How to survive the AI divide
An AI divide is here.
S&P 500 operating margins have jumped from 14 percent to 17 percent in five years, heading toward 19, but the gains are concentrated. Morgan Stanley found that the most forward-thinking AI adopters are getting twice the margin expansion of firms who are less aggressive in optimizing their business around this transformative technology.
What separates AI-native enterprises from those that merely dabble? Many firms are experimenting with AI on a small scale, but the returns are only incremental. There is only so much value that can be gotten from using generative AI to summarize emails, generate documents or prepare for meetings.
Meanwhile, a handful of forward-thinking companies are doing something much grander. The enterprises gaining disproportionate advantage from AI are actually redesigning their entire operating models around intelligence, action, operations and trust.
These AI leaders are realizing the scale of what AI makes possible. They are coming to the conclusion that AI is potentially so transformative that it asks something difficult of business leaders: the foresight to reorganize a system that already seems to work, the boldness to invest before returns are obvious and the communication to bring an entire organization along.
Does this all sound terrifyingly new? In fact, it’s history repeating itself. For guidance in brave new worlds, it can help to revisit the brave old ones.
In 1882, Thomas Edison’s Pearl Street Station in lower Manhattan began pumping electricity to customers. The first killer app was, naturally enough, the light bulb—on factory floors, it replaced oil lamps that started fires and caused injuries. And thanks to the light bulb, factories got safer and a little bit more efficient. Most factory-owners stopped there, content with these modest gains.
But a few manufacturers went much further, glimpsing that it was possible to use electricity not just for incremental gains, but for transformative ones. The key innovation was a so-called “unit drive”—a small electric motor built into each individual factory machine. Before unit drive, machines were tethered to a central rotating shaft, forcing factories to stack vertically. But in a new world where electricity infused each individual machine, machines could go anywhere. Plants no longer had to be built vertically, at great cost—they could spread along a single floor.
The early pinnacle of this new way of using electricity—re-architecting the factory from the ground up, rather than just installing new light fixtures—was Henry Ford’s famed Highland Park plant, in 1913. Ford’s moving assembly line, which was fundamentally enabled by the freedom electricity afforded, ultimately cut Model T production time from over 12 hours to just 90 minutes.Ford was able to slash Model T’s price from $850 to $260; by 1914, Ford was producing more cars than every other auto maker combined.
Bold business leaders today are taking a page from the early era of electricity. Not content to remain at “lightbulb-style” applications of AI, they are realizing that this novel technology is potentially so powerful that it may demand a tearing up of the factory floor. Put another way, they are re-architecting how they conduct business in the era of AI. There are four enabling systems that support what we might call the AI Operating Model.
The first enabling system is intelligence. The underlying point is that AI is only ever as good, useful and helpful as the freshness and relevance of the data that fuels it. There is no sense shooting your organization through with AI, if the data the AI will be drawing on is routinely stale or siloed.
Next comes action. Increasingly, enterprises are trusting AI not just to deliver insights, but to act on them. If 2025 was the year of the agentic AI pilot, 2026 is increasingly becoming the year where more and more firms are trusting agentic AI with increasing autonomy.
Next comes operations. What is the strategy by why a pilot is deemed a model that should expand into a second pilot, then a third, and finally into production across an organization? When and to what extent are humans to remain in the loop? The most AI-savvy firms are making an art and a science of the skillfully controlled expansion of AI actions at scale.
Finally comes trust. How can firms trust the AI? And how can firms’ customers trust that trust? What happens when something, as it must, goes wrong? What happens when laws change, or when a regulator comes knocking at our door? Not only must firms dial in the pace of their operational expansion; they must do so with partners, platforms and tools that are helping prevent mishaps—and correct them when they inevitably arise.
About 15 years ago, Jay Kreps was an engineer at LinkedIn when he ran into a problem. Whenever he wanted to do anything sophisticated with the company’s data, the process was archaic: He had to extract the data at the end of the day, run processing jobs overnight, then plug the results back in the next morning—a process known as batch processing. “By the time you get your results,” Kreps said later, “its tomorrow, and the customer’s gone.”
“Companies are happening continuously,” as Kreps puts it today. “They should be able to react to data continuously.” So his team built a system to do exactly that—what’s now called data streaming. With data streaming, business events (a customer complaint, a payment, a sensor reading) flow to the systems that need them, the instant they happen.
Kreps’s system grew to become Kafka, one of the most widely adopted technologies in enterprise computing. Kafka in turn became the foundation of the company Confluent (where Kreps is co-founder/CEO), which IBM acquired in March.
What problems does real-time data fix? Kreps offers one vivid example from a client. Formerly, at Citizens Bank, the nearly 200-year-old institution was running on batch systems. When a customer deposited money or made a payment, the transaction wouldn’t appear in the bank’s mobile app until processing had finished. Customers didn’t interpret the delay as a minor technical limitation; they thought the app was outright broken. Switching to real-time streaming fixed the problem.
But fixing the app, it turned out, was just the quick win. In committing to a new infrastructure undergirding intelligence through data that is up-to-the-second, Citizens Bank also brought gains to unexpected corners of the business; for instance, it has also improved fraud detection and cut processing costs. Customer satisfaction scores at Citizens Bank ultimately jumped 20 points, says Kreps.
Tight integration of Confluent’s data-streaming with IBM’s watsonx.data platform now enables operators to drill down on data, asking broader questions. Which customers are happiest or least happy? And why might that be? Which branches are having the most problems? Questions like these can be asked in plain English, rendering every employee a potential data analyst and business strategist. Intelligence is everywhere all at once.
