Move Fast Without Breaking Things

Artificial intelligence has reached nearly every company and almost none of their income statements. In McKinsey's 2026 global survey, close to nine in ten respondents said their organizations use AI regularly in at least one function. Only 37% said it had added to operating profit, the same share as a year earlier. Just 6% credited it with five percent or more of earnings.

An earlier MIT study put the gap more bluntly, finding that 95% of organizations saw no measurable return on generative AI. Its sample was small and its headline has been disputed. The direction is hard to argue with.

So is the technology overrated? The evidence points elsewhere. The companies in McKinsey's top group share a habit: nearly three-quarters of them redesigned how the work is done, where others added a tool to the old process. The opportunity is real. What separates the companies that capture it is operating discipline.

Where it pays

The dependable gains sit in work that is frequent, rule-bound and easy to check. In finance and operations that means the monthly close, reconciliations, reporting, and the document chase before a fundraise or an audit. An agent that drafts and a person who checks will finish that work in a fraction of the time, and the check is cheap because the right answer is knowable.

The gains are weaker, and the risks higher, where an answer cannot be checked quickly or a mistake cannot be undone. Knowing which kind of work is which is most of an AI strategy.

The mistakes we see

Starting with the tool. A company buys licenses, announces a mandate and waits. Meanwhile the staff have already chosen for themselves. The MIT study found that only 40% of companies had bought an official subscription, while workers at more than 90% of them used personal AI tools for work. Company data is leaving through accounts nobody approved.

Measuring nothing. In a controlled trial by the research group METR, sixteen experienced software developers took 19% longer to finish tasks when allowed to use AI. Afterwards they believed it had made them 20% faster. A team that does not time the work before and after will believe the same.

Handing over the keys. In July 2025 a coding agent on the Replit platform deleted a company's production database during a code freeze, then generated fake records to cover the gap. The agent could do that because it held the access to do it.

Leaving the output unowned. A Canadian tribunal held Air Canada liable in 2024 for a wrong answer its website chatbot gave a customer. The airline had argued that the chatbot was responsible for its own words. In 2025 Deloitte agreed to refund part of a A$440,000 fee to the Australian government after a report was found to contain a fabricated court quotation and references to papers that do not exist. In both cases a person could have caught the error. Nobody had been named to do so.

Leaving the cost out of the model. AI is priced by use. A feature that costs pennies in a pilot can take a real share of revenue at scale, and usage climbs fastest in the products customers like most. Inference belongs in cost of goods sold, with a margin target, before the feature ships.

How to move fast

Speed and control are usually presented as a trade. In practice the controls are what allow the speed. A team that knows exactly what an agent may touch can let it run.

• Rank the uses. Keep a short list, ordered by value, cost and risk, and begin where errors are cheap to find and easy to reverse.
• Set the baseline first. Time the work and count the errors before the tool arrives.
• Decide permissions before deployment. Say which tools may see which data. Keep agents away from production systems, money and customer records until there is an audit trail of what they did.
• Name the owner. One person signs whatever leaves the company, however it was drafted.
• Redesign the work. Where AI earns its place, rebuild the process around it.

None of this takes long. A plan the board can read in one sitting is enough to begin, and it can be revised every quarter as the tools change.

The companies that look fast two years from now will be the ones that set these rules early. Set them first. Then go fast.

All insights