Free resource / Version 1.0
A field guide for turning one recurring workflow into bounded AI leverage—without hiding the evidence, surrendering responsibility, or automating the wrong decision.
Print-friendly: use your browser’s Print command to save a clean field copy.
Give AI a bounded job, not a vague ambition. Make the evidence visible. Keep a person accountable for the consequence.
01 / The six-line delegation canvas
If you cannot describe these six fields, the workflow is not ready for more autonomy. Use the example language as a starting point, then replace it with the reality of your work.
Name the finished result in one sentence.
ExampleA decision log with owners, dates, and unresolved questions.
List the sources the system may use—and nothing else.
ExampleTranscript, agenda, project roster, and approved terminology.
Show what acceptable work looks like.
ExampleUse this prior decision log as the format and quality bar.
State what the system must not infer, change, or send.
ExampleDo not assign an owner who did not accept the work. Do not email anyone.
Require the output to show where factual claims came from.
ExampleAttach the source timestamp to every decision and open question.
Mark the exact step where a person decides.
ExampleThe meeting owner approves the log before it enters the project system.
02 / Seven workflow patterns
Pick one pattern. Run it alongside the current process before replacing anything. The first goal is not maximum autonomy; it is a smaller, more legible loop.
03 / The autonomy ladder
Begin at Draft for a new workflow. Move upward only when repeated runs meet the standard, exceptions are understood, and the team can reverse or correct an action.
The system watches a completed workflow and identifies the steps, inputs, exceptions, and decisions.
The system prepares work. A person checks every output and takes every action.
The system presents options, evidence, and a recommended next step. A person decides.
The system prepares an action but cannot complete it until a person authorizes it.
The system acts only inside tested limits, logs what it did, and escalates exceptions.
Track correction rate, missed exceptions, unsupported claims, time returned, and whether the human reviewer can still explain why the result is trustworthy.
04 / A seven-day pilot
Run the workflow normally. Record elapsed time, active attention, corrections, delays, and exceptions.
Complete the six-line canvas. Collect one strong example and one failure example.
Let AI prepare the work with no autonomous action. Mark every correction and unsupported inference.
Tighten inputs, standards, and stop conditions based on what failed—not what the demo promised.
Run a second real case. Confirm that the improvement survives different inputs.
Test a known edge case. The right response may be to stop and ask for help.
Keep, revise, or stop. Advance one autonomy level only if the evidence supports it.
The one-page review
Keep the useful signal
Join Issue 001 now. We’ll email you when the first five-minute Sunday brief goes live.