Delegation Is Not the Problem

Before breakfast, you may already be relying on people you have never met. Turn on the tap and you benefit from work done elsewhere: finding water, treating it, checking its quality, and keeping pipes and pumps working. Most of that effort is hidden from you. You simply expect usable water to arrive.

Switch on a light and another network of people and equipment supplies the electricity. Open the refrigerator and you find food that has passed through farms, factories, vehicles, shops, and payment systems. Your household could not easily reproduce all that work. Everyday life depends on the combined effort of people with different skills.

Because these arrangements are familiar, we seldom think of them as delegation. We think of getting through the morning. Yet they are part of what makes a complex society possible. Civilization is organized delegation: people doing work for other people within systems that let us rely on one another.

That reliance helps us do more. We hire specialists, assign jobs, ask for advice, and choose public representatives. Much of the work we send to others would be impossible for us to do alone. The challenge is to create arrangements we can trust while keeping clear which decisions and obligations remain ours.

What matters with AI is also the arrangement. Handing work to software can be sensible or careless, depending on the task, the authority granted, and the people who must answer for the result. The fact that AI has done a piece of work does not, by itself, tell us whether anyone has given up too much control.

Refusing help can also be a poor decision when it prevents useful work or leaves a task with someone less able to do it. We need to judge both the benefits and the limits of a proposed assignment. That starts with understanding what kind of delegate AI is and how it differs from the people we already rely on.

The Civilization of Delegates

A hospital shows how many people can share one task. A patient seeks help from a doctor, who may rely on laboratory technicians and other specialists. Other teams keep medicines available, maintain equipment, and organize appointments. Each does part of the work. No one person could provide all of the knowledge and effort needed to run the hospital.

In a large business, a board may set direction, executives manage the company, and managers assign work to teams and outside specialists. A customer might deal with only one employee, while many others help produce the product or provide the service. The visible encounter is one small part of a larger organization.

Government relies on a similar division of work. In a democracy, voters choose representatives, while public agencies and their staff carry out many public duties. Citizens cannot make every decision about road repairs, permits, or inspections themselves. They need ways to assign that work, limit the power it carries, and hold the people doing it to account.

At home, the same principle applies when we hire an electrician. We rely on skill we do not have. Passengers on an aircraft rely on pilots, maintenance crews, and air traffic controllers for the same reason. Trusting specialists allows us to benefit from knowledge without having to acquire all of it ourselves.

That trust needs support. Training, professional standards, contracts, supervision, and routes for complaints help people know what to expect and where to turn when something goes wrong. These arrangements can be cumbersome, and none guarantees good work. Their purpose is to make duties clear and give people a way to question how those duties are carried out.

Delegation is easier to judge when we can answer three questions: What work has been assigned? What decisions may the person doing it make? Who must answer when the arrangement fails? Those questions matter whether the work is done by a person, a team, or a machine.

Permission is often limited to one part of a task. A business owner may ask a lawyer to prepare a sale contract while keeping the decision about whether to sell. A homeowner may ask an architect to design a house while keeping control of the budget and priorities. Asking for expert help does not settle every choice.

Good delegation therefore requires more than handing over a job. Whoever assigns it must remain answerable for the goal, the limits, the standard of success, and the consequences. The specialist contributes skill and judgment within that arrangement and may have duties of their own. Several people can share accountability, provided their roles remain clear enough to follow. This is how delegation helps us do more while keeping human choices and obligations in view.

Successful delegation soon becomes part of the background of life. When a restaurant meal is good, we rarely ask to inspect the kitchen’s temperature records. After a safe flight, we do not ask for a record of every cockpit decision. Trust lets us benefit from other people’s work without doing the whole job again ourselves.

Instead of checking everything personally, we rely on training, reputation, contracts, professional rules, and supervision. Audits examine work that cannot be watched continuously. Courts, insurance, and elections provide different ways to address failures or their consequences. These arrangements can also fail or become empty rituals. They are useful only to the extent that they give us sound reasons to rely on work done elsewhere.

