: What an AI-native logistics company looks like

Everything above is diagnosis. This section is the picture at the end of it.

The defining feature of logistics today is fragmentation. Not fragmentation of market share, though that exists. Fragmentation of knowing. A single container moving from a factory in Shenzhen to a distribution centre in Ohio passes through roughly fifteen organizations, and no two of them share a system, a data model, or a definition of what "delivered" means. The shipper knows what it ordered. The forwarder knows what it booked. The carrier knows what it loaded. The terminal knows what it discharged. The drayage company knows what it collected. The warehouse knows what it received. Each of them is confident. None of them agrees. Reconciling those fifteen partial views is what most of the industry's labor is actually spent on.

An AI-native logistics company is one where that reconciliation has already happened, continuously, before anyone asks.

One model of the world

The first thing you notice about an AI-native operator is that there is a single, live, canonical picture of every shipment, and everyone touching it looks at the same one.

Not a portal that aggregates six systems and shows you six answers. One state. The container has one location, one ETA, one temperature history, one customs status, one financial position, and one set of downstream commitments attached to it. When the vessel slows down, the ETA changes, and the warehouse appointment, the drayage booking, the inventory allocation, the customer's promise date, and the working capital forecast all change with it, automatically, in the same second.

The company achieves this by treating the shipment as the primary object and everything else as a projection of it. The TMS view, the warehouse view, the finance view, the customer view: all of them are windows onto one record. There is no synchronization problem because there is nothing to synchronize.

This sounds like a data architecture point. It is actually the whole thesis. Almost every cost in this industry that people attribute to complexity is really the cost of maintaining inconsistent copies of the truth.

The shipment that nobody touched

Here is what a normal day looks like inside one.

A customer sends a purchase order. No portal, no EDI mapping project, no onboarding. They send it however they already send things, and the system reads it, understands it, and knows what it means in the context of that customer's history, contracts, and preferences. Within seconds a price exists, informed by live network state and the marginal cost of that specific movement given everything else already committed. The price is accepted or negotiated by an agent that knows the account's margin history and the company's current capacity position.

Capacity is secured before anyone would have picked up a phone. The system already knew which carriers would accept that lane at that price this week, because it has watched them accept and refuse for years. Equipment is positioned in anticipation, because the volume forecast for that lane was built three weeks ago and has been updating hourly since.

Documents generate themselves. The commercial invoice, packing list, and bill of lading are derived from one source of truth rather than retyped four times. Classification happens at the moment the product is created rather than at the moment it crosses a border, with the rationale attached and auditable. Customs filing is a formality because the data has been complete and validated since the order.

The goods move. Every transfer point is verified by vision. Every deviation is detected before it becomes an exception. When the vessel is delayed six hours, nothing dramatic happens: appointments shift, the consignee is told before they ask, the safety stock calculation adjusts, and the yard plan at the destination terminal reorders itself.

The invoice is correct the first time, because it was constructed from the same record as the operation. It is paid without dispute. The margin on that shipment is known the day it delivers rather than at month-end close.

No human touched any of it. Not because humans were removed, but because nothing went wrong, and nothing going wrong is now the normal case rather than the lucky one.

What the humans do

They handle the 5%.

An AI-native logistics company has a different shape of workforce. Fewer people, considerably more senior, doing work that is genuinely difficult. Three kinds of work survive and grow.

Exception judgment. When something breaks in a way the system has not seen, a person with deep domain knowledge decides. This is the highest-value work in the company and it is where all the interesting problems live: the shipment that is legally ambiguous, the customer relationship worth protecting at a loss, the disruption that requires improvising a route nobody has run before.

Network design. Deciding what the system should optimize for, where to put capacity, which lanes to commit to, which customers to build around. The machine optimizes within the network. Humans decide what the network is.

Relationships and trust. Freight is still a business where a shipper hands over a million dollars of goods on the strength of believing someone. That does not automate, and its value goes up as everything around it becomes commoditized.

The uncomfortable part is that the ladder into those roles has to be rebuilt. In the old model you learned the business by doing the routine work for three years. In the new one the routine work is gone, so expertise has to be manufactured deliberately: simulation, shadowing, rotation through exception queues, deliberate exposure to the hard cases. The companies that get this right will have a talent advantage that compounds as brutally as their data advantage.