But real-time data is only as valuable as what one can do with it.
Kreps demonstrates this with a food delivery scenario. Everyone has had the experience of ordering food, but 30 minutes later receiving something that doesn’t quite match the order. The app maker wants to solve the problem quickly to keep the customer happy.
In a batch-processing world, a customer complaint gets routed to a human who must check all relevant data points manually (is this a customer who always complains, gaming the system?), leading to long wait times. With real-time streaming, an AI agent can assess the situation instantly—pulling in the delivery status, the driver’s track record, and the customer’s history—and decide whether to issue a refund or flag potential fraud. The instant data enables instant (and intelligent) action.
The same logic for the relatively low-stakes scenario of food delivery also applies to higher-stakes scenarios.
Imagine a compliance officer at a financial firm belatedly discovers that a server in AWS has been missing a security control required by GDPR—the European Union’s data privacy law—for months. Customer data may have been exposed. The clock on disclosure is ticking; potential fines are piling up.
Services like IBM’s InfraGraph can now map an organization’s entire infrastructure—a sort of live X-ray of every server, application and connection. In such an instance, InfraGraph would flag a server missing a specific GDPR control.
Increasingly, IBM tools go further than just flagging: IBM’s Concert (its agentic IT operations platform) can also auto-generate a code fix (as a “pull request,” i.e. a proposed change that an engineer could review and approve).
A picture begins to form, of a digital estate that is constantly self-scanning and self-correcting, no matter its size.
A single automated action is a helpful starting point—but what is the art by which a firm decides it’s time to scale up to automate actions in the hundreds? In the thousands? In the millions?
In 2024, the question haunted Detlev Klage. Klage is the Vice President of Finanz Informatik, which runs IT infrastructure for all German savings banks—with 20 million users on the country’s top banking app. That year, European regulators made a big decision that affected Klage profoundly. Instant bank transfers (already routine in the United Kingdom, India and Brazil) would now be mandatory in Europe; not only that, but regulators would not be relaxing any anti-fraud requirements (even though instant transfers carry more fraud risk).
Klage knew the problem needed to be tackled with AI. His main question: what would be the responsible pace at which to scale from individual actions to full-scale operations?
Klage decided to begin with a relatively simple pilot, focused on risk-scoring, and with humans initially in the loop approving the AI’s work. Then Finanz Informatik graduated to real-time verification in production—classifying each transaction automatically. Only after proving the model on one high-stakes problem did the firm expand to other use cases.
That expansion is now accelerating. The company is opening its 1,600 core banking functions—moving well beyond the instant transfer problem—to AI agents. One feature already in production: a Finanz Informatik bank employee simply says “prepare my meeting,” and an AI agent pulls the customer’s account history, recent transactions, open issues and relationship notes—assembling in seconds what used to take 20 minutes of prep work.
It’s one of many use cases soon to deploy. The point, though, is the pace—Finanz Informatik did not blunder into cases that hadn’t been proven, nor did it remain caught in the endless-pilot stage. It expanded quickly yet deliberately, balancing risk and reward.
Once an operator trusts scaling up AI actions into operations, how are others brought along? How is trust built with all stakeholders?
Gaetan Willems kept hearing the same questions from his customers. Willems leads the sovereign technology effort at Cegeka, a Belgian company that manages IT systems for 2,500 customers in highly regulated European industries. The questions he heard: We want AI in critical business processes, but what happens if something goes wrong? What if the vendor changes its terms, if a government demands access to data stored on a foreign cloud? Without guardrails in place, adopting AI meant accepting a dependency that many boards were uncomfortable with.
The answer, per Willems, increasingly involves what’s called digital sovereignty, an idea that has evolved beyond simply where data is stored, increasingly encompassing questions like who runs the platform, how open the technology is and who has access.
For years, the notion of digital sovereignty remained abstract, a policy debate with no agreed-upon way to measure it. No longer: in October 2025, the European Commission published its Cloud Sovereignty Frameworkwhich scores cloud services across eight objectives on a scale from “SEAL 0” (no sovereignty) to “SEAL 4” (full digital sovereignty). What had been political rhetoric has become, increasingly, an engineering specification and procurement guideline.
Cegeka, beset with both customer concerns and regulators’ guidance, chose to use IBM’s the newly released Sovereign Core—a full technology stack that organizations can deploy on their own premises, in an air-gapped facility or on any cloud. Sovereign Core ships with 160 preloaded compliance frameworks that run continuously—a shift from annual audits.
Cegeka’s Willems can now tell fretful clients that trust is engineered into the operating environment: the control plane, identity, encryption keys, logs, audit evidence and AI workloads can all run inside a defined boundary. AI tools, models and agents can be made available through a governed catalog, so developers can move quickly without sending sensitive data or prompts outside approved systems. Trust isn’t an afterthought; it’s baked in from the start.
Feeling the full effects of the “AI divide,” like the “electricity divide” before it, may well take years. But for many business leaders, the time to act is now. When a fault line opens, there is only so much time to safely leap across.
The economists who have studied factory electrification estimate that by the 1920s (a decade of dramatic productivity growth), up to half of the acceleration in manufacturing productivity growth was due to transformative use of electricity. But the greater part of that growth went specifically to the firms bold enough to realize that electricity enabled—no, required—an entirely new way of operating. Those who failed to innovate had to content themselves with a smaller share of a growing pie.
The same thing is happening today, and the signal is there in the S&P data showing concentrated gains from firms being more aggressive with AI. Who will survive and thrive in the AI era? If history is any indication, it’s those willing to rethink their operations from the ground up.