Green, Yellow, Red: The Three Lanes of Trust

Chapter 3 placed checkpoints inside workflows. This chapter zooms out to the question those placements were quietly answering: how much do you trust AI output, for which jobs? Vague trust produces two expensive behaviors: re-reading everything, which erases the savings, or reviewing nothing, which works until the day it very much does not. The fix is to decide trust per task, in advance, in writing. Three lanes.

Green: runs and ships without review

Output whose failure costs you a shrug. Internal drafts, brainstorm lists, first-pass schedules that later steps reshape, meeting-note summaries for your own eyes, the research pass before a decision you will make yourself. If a green-lane output is wrong, the wrongness is caught naturally by the next thing that touches it, or it simply gets ignored at no cost.

The green lane is where the scatter tax dies. Owners who mistrust it, re-reading internal summaries as though a customer might see them, spend checkpoint minutes where no unwind cost exists, and those minutes are exactly what the system was built to reclaim.

Yellow: drafts freely, ships only through a checkpoint

Anything a customer, supplier, or employee will read, and anything with a number a human will act on. Quotes, invoices, marketing pieces, follow-ups, the crew schedule once it is final, the Friday analysis. Yellow is where most of your workflows live, and its rule is exactly the checkpoint discipline you built in Chapter 3: a named, bounded check between draft and send.

Yellow's failure mode is lane creep. After a month of clean drafts, the two-minute quote review starts feeling like ceremony, and the temptation is to promote quoting to green. Hold the line, and not because the drafts are secretly bad: the checkpoint is cheap insurance priced at two minutes, covering a mistake priced at a customer. Lanes are assigned by unwind cost, and a sent quote's unwind cost does not drop just because recent drafts were clean.

Red: the AI assists; a human writes

Where being wrong, or being discovered to be machine-written, does damage no correction can fully undo:

• Anything legal: contracts, warranty disputes, terminations, insurance claims
• Personnel matters: reviews, discipline, references, anything in an employee file
• Apologies for real harm: an injury, a damaged property, a genuinely botched job
• Pricing exceptions and negotiations where the relationship is the asset
• Anything you would want a lawyer, accountant, or insurer to see first

Red does not mean the assistant is banned from the room. It drafts questions you should consider, summarizes the history, pressure-tests your draft ("what could go wrong sending this?"). What it does not do is produce the text that gets sent. A red-lane recipient is owed a human author, and some of them, courts and insurers among them, may one day ask.

Write the lanes down

The lane list is a half page in your system folder, and every skill gets a pointer to it. Our HVAC version:

Two payoffs for writing it down. Employees inherit your judgment instead of improvising their own, which is Chapter 6's whole handover made possible. And disagreements become visible: when your office manager assumes review replies are green and you assumed yellow, the written list is where that assumption surfaces, before the one-star reply does.

Try it now

Draft your lane list now: every task from your audit plus the outputs of your existing workflows, each assigned a lane by one question, "what does it cost to unwind this if it is wrong?" Then find the yellow item you have been quietly treating as green, most owners have one within a week of building workflows, and put its checkpoint back. The next lesson arms that checkpoint with the checks that catch what clean-looking drafts hide.