Hello and welcome to Eye on AI. In this edition:
How Fortune 500 logistics firm C.H. Robinson became an AI success story
Apple sues OpenAI for theft of trade secrets
Economists urge policymakers to take the threat of AI seriously
A new method for making frontier AI models safer
Is data the new bottleneck to AI progress?
There are a lot of Fortune 500 C-suite executives who still complain about not being able to get ROI from AI. Dave Bozeman isn’t one of them.
The CEO of C.H. Robinson Worldwide, a 120-year old logistics company headquartered in Eden Prairie, Minnesota, says the company’s use of AI has resulted in a 45% uplift in employee productivity since 2022. Its use of AI has helped the company deliver double-digit earnings-per-share growth since 2023, despite a post-COVID slump in global shipping that has seen the company’s revenues drop some 34% over the same period.
Robinson, as the company is commonly known, is primarily a freight broker, specializing in what the industry calls LCL (less-than-container load) freight. The company now deploys hundreds of AI agents across different aspects of its business. A believer in “Lean management”—a system initially developed in Toyota’s manufacturing plants that focuses on maximizing customer value and eliminating waste—Bozeman, who has been Robinson’s CEO for the past three years, deployed teams to map out workflows and processes. Any tasks that didn’t add value were eliminated. Those that were essential but highly-routinized and repeatable, they’ve automated with AI agents. For example, these agents now deliver quotes to customers, a process that once took human specialists 20 minutes, in just 31 seconds—and they operate around-the-clock, 365 days a year.
“It provides us not just productivity,” Bozeman tells me. “This is revenue growth, margin expansion, productivity as well as customer advantage.” He says by speeding up the time it takes to give customers quotes and providing more information to the customer, customers are more likely to submit jobs for quotations to Robinson, giving it more chances to win business.
Moving employees up not out
Like many executives, Bozeman is at pains to say his company’s embrace of AI isn’t about replacing human workers. He says the company has been moving the shipping specialists who once provided quotations into higher-value work, like helping customers navigate shifting tariff regimes. But that doesn’t mean there hasn’t been some labor savings. Bozeman said the business had a natural employee turnover rate of 11% to 14% each year, and the use of AI agents means that Robinson has not had to hire new workers to replace those who have left. The AI agents mean that for certain aspects of what Robinson does, such as providing those customer quotations, headcount is now largely divorced from volume in a way that was never possible before.
AI is also letting Bozeman contemplate strategic moves that the company might have struggled to execute previously. Ultimately, his vision for Robinson is to be more than just a freight forwarder and shipping broker. He wants the company to move towards being a supply chain consultant, and perhaps ultimately taking on the entire supply chain function for its customers. “Think about it as ‘supply chain in a box,’” he says. “I want to get to the point where a customer would say it’s going to be irresponsible not to do business with C.H. Robinson, and it will be irresponsible for us to actually have a supply chain department. Why do we need that when we have this company that can really do that, do it better than us, and allow us to focus on our core?”
Bozeman is also focusing more on serving small and medium-sized customers, an area where Robinson has lost market share in recent years. Now, the CEO sees an opportunity to grab some of that back, with human sales reps assisted by AI agents. In both of these domains—the high-value supply chain consulting and the servicing of more SMEs—Robinson is hiring more employees, Bozeman says. It’s just that those workers have AI assistants helping them surface the insights their customers need.
Build don’t buy
How has Robinson been able to deploy all these AI agents without incurring crushing token costs? The answer, Bozeman says, is that it has built almost all of them in-house using its own AI models or open-source models. The company employs some 450 engineers, most of whom are steeped in the shipping industry—domain knowledge that Bozeman says has enabled the company to build better models than any third-party vendor could ever supply at a fraction of the cost. Bozeman says that the company is currently “getting hundreds of millions of dollars of benefit with a token cost of less than $2 million.” “This is a deep, wide moat,” he says. “We calculated that if you wanted to replicate what we’re doing here, you would have to partner with 15 to 20 different entities to do that.”
A key to Robinson’s success in building these in-house AI models, he says, has been the operating mode he’s brought to Robinson. When figuring out what agents to potentially build, Bozeman assembles cross-functional teams consisting of engineers, operational domain experts, and people from business departments like finance and legal. He poses questions to them using the Socratic Method and has them debate solutions. “That’s priceless when it comes to discovery. It’s priceless when it comes to ingenuity,” he says.