An old problem appears whenever we hire someone to work for us: the person doing the job may know more than we do and may want a different result. A contractor, for example, may favor the quickest repair while a homeowner wants the longest-lasting one. Economists call this the principal-agent problem. Clear instructions and useful checks help both sides understand the assignment.

Hiring an accountant does not settle the choices about how to run your business. A manager who asks an employee to prepare an analysis still needs to decide how to use it. Delegation works by assigning parts of a job while keeping clear which decisions belong to whom.

Sometimes the law names an accountable role directly. Under the U.S. aviation rule 14 CFR 91.3, the pilot in command has direct responsibility and final authority for operating the aircraft. An autopilot can perform part of the flying, but the rule identifies the pilot as the person in charge. This illustrates the distinction between equipment doing work and a person holding authority.

Corporate reporting offers another example. Under U.S. Securities and Exchange Commission rules implementing Section 302 of the Sarbanes-Oxley Act, principal executive and financial officers of companies covered by those rules must certify specified quarterly and annual reports. Their duties also include maintaining and evaluating disclosure controls, the procedures used to collect and report required information. Accountants, auditors, committees, and software may all contribute, while the certification remains attached to identifiable officers. The work is shared, and the obligation to answer remains visible.

AI enters a world that already has ways to organize delegation, although none works perfectly. We need to ask which arrangements remain useful when software does the work and which need to change. Simply copying a job title from a human role does not establish what a machine may do or who will answer for it.

Earlier tools have already changed the division of work. Calculators handle arithmetic, thermostats regulate temperature, and search engines help us find information. Other systems flag possible fraud, rank recommendations, arrange schedules, or perform repeated physical tasks. Such tools can save time or improve consistency. People still have to decide where each tool is useful and what authority it should receive.

The same practical approach helps us judge AI. We can welcome useful help while examining the assignment, the limits, and the people who remain answerable. To set those limits well, we need to understand what changes when one system can take on many kinds of work, including work that involves language, reasoning, and judgment.

The New Kind of Delegate

AI extends the old practice of delegation, but it changes what we are handing work to. A payroll program is built around payroll. A thermostat controls temperature. A general-purpose AI service may help the same person summarize a document, draft a customer reply, and sort requests from several departments. This breadth can make it useful across jobs that once needed separate tools. How well it performs still has to be checked for each task.

Another difference is how widely the same system can be used. A skilled employee has only so many hours and can handle only so much work at once. Software can serve many people without hiring and training a new person for every user. It still needs computers and, for an online service, a working connection. More use can mean more expense or waiting. Making software widely available does not make it the equivalent of a qualified professional.

Speed changes which jobs seem worth handing over. Asking a colleague to rearrange a short note might take more time than doing it yourself. If AI can produce a useful draft in seconds, asking for help becomes easier. Some tasks take much longer, and the first answer may need correction. The practical saving depends on the whole job, including the time needed to check the result.

Software can also work outside the hours when people are available. A business might arrange for AI to sort incoming requests overnight, leaving a queue for staff to review in the morning. That possibility depends on the service working and on the limits set by its operators. Outages and usage restrictions still matter. If the system is allowed to act while staff are away, someone needs to decide what it may do and how problems will be brought to human attention.

These differences help explain why AI can take on more work. They also bring us to a different question: what role does the machine occupy? A doctor has professional duties to patients. An employee works within an organization that can assign authority, supervise performance, and impose consequences. AI can produce useful work, but the software does not become the person or institution that must answer for how it is used. Those duties still need identifiable human or institutional owners.

Calling AI an assistant or adviser can make it easier to understand. The familiar name can also make us assume that familiar safeguards come with it. A title does not tell us which information the system may see, which actions it may take, or how someone affected by its work can challenge a mistake. The organization using it has to settle those questions.

