Directing AI all day? How to move toward delegated development

EDITORIAL PRACTICE

Faster output does not automatically free your attention

Consider a screen implementation: ask AI to build it, review the result, request a correction and check it against another feature. Individual steps may be shorter, yet repeated interruptions can still prevent focused work on the business.

For one task, record time spent preparing instructions, checking outputs, requesting corrections and making the final decision. Keep that separate from the time the model is running. The record makes the work you want to delegate concrete.

FROM PRIMARY SOURCES

Read productivity findings in their original context

A 2025 METR experiment involved 16 experienced open-source developers completing 246 tasks in familiar projects. In that setting, work took 19% longer with AI. This is a historical, setting-specific result, not a measure of all AI-assisted development today.

In its February 2026 update, METR described selection effects that made the follow-up experiment an unreliable measure of current productivity. Teams should also evaluate their own work.

Sources: [1][2]

EDITORIAL PRACTICE

Separate important decisions from everyday management

Before adding more AI users, agree who decides what. Delegating the daily loop to an overseas team does not require leaving business goals or additional spending undefined.

EDITORIAL PRACTICE

Hand over acceptance criteria with the request

A vague request creates repeated questions for both people and AI. Explain the user, the situation, exclusions and how completion will be checked. When Japanese requests are structured and translated for the team, keep terminology and acceptance criteria consistent.

• Purpose: whose work should become easier?
• Deliverables: working screens, code, tests or another agreed output.
• Acceptance: which actions must work before the task is complete?
• Reporting: when and how are progress and questions consolidated?
• Approval: who decides on extra spending and production release?

EDITORIAL PRACTICE

Begin with a bounded piece of work

Choose a small feature with an observable outcome. Afterwards, review client instruction time, accepted work, rework and total cost together.

Fewer conversations are not an improvement if quality drops or downstream fixes increase. Evaluate both client attention and delivered results before expanding the engagement.