Theory of Work

Framework 01In developmentRevised 12 Aug 2026v0.9

The Delegation Ladder

Canonical definition

The Delegation Ladder grades six levels of machine autonomy, from “human does the work, machine drafts” to “machine does the work, human audits samples.” Each rung carries a different verification duty.

Per-output inspection Per-policy design COMPOSITION OF DUTY · TOTAL CONSTANT MACHINE AUTONOMY → THE SEAM · INSPECTING OUTPUTS → DESIGNING POLICY 01 Machine drafts Read every line 02 Human edits Edit every line 03 Reviews all Review each output 04 Exceptions only Define the rule 05 Sampled audit Set the sample rate 06 Self-checking Audit the checker THEORY OF WORK
Fig. 01 The duty does not shrink as autonomy rises; it stops being inspection and becomes policy design, and the switch happens between rungs 03 and 04.

The problem

Watch what a team actually reads. The low-stakes draft gets read line by line, because reading it is an old habit from when they wrote it themselves. The high-stakes summary, the one that goes to the board, gets a glance. It arrived formatted and confident, so it looked finished.

Effort is being spent. It is simply not being spent where the exposure is.

Underneath that is a vocabulary problem. Teams describe machine involvement in a single word. A task is "automated" or it is not; a report is "AI-assisted" or it is not. The word says nothing about who is accountable for the output, so the question of who checks it never gets asked out loud. It gets settled quietly, by whoever happens to be holding the work when it ships.

Why current approaches fail

Two instruments are usually reached for. The first is a policy document listing approved and prohibited uses. It governs tools rather than tasks, so it dates within a quarter and says nothing about the review a given output needs. The second is a human-in-the-loop mandate, which names a participant without naming a duty: a person who signs without a defined check is a signature, not a control.

Both fail on the same point. They treat autonomy as a permission granted once, rather than a position on a scale that decides the shape of the check.

The framework

Six rungs, ordered by how much of the work the machine performs. Each rung names the verification duty that comes with it. The duty is the point: a rung you cannot discharge the duty for is a rung you are not on, whatever the tooling permits.

The duty does not shrink as autonomy rises; it stops being inspection and becomes policy design, and the switch happens between rungs 03 and 04.

  1. Machine drafts. The human does the work; the draft is raw material. Duty: read every line.
  2. Human edits. The machine produces the first version. Duty: edit every line, not approve it.
  3. Human reviews all. The machine completes the task. Duty: review each output before release.
  4. Exceptions only. The machine releases unless a rule fires. Duty: own the rule and its false negatives.
  5. Sampled audit. The machine releases; a share is checked after the fact. Duty: set and defend the rate.
  6. Machine self-checks. A second system checks the first. Duty: audit the checker on a schedule.

How to apply it Monday morning

Take one team and list the ten tasks it performs most often. Against each, write the rung it is actually on today, not the rung anyone intended. Then write the verification duty being discharged in practice. The gaps show up in one pass: tasks at rung four with a rung-one habit attached, and tasks at rung five where nobody owns the sample rate.

Fix the mismatches by moving the duty, not the rung. Naming an owner and a rate for two tasks is a week of work, and it changes how the team talks about the other eight.

Where it breaks

The Ladder assumes the task has a checkable output. For work whose quality only shows up months later, like hiring or strategy or most research, the rungs collapse. What you need there is a different instrument, one that routes by how reversible a decision is rather than by how much of the work the machine did.

It also breaks where the output is checkable but the check costs more than the work, which is common in high-volume classification. Rungs four and five are then a budget decision, and the Ladder does not make it for you.