When AI Actually Slows You Down

The marketing promises 10x productivity. The reality is more honest. There are specific kinds of work where reaching for an AI agent makes me slower, not faster. Naming them saves me the cost of learning it the hard way on the next client project.

Tasks Where the Prompt Costs More Than the Code

If I can type the solution in under two minutes, writing the prompt is slower. Renaming a variable across five files. Adding a missing import. Changing a color token. The prompt takes longer to compose than the edit takes to make. Use the editor, not the agent.

Debugging Unfamiliar Runtime Behavior

When I am staring at a stack trace from a library I have never used, the agent's first answer is often a plausible-looking fix that addresses a different bug. Debugging requires stepping through the actual state. The agent does not have access to my running process. I do. Pair the agent with the debugger, do not replace the debugger.

Architecture Decisions Early in a Project

Agents are great at executing a chosen design. They are terrible at choosing between two designs when the tradeoffs depend on business context they do not have. When I am still deciding between DynamoDB and Postgres for a new feature, asking the agent gives me confident-sounding answers with no weight behind them. That decision needs human judgment and stakeholder context.

Learning a New Technology From Scratch

When I pick up something I have never touched, the agent can scaffold it, but I cannot tell whether the scaffold is idiomatic. That reading comprehension comes from time spent with the docs and the community, not from more prompts. I budget a week of hands-on exploration before I let the agent lead on anything new. The agent is a speed multiplier for work I already understand, not a substitute for understanding.

A Simple Rule I Follow

Is this task mostly mechanical? -> Agent
Is this task mostly judgment?   -> Human
Is this task both?              -> Human decides, agent executes

The Honest Math

I timed myself for two weeks. Across 47 recorded tasks, the agent was faster on 31, slower on 9, and roughly tied on 7. The slower ones were almost all in the three categories above. Knowing which bucket a task falls into is the skill. See Anthropic's prompt engineering guide for what actually works.

Using AI well is knowing when not to use it. That took me eight months to learn.