Useful delegation needs rules that fit the actual work. We should be able to identify who authorized the system, what limits apply, and who can intervene. These arrangements need deliberate design, even when the software feels familiar and its answers sound assured.

AI can sometimes seem like a “master idiot savant.” The phrase is an imperfect metaphor for extraordinary ability sitting alongside surprising weaknesses. A system may produce sophisticated analysis, explain a difficult subject, or organize a complicated project, then miss a basic fact that changes the answer. Its abilities are broad, but its reliability can vary sharply from one task to another.

Imagine receiving a beautifully written business plan. The calculations appear careful. The recommendations sound convincing. Then someone discovers that a crucial cost estimate is wrong. The presentation made the plan easy to believe. It did not make the underlying assumption safe to use.

This unevenness makes it harder to judge what an impressive result tells us about the next assignment. We need evidence about the work we are actually asking the system to do. Chapter Eighteen examines research showing how AI assistance can help on some tasks and hurt performance on others. Success deserves attention, and so do the conditions under which it occurred.

As those abilities improve, the questions about permission remain. People and institutions must decide what the system may do, what requires further review, and who can change or stop its actions. A more capable system may justify a larger assignment. The authority surrounding that assignment still needs to be deliberately established.

An impressive answer can earn our trust before we have decided what the system should be allowed to do. That is an understandable response to useful work. It also leaves a question unanswered: what authority should follow from that success? A good result gives us a reason to consider another assignment. We still need to decide its limits.

We can welcome useful AI and take those limits seriously. If a system helps someone work faster, understand a difficult document, or do something previously beyond reach, there is a reason to use it. The benefit matters. So do the conditions of use and the people affected by its work. Supporting more delegation and asking for clear authority belong together.

Imagine hiring a bookkeeper. You would agree which records the person could see, which tasks they would handle, and which decisions needed your approval. Being good at keeping accounts would not automatically give that person permission to move money wherever they chose. The assignment needs a defined scope even when the person doing it is highly skilled.

AI calls for the same care. A business might authorize a system to match receipts to expense claims. Deciding whether to reject a disputed claim is a different assignment. Before handing over that decision, the business should establish the rules, the route for review, and who can correct an error. Greater capability can support more useful work, while making these choices more consequential.

This is why the phrase “keep humans in the loop” needs an explanation. What is the person expected to do? A reviewer needs enough information, time, skill, and authority to judge the result and change what happens next. A name on a form or a click on an approval button tells us little about whether those conditions exist.

The right arrangement will vary. For routine work that is easy to reverse, people may set written limits in advance, inspect samples, and require unusual cases to be referred to someone who can decide. Other decisions require a named person to judge the particular case, with real power to disagree. Some require independent human judgment because of what they decide for another person.

The design should reflect what is at stake, whether a mistake can be undone, and how many people could be affected. Sometimes the decisive human choice happens before the machine starts work, when someone sets its permissions, checks how it will be watched, and establishes a workable way to stop it. Later chapters examine these choices in detail.

Handing work to AI is not, by itself, a failure of judgment. People have long depended on delegates. AI extends that practice through software that can be used widely and built into everyday services. The task is to make useful delegation work while keeping authority and answerability clear.

Current Evidence: Delegation Becomes Routine

Reviewed September 20, 2026.

AI adoption numbers become easier to understand when we ask three questions: who was surveyed, when, and what counted as use? A company trying a chatbot, a worker relying on it every day, and a business letting an agent carry out a sequence of tasks are making different kinds of delegation.

Stanford’s detailed 2026 AI Index reports that 88 percent of respondents to a McKinsey survey said their organizations used AI in at least one business function in 2025, compared with 78 percent in 2024. It puts regular generative AI use at 79 percent. These are self-reports, offering a guide to the pattern rather than a count of all organizations.

The same report describes scaled agent use as being in single digits across nearly all business functions in 2025. Experiments and pilots were counted separately. That distinction matters: trying a delegate and relying on it widely are different stages of adoption.

