How much a department may do alone

Every company that brings in AI meets the same tension. Let it do too much on its own and people stop trusting it. Let it do too little and you have built a slow way of doing what a person could do faster. The answer is not an on/off switch. It is a scale.

Most tools treat autonomy as one setting for everything. That works when one assistant does one thing. It falls apart when several departments work at once, each with its own risk. A marketing department writing a post and a legal department drafting a contract clause should not be trusted with the same amount of independence.

Higher means more human

Vepartment's scale — the Human Agency Scale, HAS², part of its hierarchical agentic system — has five levels, and one rule: a higher number means more human control. That is the opposite of the scales used for robots, where higher means more machine independence. In a company, the person is the authority, and the scale says so.

The five levels

• Level 1 — Routine on its own. The department does the small, low-risk, repeated steps it has been allowed, and reports: scheduling what you already approved, keeping records up to date, the weekly numbers.
• Level 2 — Do, and flag. The department does the work and stops at anything unusual: a number out of range, a customer it has not seen, a rule it is not sure about.
• Level 3 — Propose. The default. The department drafts and recommends; you choose, change and approve before anything happens.
• Level 4 — You lead. The department prepares the analysis and the material; you make every decision.
• Level 5 — Every step on your instruction. For legal filings, money and anything regulated.

Per department, per task, per risk

The level is not one setting for the company. A marketing department may start at level 3 and a legal department at level 5, because that is what their work deserves. Inside marketing, a blog post may run at level 2 while a paid campaign needs level 4. And when something looks wrong mid-task — an unexpected cost, a brand-safety concern, a rule that applies — the level rises for that task before it goes on.

A yes is built into the engine, not into the screen

From level 3 up, a department's work stops at an approval before it moves. That stop is not a button someone could forget to draw. It is in the engine that runs the work: a department cannot skip it, and the only thing that opens it is a person's yes, recorded with the time. Whatever the level, anything sent to someone outside the company waits for that yes.

Trust is a property of the design

If you cannot say exactly when the system will stop and ask, you do not have governance; you have hope. The scale makes it plain. Every department shows its level. Every task shows what it needs a yes for. Every escalation is written down. That is how a regulated, careful company can use this — not because the model is promised to be safe, but because the person is always the one in charge.