Why Do AI Agents Create More Work for Humans?

TL;DR

AI agents often shift human labor from direct execution to planning, context provision, review, intervention, and recovery. As agents become more efficient, organizations may run more of them and generate more output, so the total management burden can grow. Individuals can absorb this work personally, while small businesses often need vendors and enterprises can fund dedicated engineering and deployment teams.

Transcript

Look, we were told agents would take our stuff off our plate. Some of us were told agents would take our jobs. It turns out agents aren't going to take our jobs because agents are generating so much work for humans. They're giving us more work to do. Everybody buying an agent is doing the same math. They're thinking the agent does the work, I need ... Read More

Key Insights

• AI agents create management work because humans must select tasks, supply context and permissions, define success, check outputs, and intervene when execution goes wrong. The work removed from direct execution can reappear as allocation, supervision, evaluation, and recovery.
• Agent usage is expanding rapidly, with OpenRouter usage described as rising about 14-fold between February and August. Agents reportedly consume more than five tokens for every token consumed by a human, indicating that automated execution can generate an enormous volume of activity requiring selective oversight.
• The Jevons effect describes the observed pattern in which improved efficiency increases total usage. As each agent requires less supervision, people may operate more agents and initiate more work, causing the overall human coordination burden to grow rather than disappear.
• Human and agent responsibilities are divided unevenly in coding sessions. Across about 400,000 Claude Code sessions, humans made roughly 70% of planning decisions, while agents made nearly all execution decisions, including locating files, modifying code, and running tests.
• Experienced users supervise agents by focusing on direction rather than every small action. They interrupted agents on about 9% of conversational turns, compared with about 5% for newer users, while also allowing more actions to proceed automatically.
• Domain expertise increases the amount of useful work an agent can perform from each instruction. Expert-led sessions averaged about 12 agent actions per instruction, compared with about five for novice-led sessions, because experts could specify jobs and assess results more precisely.
• Verifiable domains are well suited to agentic work because outputs can be checked against clear standards. Legal use measured through Codex reportedly increased roughly 108-fold since January, reflecting the value of domain knowledge and reviewable results.
• Small businesses struggle to obtain operational outcomes from inexpensive AI subscriptions. Across 4.6 million small businesses, almost two-thirds of AI-paying firms spent about $40 per month, while another survey found that 73% of owners needed more implementation and evaluation support.
• More videos from this channel:
  • How to roll out AI with team buy-in and minimal resistance
  • How to use Claude code interpreter for Excel and PowerPoint
  • How to spot agent shaped work in one minute
• More from this channel: AI News & Strategy Daily | Nate B Jones

Questions & Answers

Q: Why do AI agents create more work for humans?

AI agents create more work because automated execution still requires human planning, context, permissions, quality standards, review, and recovery. When agents become faster or easier to operate, users can launch more runs and produce more output. Humans then spend less time performing each task directly but more time deciding what should run, checking results, correcting direction, and handling failures.

Q: What work remains for humans when AI agents execute tasks?

Humans remain responsible for choosing appropriate jobs, starting runs, supplying necessary files and permissions, and explaining what a successful result should look like. They must also judge whether an answer is usable, compare it with relevant standards or tests, interrupt work that is heading in the wrong direction, and change instructions or rules when the same failure keeps recurring.

Q: How do experienced users manage AI agents differently?

Experienced users allow more actions to proceed automatically while monitoring the overall direction of a run. In the cited study of about 400,000 Claude Code sessions, experienced users interrupted the agent on about 9% of conversational turns, compared with about 5% for newer users. Their advantage was recognizing sooner when the complete run was moving toward an unusable result.

Q: Why does domain expertise matter when using AI agents?

Domain expertise helps a person describe a task precisely, identify work the agent can finish, and distinguish a plausible answer from a correct one. Expert-led sessions averaged about 12 agent actions per instruction, compared with about five for novice-led sessions. The expert can provide better context, recognize subtle mistakes, verify results, and decide when intervention is necessary.

Q: Why are verifiable domains suitable for AI agents?

Verifiable domains provide standards against which an agent's output can be checked. Coding can use tests, while legal work can be compared with laws and other relevant requirements, even though interpretation remains part of legal practice. This ability to assess correctness helps knowledgeable users define acceptable results, detect errors, and decide whether the agent's work can be used.

Q: Why do small businesses struggle to integrate AI agents?

Small business owners are often short on time, capital, and implementation resources. They want concrete outcomes such as answered calls, timely quotes, collected invoices, and correctly booked appointments, but they do not want an additional job managing agents. In a survey of 1,256 owners, only 14% reported full integration into core operations, while 73% wanted more training and resources.

Q: What does a $40 monthly AI subscription provide a small business?

A subscription costing about $40 per month is characterized as providing a chatbot-style assistant rather than a fully managed operational system. It is not expected to independently handle appointment booking, discover that a process broke three days earlier, or proactively repair the problem. Achieving those outcomes requires workflow selection, software connections, monitoring, rule changes, and ongoing human management.

Q: How does AI agent management change with company size?

An individual can manage an agent personally because the same person often holds the goal, context, permissions, and domain knowledge. A small business may require multiple agents but lack the time and capital to supervise them, so a vendor can assume management duties. An enterprise can hire engineers and deployment teams to build customized agent management into company workflows.

Summary & Key Takeaways

• AI agents do not simply remove work. People still decide which jobs should run, provide files and permissions, define acceptable results, review outputs, interrupt failures, and repair recurring mistakes. Increased agent capacity can therefore move humans above the execution loop while expanding the total volume of work requiring judgment and oversight.

• Individuals can often manage an agent because one person holds the goal, context, permissions, and relevant domain knowledge. Experienced users supervise differently from novices: they approve more actions automatically but interrupt misguided runs more frequently. Expertise helps users specify tasks precisely, recognize plausible mistakes, and determine whether an output is genuinely usable.

• Small businesses face a mismatch between desired outcomes and available resources. Owners want calls answered, quotes sent, invoices collected, and appointments booked, but most cannot dedicate time to agent management. Low-cost subscriptions largely provide chatbot assistance, so vendors may need to choose workflows, connect software, inspect results, and revise operational rules.