There’s no success like failure
He also credits the AI success to other aspects of a cultural transformation he’s tried to implement at the company. His teams use a FMEA (Failure Mode & Effects Analysis) methodology to game out how the AI systems they are building might fail and to mitigate those risks. Bozeman has also pushed Robinson’s employees to embrace failure as a waypoint on the path to success. “Failure is part of what we do,” he says. He notes that when his teams report progress towards goals, they use a modified “traffic light” methodology that only allows two colors: green (on track) or red (off track.) There’s no yellow; Bozeman says ‘yellow’ is usually really a red but the manager is afraid to say so. Instead, he has tried to take the fear out of reporting a red. “We say we celebrate the red. If you’re red, you get the full weight of this organization to get you back to green,” he says. “But you have to think about it, and how you problem-solve to get back to green is super important.”
It’s something I hear a lot from executives who report success deploying AI at scale: success is never just about the technology or about engineering talent. It’s about operational design and culture too.
With that, here’s more AI news.
Jeremy Kahn
[email protected]
@jeremyakahn
Before we get to the news, just a reminder to check out our new vodcast, Fortune AI Weekly. This week, Bea Nolan and I break down major global AI developments, including the policy implications of OpenAI’s GPT-5.6 rollout, backlash over Meta’s smart glasses, and Illinois’ landmark AI safety law. We also discuss China’s proposed open-source restrictions and new Anthropic research reigniting the debate over AI consciousness. You can check out the vod here on YouTube.
FORTUNE ON AI
Apple accuses OpenAI, and former design star Jony Ive’s io Products firm, of stealing hardware trade secrets in blockbuster lawsuit—by Sebastian Herrera
Stolen laptops, data breaches, secret moles, and recruiting-as-espionage. Here are the wildest claims in Apple’s lawsuit against OpenAI—by Emily Forlini
Fidji Simo steps back from OpenAI—and exposes the fragile hold women still have on power—by Emma Hinchcliffe
OpenAI’s latest AI model likely has similar cyber vulnerabilities to one that led to U.S. export controls on Anthropic’s Fable, British agency says—by Emily Forlini and Jeremy Kahn
Exclusive: Google’s former ‘click fraud czar’ emerges from stealth with an on-device AI shield against AI-powered phishing, deepfakes, and other scams—by Jeremy Kahn
Companies are shifting toward cheaper open‑source AI models to rein in costs, Amazon CTO says—by Beatrice Nolan
AI IN THE NEWS
More than 200 economists, including Nobel winners, call for policy makers to urgently prepare for AI-driven economic impacts. The economists, who included Nobel laureates Daron Acemoglu and Paul Krugman as well as the chief economists of OpenAI and Anthropic, signed an open letter urging policymakers to do more to prepare for AI’s potentially transformative economic effects. The letter warned that AI could reshape the economy more dramatically and far more quickly than the Industrial Revolution, bringing both major productivity gains and the risk of widespread job displacement. Rather than predicting a specific outcome, the letter calls for more research, as well as new incentives, guardrails, and institutions to ensure AI complements human workers and benefits society. Acemoglu’s presence among the signatories is significant because the prominent MIT economist had previously voiced skepticism that AI would have profound economic impacts. You can read more from the New York Times here.
New York issues a one-year moratorium on data center construction. New York Governor Kathy Hochul issued a one-year executive order halting approval of new hyperscale data centers requiring 50 megawatts or more of power, the New York Times reported. The decision makes New York the first U.S. state to impose such a statewide pause. The move reflects growing concerns over AI-driven data centers’ demands on electricity, water, and public infrastructure, and will require future projects to help fund grid upgrades. Facilities that already have permits are exempted. The moratorium was praised by environmental groups and some lawmakers but criticized by tech industry groups and construction unions, which warned it could cost jobs and slow AI-related investment.
Meta plans to start making a new in-house AI chip. That’s according to a story from Reuters, which cited an internal company memo it obtained. The chip is code-named Iris and will begin production in September as part of an ambitious push to double Meta’s AI computing capacity to 14 gigawatts by 2027. The in-house chip will reduce Meta’s current dependence on Nvidia and AMD GPUs. The Broadcom-designed, TSMC-manufactured Iris is part of Meta’s planned four-generation series of in-house AI chips for both training and inference, with the company planning to release a new version every six months through 2027.
