One U.S. regional bank spent four months on the technical build after adopting an AI loan-underwriting system—and then needed another fourteen months to actually put it into production. Six of those months went to defining the criteria under which underwriters would trust the AI's judgment, five to pinning down who was accountable when something went wrong, and three to preparing the documentation regulators required. Even once the technology is ready, if the organization doesn't move, the AI just sits there.
In a recent McKinsey analysis of AI transformation in banking, three experts say they saw this pattern again and again. At the moment AI shifts from "assistant" to "operator"—where it no longer waits for a person to review and decide but handles the work and produces the outcome itself—technical readiness is no longer the bottleneck. The blockage sits somewhere else. For Korea's middle managers and solo entrepreneurs, the pattern won't feel unfamiliar either.
When AI Becomes the Operator, the Nature of Error Changes
When AI is used as a tool, a person makes the final call. It drafts an email, summarizes a contract, tidies up data. A person looks at the result, judges it, and signs off. If the AI slips at this stage, the final reviewer catches it. The lines of accountability are relatively clear.
When AI becomes the operator, the flow changes. In banking, models are increasingly handling everything from the moment a loan application arrives to the approval or rejection notice—on their own. Triaging customer inquiries, detecting anomalous transactions, generating compliance documents: all of it is moving the same way. People step back into exception-handling and oversight roles. The AI makes decisions in bulk, and a person steps in only when something looks off.
Banks report internally that this shift cuts the time spent on repetitive document processing by 60 to 70 percent. At the same time, problems that didn't exist before begin to surface.
When AI takes over execution, the nature of error changes. When a person makes a mistake, it's that person's mistake. When the AI makes one, whose mistake is it—the team that built the model, the executives who decided to deploy it, or the employee assigned to supervise it? An organization that can't answer that question clearly can't put AI into live operations. The McKinsey report names this question of accountability as the very first design problem that has to be solved.
What Blocks You Isn't a Technical Limit—It's a Gap in Design
The items the McKinsey experts flag all converge on organizational design, governance, and talent placement: how far AI is allowed to decide on its own, who sets that boundary, and through what path accountability for the outcome flows. An organization where these are unclear can't move AI up to the operator stage no matter how ready the technology is.
Regulators' demands run in the same direction. In the financial sector, you have to be able to explain the basis for an AI's decision. If you can't tell an auditor or a regulator why the model rejected that loan, or why it flagged that transaction as anomalous, you won't get clearance to operate. This explainability isn't a job for the technical team alone. It's an organization-wide task that has to weave in legal, compliance, and risk management.
The state of the data is tangled up in the same problem. An AI execution system needs the data it bases its judgments on to be clean and consistent. What banks actually found was data scattered across decades-old legacy systems, not even uniform in format. They had to clean it up before they could put AI on top of it—and that work swallowed more time than the technical build itself.
The skeptical view of moving AI to the operator stage deserves an honest hearing too. The argument is that in domains where the cost of error is high—like finance—shrinking the human's final judgment may carry more risk than the efficiency is worth. In 2023, an AI loan-underwriting system at one U.S. financial firm was found to be systematically scoring applicants from certain ZIP codes lower. A human underwriter would have sensed something was off about the pattern; the AI simply repeated it. The case shows how bias can be amplified when AI is promoted to operator without an oversight structure in place. The claim that design has to come before speed is borne out by this 2023 bias case—and the McKinsey report, too, acknowledges the point indirectly.
What to Check Before You Promote AI to Operator
It's hard to apply the banking case directly to other industries. But the push to move AI from assistant to operator is showing up the same way outside finance. Generating marketing content, fielding customer inquiries, processing tax invoices, coordinating schedules—plenty of solo operators have already started handing this work to AI.
At this point there are things you need to check first. When something the AI handled turns out wrong, by what procedure will you catch and correct it, who bears the responsibility, and how will you explain it to the customer? Hand execution to AI without settling that flow in advance, and the same bottleneck McKinsey observed in banks comes for you too. A smaller organization doesn't make the problem any simpler.
There's a way of making financial decisions that draws the loss scenario before running the return calculation—a thinking order that sketches out what damage occurs if you're wrong before figuring out how much you stand to gain if it works. You can take the same order with the shift to AI execution. Instead of the efficiency of AI when it runs well, draw first what collapses when the AI is wrong. Only an organization that has finished that picture can move AI from assistant to operator.
If your AI-adoption checklist is missing the error-detection cycle, the chain of accountability, and an explainable processing flow, then no matter how far ahead the technology is, the organization spends fourteen of those eighteen months waiting. I'd like to call this "loss-first design."
The more work a tool takes on, the more you need a person and a procedure to decide when to make it stop. The things that block you after the technology is ready are sitting right there.