A newer McKinsey survey, published in August 2026, shows movement beyond that earlier picture. Among respondents from organizations with at least $1 billion in annual revenue, 40 percent reported scaling agents in one or more functions, up from 27 percent the previous year. Among smaller organizations, the share remained at 22 percent. These are rates for organizations using agents somewhere in the business, not rates for each individual function. The distinction prevents a misleading comparison with the earlier figures.

The U.S. Census Bureau looks at a broader population: U.S. businesses with employees. Its May 2026 report put AI use between 17 and 20 percent from December 2025 through early May 2026. For the period ending May 3, the national rate was 19.8 percent. The question asked about use in the previous two weeks. The report also found a size gap: 37 percent of firms with at least 250 employees used AI, compared with fewer than 20 percent of firms with four or fewer employees.

Those different adoption numbers can coexist. A worldwide survey of organizational respondents and a representative survey of U.S. employer businesses reach different groups and ask different questions. Before using a headline percentage to describe the economy, we need to know whose experience it represents.

Anthropic’s June 2026 study offers a closer view of Claude users. It reports more extended agent tasks and distinguishes working with AI from asking it to complete work with little user input. People who handed over more complete tasks also believed AI could do more of their work. The direction of cause remains uncertain: confidence may encourage delegation, or delegation may build confidence.

Surveyed users reported benefits: 86 percent said AI helped them work faster, 82 percent said it expanded what they could do, and 69 percent said it improved the quality of their work. These are self-assessments from users of the company’s own products. They help explain the appeal, while leaving actual performance and wider productivity effects to be tested.

My conclusion is practical. When people find that AI saves time or helps them do more, they have a reason to hand it more work. That makes the terms of delegation an immediate responsibility. Someone still has to decide what the system may do, check whether it works, and answer for its effects on other people.

Accountability Analysis

The first accountability question is easy to skip: who answers? In ordinary human delegation, the answer is often embedded in a role. A manager assigns the work. A physician orders the test. A board appoints an executive. A pilot remains responsible for the aircraft. A framework developed later in this book will turn this into a fuller Accountability Test, but the first link is already visible here.

AI can obscure that starting point because the system may look like the actor. It writes the memo, ranks the candidates, selects the next task, sends the message, or recommends the treatment. The visible action comes from the machine, while the real authorization may sit several layers upstream.

The point here is narrower: the mere presence of delegation does not tell us that accountability has failed. We have to ask what was delegated, who authorized it, what role the machine occupies, and whether the institution around it still knows where answerability begins.

This distinction matters in both directions. A company should not be able to excuse a harmful decision by saying the machine made it. But a critic should not be able to condemn a useful system merely by pointing out that a machine performed work once done by a person. Both positions avoid the harder task of designing legitimate delegation.

Accountability Verdict

Delegation is one of civilization’s basic operating systems. It lets individuals benefit from expertise beyond their own and enables institutions to become larger than any one person’s attention, memory, or skill. Artificial intelligence extends that logic.

AI changes who, or what, does the work. The same software can help many people, work outside office hours, and move between different tasks. Its answers may sound like those of a skilled colleague. That does not tell us what duties the supplier or the organization using it has accepted.

Those duties must be made clear. Who approved this use? Which decisions may the software make? Who can explain or correct a mistake? Asking only whether AI was involved tells us too little. We need to follow the authority that came with the assignment.

Check the assignment

Think of the bookkeeper discussed in this chapter. Keeping records and moving money require different permissions. Apply the same distinction to an AI tool: describe the task, identify the decisions it may make, and name who must answer for the arrangement. Judge the result on the task you assigned; success at one job does not establish reliability at another.

The next chapter begins where this one ends. Delegation is normal. The harder question is the point at which normal delegation becomes abdication.