Google DeepMind CEO calls for U.S.-led, industry-funded AI standards agency. Demis Hassabis, the Google DeepMind cofounder and CEO, wrote a blog post calling for the U.S. to establish an industry-funded, federally overseen standards body to evaluate frontier AI models for risks including cybersecurity, biological threats, and deceptive behavior before deployment. He said the new agency could be modeled on the Financial Industry Regulatory Authority and that it could delegate model testing to the U.S. national laboratories in areas directly pertaining to national security. He proposed voluntary pre-release testing that could eventually become mandatory for the most capable models, saying global coordination on AI safety is essential if society is to realize AGI’s benefits while avoiding its most serious risks.
OpenAI’s safety head departs. Johannes Heidecke, OpenAI’s head of safety systems, is leaving the company following a reorganization that merges the company’s safety and research teams, Wired reports. Mia Glaese, vice president of research, is assuming expanded responsibility for both areas, while Saachi Jain will serve as interim head of safety systems. Heidecke’s departure comes along with other leadership changes at OpenAI. “Chief Futurist” Joshua Achiam announced his departure, while Fidji Simo, CEO of Applications, announced she was leaving for medical reasons (see ‘Fortune on AI’ section above.)
EYE ON AI RESEARCH
Want to make AI models safer? Switch off the dangerous bits. That’s the idea behind a new method pioneered by researchers at AE Studio, a startup product design and AI studio, working with Anthropic. The technique lets a single AI model be built with « switches » for its riskiest knowledge—like detailed virology or cybersecurity know-how—so that these capabilities can be turned on for vetted, trusted users and turned off for everyone else. That solves the current problem, where models have to be given guardrails, such as training them not to answer these kinds of questions, or protected by classifiers that filter out prompts trying to ask the model these sorts of queries, but these protections can be “jailbroken,” giving users access to the risky capabilities. Another solution is to train entirely separate models with the risky data stripped out — which is effective, but expensive to do repeatedly.
The new method, called GRAM, works by routing sensitive topics into separate, removable mini-modules inside the model during training. In tests spanning models from 50 million to 5 billion parameters, a single GRAM-trained model was able to mimic several separately-trained, « filtered » models at once, and stayed just as reliable even when researchers tried to fine-tune the dangerous knowledge back in. The approach isn’t yet used in any live products—the authors call it early-stage work—but it points to a future where AI labs could sell one model with different capability tiers unlocked for different customers, instead of training and maintaining a whole fleet of separately restricted models. You can read the research paper on the AE Studio website here.
AI CALENDAR
Aug. 4-6: Ai4 2026, Las Vegas.
Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.
Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.
Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.
BRAIN FOOD
Do we really need a Stargate for Data? A few weeks ago ex-OpenAI researcher Will Depue wrote a long essay-like X post in which he claimed data, not compute, was becoming the real stumbling block to AI progress. In the past, most AI labs got the majority of their data for free, by simply scraping the internet. But today’s massive frontier models require much more data than that—and also much more specialized data from specific professional domains that is often difficult and expensive to obtain.
Depue argues that data, not compute, has become the real bottleneck to further AI progress. (He notes that data collection costs are accelerating rapidly, although they remain well below what AI companies are spending on compute.) Countries that want to excel in AI should invest in national data collection strategies, he argues. He also postulates that perhaps AI labs will start purchasing other businesses and operating them at a loss simply to harvest the data they generate.
It certainly sounds provocative. But there are some real problems with what Depue is saying. He is not the first to say data has become the bottleneck to AI progress. AI luminary Ilya Sutskever said pretty much the same thing on the Dwarkesh podcast in November 2025. But unlike Depue, Sutskever didn’t conclude that we just had to find more ways to harvest new pools of human-generated data in order to keep scaling our current large language models. Instead, he suggested that we needed to find new architectures that would be much more data efficient. He called it a return “to the age of research.” Depue also dismisses far too easily the idea of using reinforcement learning in simulated environments and “self-play” to generate completely new data from which an AI model can learn. That’s the bet that former Google DeepMind researcher David Silver is making with his new startup Ineffable Intelligence. Depue doesn’t really engage with either of those ideas. So if you were hoping to sell your taco stand to OpenAI “for the data”—you might want to come up with a Plan B.
