The AI Divide Will Be Between Organizations That Learn and Organizations That Report

In 1988, Arie de Geus, then head of planning at Royal Dutch Shell, wrote in Harvard Business Review that "the ability to learn faster than your competitors may be the only sustainable competitive advantage." Peter Senge developed the idea further in The Fifth Discipline in 1990. For many organizations, it has remained a widely accepted principle rather than an operating practice. Broad access to AI is likely to change that, because it removes several other sources of advantage.

Why using AI will not differentiate

In McKinsey's 2026 State of AI survey, nearly nine in ten respondents said their organizations use AI regularly in at least one business function. The leading models are available to most companies at similar prices, and those prices continue to fall. As with computers a generation ago, using AI is becoming a basic requirement rather than a source of advantage.

What remains difficult to copy is an organization's ability to absorb new capabilities, learn from using them and adjust how it works as a result. That ability is where the differences between organizations are beginning to show.

Two ways of handling the same signals

Consider two insurers using the same AI tool to support claims handling. In both, experienced adjusters regularly override the tool's recommendations on a particular type of claim.

In the first insurer, the overrides are logged and counted each month. The override rate appears in a dashboard and is reported to a governance committee, which notes that adoption is below target and asks for more training. The way decisions are made does not change.

In the second, the same overrides are treated as information about the tool. A concentration of overrides on one claim type suggests that the model is wrong in that area, or that adjusters do not yet have reason to trust it there. Within a few weeks, the manager responsible for the workflow reviews the cases, talks with the adjusters and makes a change, such as adjusting a threshold, revising a business rule, improving the data the model learns from or removing that claim type from the tool's scope. Decisions in the following month are better as a result.

Both insurers collect the same data. The difference is that the second has assigned someone to act on it, on a defined cycle, with the authority to change the workflow.

Why the gap widens over time

Companies built on software from the start tend to have these feedback loops by design, because customer interactions generate data that informs the next decision and changes are deployed through code. Established institutions usually have to build them deliberately. Outcome data has to be captured and routed back to the people who make decisions, overrides have to be analyzed rather than only counted, and recurring exceptions have to lead to changes in the workflow rather than only to incident reports.

Each cycle of improvement also lowers the cost of the next. Data cleaned for one use can serve the next one, a control approach approved once can be reused, and a method for measuring benefits that finance has already accepted does not need to be negotiated again. As a result, the advantage of organizations that learn effectively tends to grow over time, and other organizations cannot close the gap simply by spending more.

Organizations that learn well also keep a record of what they learn. After a significant pilot, launch or decision to stop an initiative, they document what they expected, what happened and what can be reused. This takes little time and keeps different teams from repeating the same mistakes later at full cost.

McKinsey's 2026 data is consistent with this. The respondents it classifies as high performers, about 6 percent of the total, attribute at least 5 percent of their EBIT to AI. Nearly three-quarters of them have fundamentally redesigned workflows because of AI, compared with about a quarter of other respondents.

Assessing your own organization

Three questions give a reasonable indication of whether an organization is learning from its use of AI.

1. Is the tenth AI initiative easier to deliver than the first? If each new initiative still has to resolve data definitions, negotiate risk reviews and build its own monitoring, effort is being repeated rather than capability built.

2. Which initiatives were stopped in the past year, and what was learned from them? An organization that never stops anything is probably not learning much from its results.

3. When front-line staff find that the AI is wrong, how long does it take for something to change? In organizations that learn well, the answer is measured in weeks.

Conclusion

As AI becomes widely available, the organizations that benefit most are unlikely to be those that adopt the most tools. They will be the ones that consistently turn what they observe in daily operations into better decisions and better workflows, and that do so quickly.