The larger competitive advantage comes from turning cheaper thinking into faster experiments, decisions and improvements that customers value.

A CEO can make a great deal of money using artificial intelligence to reduce labor costs. We should acknowledge that before arguing about what comes next.

In June 2025, C.H. Robinson reported that its AI orders agent was interpreting emailed shipping requests and building orders in about 90 seconds. The company said the agent handled 5,500 truckload orders a day and saved 600 hours of daily labor. Those are company estimates of work saved, but the economic opportunity is easy to see.

Now imagine two competitors capturing similar savings. One improves its margins. The other also invests in understanding customer problems, testing solutions and changing how the business works. Over successive cycles of improvement, their capabilities can diverge even if they use the same models.

That is the strategic issue I’m exploring in my book, The Generative Organization. AI gives companies a chance to learn faster than the competition on things that matter to customers. Labor savings can help finance that advantage.

Cheaper thinking changes what a business can attempt

Computing has been getting cheaper for decades. What makes this period different is the range of thinking that can become economical.

People can describe what they need in their own language, inspect a result and refine it while the details are still fresh. An expert’s knowledge has a shorter journey into a working artifact. Models can also connect words, images and other forms of information that were expensive to analyze together. And they can explore enough alternatives to make previously impractical experiments worth trying.

A field experiment at Procter & Gamble illustrates the potential. Researchers assigned 776 experienced professionals to product-innovation tasks, individually or in pairs, with or without AI. Individuals using AI performed comparably to teams without it. Commercial and R&D specialists using AI also produced proposals that more evenly combined technical and commercial considerations. The study concerned bounded innovation tasks; it does not establish that all teams are replaceable.

For a CEO, the interesting possibility is that more people can consider a problem across functional boundaries. Useful expertise becomes easier to bring into a decision.

FM Logistic pursued a related opportunity in warehouse operations. It supplied Google’s AlphaEvolve with an existing routing algorithm and an evaluation based on actual picking tours. The system generated alternatives and tested them against operational constraints. FM Logistic and Google reported a 10.4% improvement in routing efficiency beyond the previous best solution, with the pilot subsequently running in production.

A company’s method for doing the work can itself become a subject of regular experimentation. Think of the experimental line in a manufacturing plant: a place to improve the core operation. Consulting, customer service, purchasing and other information-intensive activities need that capacity too.

The management cycle has to accelerate too

Faster analysis alone will not produce a faster company. In our management information systems teaching at Harvard Business School, we distinguished the operating cycle from the management cycle.

The operating cycle is how work gets done: an order is fulfilled, a customer is served, a product is made. The management cycle is how the organization recognizes a need for change, makes a decision, allocates resources and authorizes a response.

If customer complaints can be analyzed daily but changing the policy requires a quarterly meeting, the delay has moved. If management makes decisions faster but operations cannot implement them, decisions simply accumulate. Both cycles must adjust.

Hapag-Lloyd provides a useful example . Its AI system ingests customer feedback daily and produces reports for the teams’ two-week planning rhythm. Recurring requests for a preview function in Shipping Instructions helped the team prioritize and release that feature. Later feedback indicated that the specific request had been addressed. This is a company account of an improvement, not a controlled study of revenue or retention.

The significance is the connection between listening and changing the service. A dashboard by itself cannot make that connection.

I have encountered companies that welcome islands of automation, then expect the people who created them to undertake broader redesign off the side of their desks. The team has found an opportunity, but there is no clear route to funding, expertise or decision authority.

Major transformation cannot be staffed indefinitely as extracurricular work.

An employee close to the customer should be able to invoke an improvement process and summon support. The source of the problem may sit several departments away. AI can help investigate it; management must make it possible to act.

Budget for learning, and measure the whole journey

I would dedicate a portion of the AI budget to experiments that improve the business. For operating experiments, a reasonable ambition is a portfolio that pays for itself. Some attempts will fail. Others should generate enough value to cover those failures and the cost of learning.

Count evaluation, expert time, integration and implementation alongside the model bill. A demonstration that looks promising is the beginning of the economic test.

Longer-term R&D needs a different investment logic. A call-center experiment may produce evidence in days. A scientific discovery may take years to become a commercial product. Judge each against its potential business impact and the time over which value can reasonably emerge.

Then measure the journey from customer signal to implemented improvement. How long until the information is usable? How long until someone can decide? How long until the change reaches the field? Did it actually help the customer? Retention, share of wallet, revenue per employee and asset utilization can show whether those improvements are translating into business results.

Finally, protect the expertise that makes the system useful. Experts must remain involved in updating its knowledge. Novices need time to develop judgment. I call one approach Socrates mode: have the model question a learner, introduce exceptions and ask for a defense of the reasoning against expert standards. Allocate time for it; people serving customers should not have to turn every transaction into a training session.

The competitive advantage lies in what a company does with the capacity AI releases. A rival that repeatedly discovers better ways to serve customers can become a different kind of competitor.

You can spend a great deal of ingenuity optimizing a clipper ship in the age of steamships. Make sure some of your AI investment helps you discover what the customer will need next—and equips your organization to build it.