The boundaries between companies get thin

The deepest change is that the seams between organizations stop being where information dies.

Right now every handoff between a shipper, a forwarder, a carrier, a terminal, and a warehouse is a translation event, and translation events lose data. In an AI-native industry the handoff is an API call against a shared standard, or increasingly an agent-to-agent negotiation where two systems settle a booking, a rate, an appointment, and a liability position between them in seconds without either company's staff being involved.

That is where the electronic bill of lading work, the DCSA interoperability standards, and the MLETR legislation all converge. Once title, custody, and condition are all verifiable digital facts, the supply chain stops being a relay race of documents and becomes something closer to a single distributed system with multiple owners.

The consequence is that "end to end visibility", which the industry has been selling as a feature for fifteen years, stops being a product. It becomes an assumption. You will no more pay for visibility than you pay for your bank to know your balance.

The company becomes an instrument, not an intermediary

An AI-native logistics company does not make its money on information asymmetry. That is worth stating plainly, because most of the industry's current margin quietly depends on it. The forwarder knows the carrier rate and the customer does not. The broker knows the market and the carrier does not. When information is cheap and universal, those spreads compress.

What replaces them:

• Guaranteed outcomes. Selling a delivery date with real financial consequence attached, priced off a prediction model good enough to underwrite. That is an insurance business layered on a logistics business, and it is only possible if your forecasts are genuinely better than the market's.
• Utilization. Running assets at occupancy rates that were previously impossible. A terminal, a fleet, or a building operated at 90% instead of 65% is a completely different return profile, and coordination quality is the only thing standing between those two numbers.
• Working capital. When custody and condition are verifiable in real time, financing the goods in motion becomes low risk and highly automatable. Logistics companies become the natural providers of that capital because they hold the data that prices it.
• Specialization. Deep, narrow expertise in a trade lane, a commodity, a regulatory regime. This gets more valuable as generic execution becomes free, because the remaining hard problems are all specific.

What it feels like from outside

For a customer, the endgame is that logistics disappears as a thing they think about.

They do not track shipments, because they are told what they need to know before they wonder. They do not chase documents, because documents are not a separate artifact from the transaction. They do not negotiate rates quarterly, because pricing is continuous and fair by construction. They do not carry six weeks of safety stock to insure against a 5.5 day average delay, because the delay is predicted with enough precision that the buffer can be two days instead of fourteen.

That last one is the biggest number in this entire report and it does not appear in any market sizing model. Global inventory exists in large part as insurance against logistics uncertainty. Trillions of dollars of working capital sit on pallets around the world because nobody trusts an ETA. Reduce the uncertainty and you do not just make logistics cheaper. You release capital from the entire physical economy.

The honest caveat

Nothing in this section is science fiction. Every component of it exists somewhere today, in production, at a real company, at partial scale. A million coordinated robots. A million AI-generated quotes. 98.2% ETA accuracy. Cross-platform digital title transfer. Driverless freight on a thousand-mile lane. Warehouse platforms standardized across a thousand sites in a hundred countries.

What does not exist yet is all of it in one company, connected, on one model of the world.

That is the actual prize, and it is a systems integration problem rather than an AI problem. The industry will get there by 2040, probably sooner in pockets. The companies that get there first will not have invented anything. They will have connected things everyone else bought separately.

That has always been what logistics is.

Sources referenced include: BCG (2026 logistics AI survey with Alpega), Gartner (SCM agentic AI forecast, 2026 Supply Chain Technology Trends, Future of Supply Chain 2026), McKinsey (supply chain AI, eBL economics, delivery cost), IATA (2025 full-year and 2026 cargo outlooks), Sea-Intelligence (Global Liner Performance), C.H. Robinson (AI agent disclosures 2024 to 2026), Amazon (DeepFleet and robotics), INFORMS (UPS ORION), NRF and Happy Returns (2025 Retail Returns Landscape), Verisk CargoNet and FBI IC3 (cargo theft), DCSA (eBL standards and adoption), FMC (detention and demurrage), ATA and ATRI (trucking), Pitney Bowes (Parcel Shipping Index), Mordor Intelligence, GMInsights, Grand View Research, Precedence Research, Fortune Business Insights, Interact Analysis, MHI, and BLS.

Where sources conflicted, ranges are given rather than a single figure. Market sizing across research firms varies by more than 3x for the same category and year, which is itself worth knowing.