C.H. Robinson’s 90-second AI agent is not the moat

C.H. Robinson says its AI agent can turn emailed freight requests into truckload orders in about 90 seconds, processing 5,500 orders a day and saving 600 hours of labor daily. The figures are the company’s own estimates, but the strategic point is broader: cheaper execution is not the same as a stronger business. If rivals use similar models to cut comparable costs, the advantage will go to the company that uses the savings to learn faster, test more ideas and change how it serves customers.

This distinction sits at the center of the argument in The Generative Organization. AI can reduce the cost of thinking, not just the cost of routine execution. That gives companies a chance to run more experiments, involve more expertise and improve their operations in cycles that were previously too expensive.

The resulting gap may appear even when competitors use the same models. One company can simply add the savings to its margin. Another can use the money and freed-up time to investigate customer problems, test solutions and redesign its business. After several rounds, the two companies may have very different capabilities.

The most important change is the widening range of work that becomes economical. People can describe a need in ordinary language, inspect the result and refine it while the details remain fresh. Expert knowledge can become a working output more quickly. Models can also combine text, images and other data types whose joint analysis was once costly. They can generate enough alternatives to make previously impractical experiments worth attempting.

A Procter & Gamble field experiment illustrates the potential without proving too much. Researchers assigned 776 experienced product-innovation professionals to different conditions: some worked alone, some in pairs, some used AI and some did not. Individuals working with AI produced results comparable to teams without AI. Commercial and research-and-development specialists using AI also produced proposals that balanced technical and commercial considerations more evenly.

The study covered a limited set of innovation tasks. It does not show that any team can be replaced. Its more useful lesson is organizational: more people may be able to examine a problem from several functional perspectives at once, making relevant expertise easier to bring into a decision.

FM Logistic used a related approach in warehouse operations. The company gave Google AlphaEvolve an existing routing algorithm and an evaluation system based on real picker routes. The system generated alternatives and tested them against operational constraints. FM Logistic and Google reported a 10.4% improvement in routing efficiency compared with the previous best solution. After the pilot, the system continued operating in production.

That example points beyond automation. The work itself becomes an object of repeated experimentation. Manufacturing has long used experimental lines to improve a core operation. Consulting, customer service, procurement and other information-heavy functions need an equivalent capability.

But faster analysis does not automatically make a company faster. In their teaching of management information systems at Harvard Business School, the authors distinguish between the operating cycle and the management cycle.

The operating cycle is the work itself: delivering an order, serving a customer or producing a product. The management cycle is how an organization notices the need for change, makes a decision, allocates resources and authorizes action.

If customer complaints can be analyzed every day but changing a rule requires a quarterly meeting, the delay has merely moved to another stage. If executives decide quickly but operating teams cannot implement the decisions, the organization accumulates decisions instead of improvements. Both cycles have to move faster.

Hapag-Lloyd offers a practical example. Its AI system receives customer feedback daily and prepares reports for teams’ two-week planning cycle. Repeated requests for a preview feature in Shipping Instructions helped the team prioritize and release that feature. Later feedback showed that the specific request had been addressed. This is the company’s account of an improvement, not a controlled study of revenue or customer retention.

The useful part is the connection between collecting feedback and changing the service. A dashboard alone does not create that connection.

The question the announcement is quiet about is what happens after an automation project finds an opportunity. The author has encountered companies that welcome isolated automation efforts but expect the employees who built them to redesign the wider business in their spare time. The team identifies a possibility, then discovers that it has no clear route to funding, expertise or decision-making authority.

A major transformation cannot remain side work indefinitely. An employee close to the customer should be able to initiate an improvement process and bring in the support required to pursue it. The source of a problem may sit in another department. AI can help diagnose it, but management must make the next action possible.

That changes how an AI budget should be judged. Some of it should fund experiments intended to improve the business. For operational experiments, a sensible target may be a portfolio that pays for itself: some attempts will fail, while others should create enough value to cover the losses and the cost of learning.

The accounting must include evaluation, expert time, integration and deployment, not just the model bill. A promising demo is the beginning of the economic test, not its conclusion.

Longer-term research and development requires a different investment logic. A call-center experiment may produce data within days. A scientific discovery may take years to become a commercial product. Each project needs to be assessed against its potential business impact and the period before results might appear.

Companies should measure the full path from a customer signal to an implemented improvement: how quickly information becomes usable, how long a decision takes, when the change reaches practice and whether it actually helps the customer. Customer retention, the share of a customer’s spending captured by the company, revenue per employee and asset utilization can indicate whether operational improvements are becoming business results.

They also need to preserve the expertise that keeps the system useful. Experts should participate in updating its knowledge, while newcomers need time to develop professional judgment.

The author calls one approach the Socratic mode: the model asks the learner questions, proposes exceptions and asks the learner to defend the reasoning against expert standards. That requires dedicated time. Employees who work with customers should not have to turn every transaction into a training exercise.

AI therefore creates a choice that a labor-saving calculation conceals. A company can keep the benefit as margin, or spend part of it building a faster system for learning what customers will need next. The second path carries costs and failed experiments, but it can alter the company’s capabilities rather than merely its expense base.

In the age of steamships, a great deal of ingenuity could be spent optimizing a clipper. The same danger exists now: companies may become exceptionally efficient at a form of work whose replacement is already becoming cheaper. The durable advantage will belong to organizations that use AI savings to discover the next service and build the authority, expertise and operating machinery to deliver it.