← Back to Concept Contrast

Concept Contrast Fable

The rules for how companies use AI and the rules for how agents do work

Conclusion first:Both are called policy, but they govern different levels. Company-level rules determine which businesses may use AI, who approves it, and how data is handled; agent-level rules determine what tools a given agent can call and how far it can act on someone's behalf.

Both are called policy, but they govern different levels. Company-level rules determine which businesses may use AI, who approves it, and how data is handled; agent-level rules determine what tools a given agent can call and how far it can act on someone's behalf.

Fable

The city's public notices and the machine's manual

Many shops opened in Riverbend City—mills, pharmacies, and money houses all bought machines that turn automatically. The city hall worried about trouble, so it posted a notice at the city gate: which industries may use machines, which tasks must have a person watching, and which ledgers must not be fed to machines. Shopkeepers entering the city each day first read this large notice.

After the notice was posted, the mill owner was delighted. Seeing that grain processing was listed as permitted, he immediately moved the new machine into the back yard and even tore down half of the old stone mill. But hanging beside the machine was a thin manual that read: only dry grain may be milled, not wet wheat; it must not be left unattended at night; if stones are encountered, the machine must be stopped.

The shopkeeper thought city hall had already nodded, that the manual was just fine print, so he had the workers run the machine at night too. In the middle of the night damp wheat jammed the gears, the whole belt snapped, and the warehouse was covered in flour. The next day he went to city hall to argue, and the official asked only one question: did you use that machine according to its own rules?

The pharmacy also encountered the opposite problem. A city hall notice said that herbal sorting could use machines, but prescription decisions must be handled by the attending physician. Yet the pharmacy's sorting machine manual was very strict: anything with similar drug names, similar colors, or unknown origin must be placed in the manual review basket. The assistant, finding it slow, emptied the review basket and nearly mixed up two medicines.

Only then did the shopkeepers realize that the city gate announcement governs the boundaries of the whole city: which industries can use machines, who is responsible, whom to seek when an accident occurs. The machine manual governs the actions of one machine: what it can touch, where it can stop, and whether to call a person when it encounters a difficulty. The announcement gives entry permission; the manual gives the measure of action. Neither rule alone is enough.

Later the city hall held the briefing again. Officials hung the announcement on the wall and placed the manuals for each machine beside it. Shopkeepers first had to confirm that their industry was permitted, then confirm whether the machine in hand could perform that specific task. The city hall did not start the machines, and the manuals could not approve a business for the city either.

From then on, Riverbend City had far fewer disputes. The grand proclamation decided whether a category of things could enter the city, and the manual decided how a particular machine should operate in the shop. The shopkeepers finally stopped using a city gate proclamation to replace the few pages of detailed rules beside each machine.

AI policy

Organization-level AI usage rules that explain which business activities may use it, how approval and responsibility are arranged, and where the data and compliance boundaries lie.

agent policy

The operating rules of a single agent, limiting which tools it can call, which data it can access, and which situations require it to stop or escalate to a human.

Story Mapping

  • Town hall announcement → organization-level usage boundaries for AI policy
  • Which industries can use machines → which business scenarios are allowed to use AI
  • Machine manual → monolithic behavior rules of agent policy
  • Wet wheat jams the gears → looking only at organizational permission and ignoring execution limits causes accidents
  • Review basket → scenarios where the agent must escalate to a human
  • Read the announcement first, then the manual → company-level rules and agent-level rules must both be satisfied

Put into governance design

When making an AI policy, first answer company-level questions: which departments can use it, how data is classified, who approves, and who is accountable. When designing an agent policy, then get down to the specific agent: whether it can send messages, whether it can modify databases, whether it can spend money, and in what situations it must ask a human for confirmation. The former draws boundaries for the organization; the latter installs guardrails for the executor.

京ICP备2026014272号