Sources and Notes

1. Stanford Institute for Human-Centered Artificial Intelligence, “Economy,” The 2026 AI Index Report, 2026, pp. 193, 197 and 198. https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_4_economy.pdf. Accessed September 20, 2026. Page 193 and Figure 4.3.1 give 79 percent for generative AI, although the summary says 70 percent. The detailed figure is used here. Agent figures concern scaled use, with experiments and pilots separate. Source record: DA-CH2-S1. Archive reference: pending.

2. U.S. Census Bureau, “Large Firms With at Least 20 Employees Biggest AI Users,” May 26, 2026. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html. Accessed September 20, 2026. The Business Trends and Outlook Survey represents U.S. employer businesses. These figures cover December 2025 through May 3, 2026; the use question concerns the preceding two weeks. Source record: DA-CH2-S2. Archive reference: pending.

3. Anthropic, “Anthropic Economic Index report: Cadences,” June 26, 2026. https://www.anthropic.com/research/economic-index-june-2026-report. Accessed September 20, 2026. Vendor platform evidence and user self-reports do not establish population-wide productivity gains or causal effects of delegation. The author uses Anthropic products as part of the Living Edition’s external critique process; see the project transparency statement. Source record: DA-CH2-S3. Archive reference: pending.

4. Electronic Code of Federal Regulations, 14 CFR 91.3(a), “Responsibility and authority of the pilot in command.” https://www.ecfr.gov/current/title-14/chapter-I/subchapter-F/part-91/subpart-A/section-91.3. Accessed September 20, 2026; Title 14 displayed as up to date through September 17, 2026. The rule names the pilot’s responsibility and authority; its use as a delegation example is the author’s analysis. Source record: DA-CH2-S4. Archive reference: pending.

5. U.S. Securities and Exchange Commission, “Certification of Disclosure in Companies’ Quarterly and Annual Reports,” Release Nos. 33-8124 and 34-46427, implementing Section 302 of the Sarbanes-Oxley Act of 2002. https://www.sec.gov/rules-regulations/2002/08/certification-disclosure-companies-quarterly-annual-reports. See also the SEC’s “Exchange Act Rules” interpretations, sections 161 and 162, for officer certification and disclosure controls: https://www.sec.gov/rules-regulations/staff-guidance/corporation-finance-interpretations/exchange-act-rules. Both accessed September 20, 2026. Source record: DA-CH2-S5. Archive reference: pending.

6. Anthropic, “Guides to common use cases,” Claude Platform Docs, undated. https://platform.claude.com/docs/en/about-claude/use-case-guides/overview. Accessed September 20, 2026. The vendor documents legal summarization, customer support, and request routing as uses of Claude. This supports the examples of breadth, not equal reliability across tasks or equivalence to a qualified professional. The author uses Anthropic products as part of the Living Edition’s external critique process; see the project transparency statement. Source record: DA-CH2-S6. Archive reference: pending.

7. Anthropic, “Pricing,” “Rate limits,” “Reducing latency,” and “Claude API errors,” Claude Platform Docs, undated. https://platform.claude.com/docs/en/about-claude/pricing; https://platform.claude.com/docs/en/api/rate-limits; https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-latency; https://platform.claude.com/docs/en/api/errors. Accessed September 20, 2026. These vendor documents describe usage charges, capacity limits, response delays, and service errors. They illustrate conditions on scale, speed, and availability; they do not establish general savings against equivalent human labor. The whole-job cost comparison and the overnight business example are the author’s analysis and illustration. Source record: DA-CH2-S7. Archive reference: pending.

8. McKinsey & Company, “The state of AI in 2026: On the road to ROI,” August 25, 2026, Exhibit 2 and “About the research.” https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai. Accessed September 20, 2026. Survey conducted May 4 to June 8, 2026, with 1,719 respondents in 97 nations; results weighted by national GDP. Scaling agents in any function differs from a rate within one function. Source record: DA-CH2-S8. Archive reference: pending.