C.H. Robinson: 45% Productivity Gain From Hundreds of In-House AI Agents

C.H. Robinsonthe 120-year-old logistics giant headquartered in Eden Prairie, Minnesota, built hundreds of AI agents in-house and posted a 45% productivity gain since the end of 2022cutting customer quote generation from 20 minutes to 31 seconds along the way. CEO Dave Bozeman detailed the results in a Fortune interview published on July 14, 2026: an annual token bill below $2 millionset against “hundreds of millions of dollars of benefit.”

Before the agents: 20-minute quotes in a shrinking freight market

C.H. Robinson, listed on NASDAQ as CHRW, runs one of the largest logistics networks in the world, brokering freight between shippers and carriers across hundreds of thousands of transactions. For decades, the core of that business was manual knowledge work at enormous volume. Freight orders arrived as unstructured emails that human brokers opened, sorted and classified by hand. Pickup and drop-off appointments required phone calls and follow-ups. A single customer quote absorbed around 20 minutes of an experienced broker’s attention, and requests landing outside office hours sat in a queue until the next morning.

The market backdrop made that cost structure untenable. The freight recession that began in 2022 dragged revenue down 34% from its peakcompressing margins across the entire sector. Management faced a classic squeeze: service expectations kept rising while pricing power evaporated. The company chose to attack cost per transaction with AI, because its most expensive workflows shared a precise profile, high-volume, routine, rules-heavy and text-based. That profile is exactly the terrain where large language models perform best.

The decision: 450 in-house engineers instead of a vendor contract

Most enterprises buy AI through procurement: a vendor platform, a system integrator, a rollout plan. C.H. Robinson inverted that playbook. The company put roughly 450 in-house software engineers and data scientistsmost with years of shipping-industry experience, to work building agents directly on proprietary and open-source models. Cross-functional teams paired those engineers with operational domain experts, finance staff and lawyers, so every agent encodes how freight actually moves rather than how a demo imagines it.

The engineering culture borrows from lean manufacturing. Each agent passes FMEA, Failure Mode and Effects Analysis, a discipline lifted from industrial quality control, before touching production. Progress reviews proceed by Socratic questioning and use a two-color traffic light: green or red, with yellow deliberately abolished to force honest status calls. The company’s own newsroom describes more than 30 specialized agents deployed along the shipment lifecycleoperating inside strict guardrails, with human oversight reserved for critical judgments and closed-loop learning fed by 100 trillion data points. Bozeman’s reasoning is blunt: software is purchasable; institutional knowledge, proprietary data and operational muscle grow from within.

The result: 45% productivity, double-digit EPS growth, a $2 million token bill

Four years into production, the numbers are unusually clean. Productivity across the business is up 45% since the end of 2022. The quoting agent compresses a 20-minute task into 31 seconds and answers around the clock, seven days a week, so after-hours requests get priced the moment they arrive. Agents now open and sort emailed freight orders, classify freight and book pickup and drop-off appointments at a scale that would demand far larger human teams.

The financial trail is visible in public filings. Adjusted earnings per share have grown at double-digit rates since 2023during a period when revenue fell 34%, and the first quarter of 2026 delivered 15% year-over-year adjusted EPS growthwith management projecting double-digit productivity improvements across both its North American surface transportation and global forwarding units. The cost side is the striking part: Bozeman puts cumulative token spend below $2 millionagainst benefits he values in the hundreds of millions of dollars. That ratio, roughly two orders of magnitude, is the cleanest cost-benefit figure any large enterprise has published this year.

The workforce story deserves honest framing. Headcount fell through natural attrition of 11% to 14% a yearwith vacated positions left unfilled; Fortune stresses the absence of mass layoffs. Local reporting adds nuance: the Minneapolis Star Tribune documented cuts among high-level managers tied to the AI-driven reorganization. Both facts hold together, the reduction was real, and it was managed gradually. Displaced specialists moved to higher-value work such as tariff navigation, and the company keeps hiring in supply-chain consulting and in services for small and medium-size shippers.

What other organizations can learn

Three lessons transfer. First, domain depth beats model novelty: C.H. Robinson’s agents work because 450 engineers who understand freight encoded decades of institutional knowledge and proprietary data into them, while competitors bolt generic tools onto workflows they master far less deeply. Second, inference is cheap where workflows are routine and voluminous: a sub-$2 million token bill against hundreds of millions in benefit means the binding constraints are process knowledge and engineering capacity rather than compute budgets. Third, manufacturing discipline turns probabilistic models into dependable operations, FMEA before deployment, guardrails in production, red-green reporting that forbids ambiguity.

The conditions for replication are equally clear: high-volume, rules-heavy workflows; proprietary data accumulated over years; an engineering bench that sits close to the business; and executive patience measured in years rather than quarters. Bozeman calls the outcome “a deep, wide moat” and estimates that a rival assembling the same stack from vendors would need 15 to 20 separate partners. For the many industries that run on repetitive, document-heavy transactions, the C.H. Robinson case reads less like an outlier and more like a blueprint.