Barings, Boeing, and Bots

For decades the accepted management wisdom has been for managers to delegate tasks and for organizations to outsource activities that aren’t core to their operations. Yet, when the proposed delegate or supplier is an AI agent, these established practices are suddenly portrayed as irresponsible, dehumanizing, or even immoral. Much of this resistance has all the marks of a moral panic: a novel technology is judged through its most alarming failures, its benefits are discounted, and its use is treated as crossing a boundary rather than as another management decision. Here’s the thing: delegating work to an AI agent is neither inherently wise nor inherently reckless. As with human delegation and organizational outsourcing, what matters is the task, the controls, and who remains accountable.

Managers delegate because their time, attention, and expertise are limited. A capable subordinate may know more about a task, be closer to the relevant information, or simply have the time required to perform it properly. Delegation can also develop employees by giving them progressively more demanding responsibilities. However, managers delegate execution, not accountability. Effective delegation requires a clear outcome, sufficient authority and resources, appropriate constraints, progress checks, and an escalation path should things go wrong. A task shouldn’t be delegated when its consequences exceed the delegate’s authority or competence, when the manager can’t evaluate the result, or when the work is so sensitive that failures can’t be detected and corrected in time. We would rightly be appalled if an airline pilot delegated landing to an untrained flight attendant and remained in the cabin chatting with a passenger. On the other hand pilots routinely delegate well-defined flight phases to the autopilot.

Poor delegation can amplify information asymmetry, misaligned incentives, and misplaced confidence. A delegate may hide problems, optimize the measured target rather than the desired outcome, or continue along a failing path because admitting failure is costly. The 1995 collapse of Barings Bank is a striking example. Nick Leeson effectively controlled both trading and the corresponding back-office operation, while ambiguous reporting lines left nobody clearly responsible for supervising him. This allowed unauthorized trading and concealed losses to continue until they brought down a 230-year-old institution. The resulting inquiry, rather than blaming delegation, found that management had failed to understand the delegated activity, separate duties, establish controls, and act on warnings.

Organizations also outsource activities, again for reasons similar to those that lead managers to delegate. An external supplier may possess specialized knowledge, infrastructure, economies of scale, or access to talent that would be costly to develop internally. Outsourcing can turn fixed costs into variable ones, increase flexibility, and allow an organization to focus its management attention and resources on activities that differentiate it. However, “non-core” shouldn’t be confused with tedious, expensive, or poorly understood. An activity may be strategically important because it preserves domain knowledge, controls a critical interface, or provides the ability to evaluate suppliers. Sensible outsourcing therefore retains sufficient internal expertise to specify the work, integrate the result, assess its quality, and replace the supplier when necessary.

Outsourcing isn’t without risks. It introduces coordination costs, contractual rigidity, supplier dependence, security exposure, and the gradual erosion of internal capability. Boeing’s 787 Dreamliner illustrates what can happen when an organization outsources too much of the knowledge required to coordinate production. Boeing distributed major aircraft sections across a large global supplier network, expecting partners to deliver completed assemblies. The program ended up nearly three years behind schedule and billions of dollars over budget. Boeing itself acknowledged that it had attempted too much outsourcing alongside too many other innovations. Outsourcing complicated integration work, while weakening Boeing’s ability to see and fix problems at their source.

Technology-induced moral panics follow a familiar pattern. A new medium or technology appears, becomes popular — especially among younger people — and is blamed for intellectual decline, moral corruption, social isolation, unemployment, or violence. Novels, radio, television, comic books, video games, and smartphones have all occupied this role. Sometimes entrenched interests amplify these fears. Research on these recurring technology panics warns that emotionally charged claims often run ahead of the available evidence and lead to poorly targeted research and policy. Past alarms often look quaint because the eventual harms differ markedly from those originally predicted, while public concern shifts to the next technology. A particularly striking example is the red-flag traffic laws, which required early automobiles to be preceded by a person carrying a red flag. The absurdity of the prescribed remedy didn’t mean that automobiles were harmless: approximately one million people die each year as a result of road traffic crashes. Novels can spread bad ideas, television can waste time, and smartphones can be addictive. However, harms should be identified and mitigated rather than used to condemn an entire technology.

Delegating to an AI agent resembles human delegation and outsourcing in many important management respects. The principal must specify the desired result, provide the authority necessary (and not more), monitor progress, verify the output, and remain accountable. However, AI agents introduce very different operational characteristics. They operate extraordinarily fast, can make thousands of changes before a person would complete one, and can produce convincing but subtly wrong results. Their effectiveness and currently low cost can also encourage us to delegate tasks that weren’t worth doing in the first place. These differences call for tighter technical controls: tests, access restrictions, review gates, audit trails, spending limits, and easily reversible actions.

AI agents also differ from both employees and conventional suppliers in several consequential organizational ways. Employees are recruited, trained, socialized, and gradually entrusted with broader responsibilities; through this process they acquire institutional knowledge, develop judgment, form relationships, and may grow into roles the organization will need later. An AI agent doesn’t normally undergo this development, retain dependable memories of past work, or become committed to the organization’s success. Replacing people with agents may save costs in the short term, but eliminate future expertise, managerial capacity, resilience, and the informal knowledge through which organizations often function. Nor are we necessarily equipped to supervise our new delegates. We have learned to recognize evasion, inexperience, and poor judgment in people, but not fabricated evidence, subtly corrupted code, sycophantic agreement, or fluent reasoning built on a false premise. Worse, we’re also likely to lack the metacognitive monitoring and judgment required to evaluate AI agents. Furthermore, agents operate at a scale and speed that can turn a small misunderstanding into thousands of consistent errors, while dependence on externally controlled models exposes the organization to silent changes in behavior, pricing, availability, and data-handling policies. Unlike a conscientious employee, an agent is often unlikely to hesitate, object, protect a colleague, report an ethical concern, or notice that a formally correct instruction is organizationally absurd. These differences strip AI delegation of many human safeguards and multiply its effects by the agents’ speed and reach.

We should therefore delegate to AI agents tasks that are bounded, observable, reversible, and supported by objective checks. We should let them search, summarize, classify, draft, test, transform, and implement changes where failures have a limited impact and where a competent person can review the result. We shouldn’t delegate final responsibility for safety-critical decisions, employment and legal judgments, irreversible financial transactions, or important work whose quality nobody in the organization can assess. We should also avoid delegating the very thinking we’re trying to develop. Emerging research associates heavy AI-mediated cognitive offloading with reduced critical-thinking effort and weaker independent evaluation. Just as a manager who delegates everything eventually loses touch with the work, a professional who delegates all analysis may gradually lose the ability to perform—and, more dangerously, to judge—it. Rather than resist AI delegation, we should delegate deliberately, retain the expertise needed to supervise it, and never outsource accountability.

Comments   Post Toot! Tweet Share