YoZmiTe Lists Four Rules for Delegating Work to AI

YoZmiTe analyzed nine case studies published from May through October 2026 and identified four rules for deciding which tasks AI should handle and which should stay with people: choose work worth delegating, use rules for tasks with set answers, reserve hard-to-reverse decisions for people, and design checkpoints to review results. Contributors included a business management employee who automated repetitive work, a team building a webpage diagnostic service, a designer behind a legal AI service, a live-game planner, and developers.

Writer Cho, an assistant manager in business management, built 12 work tools over four months and concluded that choosing what to automate was harder than coding. He considered how often a task recurred, how long each run took, the risk of mistakes, and whether a tool could be reused. He automated the search for missing signatures in monthly inspection binders exceeding 400 pages: the task took 30 minutes to 1 hour each time, and missing even one signature could jeopardize a deadline. He did not build tools for tasks done only once or twice a year. Kim Hyun-min said many tasks can be automated with Apps Script or n8n, and advised asking whether something that already works needs to be rebuilt with Claude Code. In a live-game organization, writer Sookyungsoo started with Excel entry errors that did not directly affect the service and maintenance-document organization. He said that finishing a document task in two days, instead of the minimum two weeks expected, changed the skeptical team’s response.

The team behind a webpage diagnostic service initially left almost all diagnosis to AI, but scores for the same page varied across three runs: 72, 85, and 61. It divided the work so code checked verifiable facts, such as the robots.txt file and specific tags, while AI handled judgments such as content credibility. Front-end developer Kim Eui-jung proposed automating deterministic processes, where the same input always produces the same output, and assigning tasks that require inference to AI. In a comparison using an OpenAPI specification—a document defining data formats between a server and a user interface—Orval’s code generator produced the same API code for free in about 1 second. AI took 2 to 4 minutes for the same task, cost about $1.4 per run, produced different code each time, and left 4 type errors on one run. Cho also built a tool to assess the amount of ink in signature fields in inspection binders exceeding 400 pages according to set rules, with help refining those rules through conversations with Claude Code.

Designer Sarah, who created the legal AI service Lawwin, said she delegates work only up to the point where a person can reverse an AI mistake. The service finds legal precedents similar to a user’s account and organizes the issues and response strategies, but directs decisions about whether to sue to a lawyer consultation. It does not promise a win and connects users with a lawyer, rather than AI, at the final stage where they incur the greatest cost. Sarah said AI has a wider role in a workplace-survival-style assessment service, where errors carry little burden. Park Dong-hyun, an HR employee at Classmethod, said AI agents automate repetitive tasks such as organizing performance reviews and processing salary data. For employee conflicts or sudden resignation consultations, staff use AI only to seek input; people make the final judgment and communicate with employees.

Writer Kkomi built a development tool that extracts requirements from planning documents and design drafts, then carries work through implementation and testing. Of 40 requirements, only 25 made it into the code, but the tool reported development complete because all written tests passed. Kkomi then had it create a traceability table linking each requirement to its acceptance criteria, tests, and code, and manually checks for missing links. If the same scenario fails nine times, the AI brings its logs and reasoning to a person for guidance. Cho’s tool found 68 cover pages in a 407-page binder, automatically verified 66, and placed the 2 ambiguous cases at the top of a report for human review. A tool comparing attendance records with leave requests likewise flags only mismatches or items needing review. In Kim Eui-jung’s development workflow, after a specification changes, a code generator updates types and an AI agent fixes compilation errors, while a rule-based tool first identifies where the errors are. Across the nine cases, the shared approach was to decide task recurrence, whether an answer is fixed, whether errors can be reversed, and where results should be checked before judging how smart AI is.

YoZmiTe analyzed nine case studies published from May through October 2026 and identified four rules for deciding which tasks AI should handle and which should stay with people: choose work worth delegating, use rules for tasks with set answers, reserve hard-to-reverse decisions for people, and design checkpoints to review results. Contributors included a business management employee who automated repetitive work, a team building a webpage diagnostic service, a designer behind a legal AI service, a live-game planner, and developers.

Writer Cho, an assistant manager in business management, built 12 work tools over four months and concluded that choosing what to automate was harder than coding. He considered how often a task recurred, how long each run took, the risk of mistakes, and whether a tool could be reused. He automated the search for missing signatures in monthly inspection binders exceeding 400 pages: the task took 30 minutes to 1 hour each time, and missing even one signature could jeopardize a deadline. He did not build tools for tasks done only once or twice a year. Kim Hyun-min said many tasks can be automated with Apps Script or n8n, and advised asking whether something that already works needs to be rebuilt with Claude Code. In a live-game organization, writer Sookyungsoo started with Excel entry errors that did not directly affect the service and maintenance-document organization. He said that finishing a document task in two days, instead of the minimum two weeks expected, changed the skeptical team’s response.

The team behind a webpage diagnostic service initially left almost all diagnosis to AI, but scores for the same page varied across three runs: 72, 85, and 61. It divided the work so code checked verifiable facts, such as the robots.txt file and specific tags, while AI handled judgments such as content credibility. Front-end developer Kim Eui-jung proposed automating deterministic processes, where the same input always produces the same output, and assigning tasks that require inference to AI. In a comparison using an OpenAPI specification—a document defining data formats between a server and a user interface—Orval’s code generator produced the same API code for free in about 1 second. AI took 2 to 4 minutes for the same task, cost about $1.4 per run, produced different code each time, and left 4 type errors on one run. Cho also built a tool to assess the amount of ink in signature fields in inspection binders exceeding 400 pages according to set rules, with help refining those rules through conversations with Claude Code.

Designer Sarah, who created the legal AI service Lawwin, said she delegates work only up to the point where a person can reverse an AI mistake. The service finds legal precedents similar to a user’s account and organizes the issues and response strategies, but directs decisions about whether to sue to a lawyer consultation. It does not promise a win and connects users with a lawyer, rather than AI, at the final stage where they incur the greatest cost. Sarah said AI has a wider role in a workplace-survival-style assessment service, where errors carry little burden. Park Dong-hyun, an HR employee at Classmethod, said AI agents automate repetitive tasks such as organizing performance reviews and processing salary data. For employee conflicts or sudden resignation consultations, staff use AI only to seek input; people make the final judgment and communicate with employees.

Writer Kkomi built a development tool that extracts requirements from planning documents and design drafts, then carries work through implementation and testing. Of 40 requirements, only 25 made it into the code, but the tool reported development complete because all written tests passed. Kkomi then had it create a traceability table linking each requirement to its acceptance criteria, tests, and code, and manually checks for missing links. If the same scenario fails nine times, the AI brings its logs and reasoning to a person for guidance. Cho’s tool found 68 cover pages in a 407-page binder, automatically verified 66, and placed the 2 ambiguous cases at the top of a report for human review. A tool comparing attendance records with leave requests likewise flags only mismatches or items needing review. In Kim Eui-jung’s development workflow, after a specification changes, a code generator updates types and an AI agent fixes compilation errors, while a rule-based tool first identifies where the errors are. Across the nine cases, the shared approach was to decide task recurrence, whether an answer is fixed, whether errors can be reversed, and where results should be checked before judging how smart AI is.