---
title: "Write Down Manual Decision Criteria Before Delegating to an AI Agent"
description: "A note on why you need to audit what operators actually look at before you start agentifying a workflow."
lang: "en"
canonical: "https://llm-lab.dev/en/posts/manual-decision-before-agent-note/"
source: "https://llm-lab.dev/en/posts/manual-decision-before-agent-note.md"
publishedAt: "2025-01-16"
updatedAt: "2025-01-16"
category: "技術メモ"
tags:
  - "agent"
  - "workflow"
  - "operations"
---

# Write Down Manual Decision Criteria Before Delegating to an AI Agent

> [!NOTE]
> What this article confirms
>
> Before delegating a workflow to an AI agent, the first useful artifact is not a tool list but a written map of the human decision criteria. Always-checked items, past information, hold conditions, escalation conditions, and the impact of mistakes define what can be automated and what should return to a human.
>
> The practical takeaway is to audit the existing manual judgment before designing agent behavior. If the operator is using context that is not visible on-screen, the agent design needs that boundary before any tool-calling implementation.

When I ask what tasks people want to hand over to an AI agent, I often hear something like, "I just check and judge things each time, so it seems automatable."

But when I dig into what that "just checking and judging" actually involves, it turns out to be surprisingly complex. Amounts, deadlines, counterparties, past exchanges, exceptional contracts, how busy the handler is. People are combining information that does not even appear on-screen inside their heads.

Before thinking about agentification, it seems better to first write out the manual decision criteria.

```text
Items always checked
Past information used for decisions
Conditions for putting on hold
Conditions for escalating to a human
Impact of getting it wrong
```

Doing this separates what can be automated from what should remain with humans. You do not have to hand everything over to AI. Even just making hold conditions and escalation conditions explicit has value.

I feel like designing an AI agent by starting with "let's make it capable of calling tools" is risky. What you need first is making visible what judgments people were making during the actual work.