Lessons 89.1–89.3 made your operation robust, affordable, and survivable. Now the final restraint — knowing where automation stops. Some tasks get worse the faster they run: a wrong public claim published at scale, a refund decided by a regex, a conflict answered with confident polish. Speed multiplies judgment errors; it never fixes them. This lesson draws the lines before scale tests them.

What stays human, and why

Six territories, each with a reason speed can't help:

• Strategy — which audience, which promise, which trade-offs. AI can brief the options; choosing the direction is ownership. Automating strategy produces efficient motion toward the wrong destination.
• Unusual conflict — the angry partner, the edge-case dispute, the situation with no precedent. Judgment here is the product; a template reply is an insult.
• Reputational repair — apologies, corrections, retractions. Trust returns through a person with authority accepting responsibility — never through an auto-generated "sorry you feel that way."
• Public claims — anything stated as fact to the world (pricing, capabilities, results, partner names). One hallucinated claim at scale is a correction campaign (Lesson 89.1's gates exist for this).
• Irreversible money movement — refunds, payouts, contract acceptance, account closure. Reversible drafts are fine; irreversible execution stays human (Lessons 84.4, 86.4, 88.3).
• Ambiguous high-stakes choices — low information plus real consequences: firing a vendor, changing access, medical/legal-adjacent content. When uncertainty changes what to do, a person decides.

The pattern across all six: consequences outlive the run. If the cost of being wrong exceeds the savings of being fast, it stays human.

AI's proper role: prepare, never decide

Around every red line, AI is still useful — in the assistant's seat:

• Prepare facts — gather the ticket history, the contract clause, the source passages, the cost numbers. Evidence in, opinion out.
• Lay out options — two or three paths with trade-offs and precedent, each labeled with confidence and gaps.
• Draft — the apology letter, the refund note, the strategy memo — clearly marked draft, awaiting the owner's edits and signature.

This is Lesson 84.4's "AI-prepared / human-approved" step made permanent for red lines: the machine does the reading, the human does the deciding, and the log shows which was which.

Exercise: declare your red lines

Create HUMAN-OWNERSHIP.md with at least three red-line decisions for your operation:

# HUMAN-OWNERSHIP.md — [Operation]

## Red lines (AI must never decide; human owner signs)
1. ___ (e.g. issuing refunds over $___) — owner ___ / AI prepares: ___
2. ___ (e.g. publishing public claims) — owner ___ / AI prepares: ___
3. ___ (e.g. closing/escalating conflict threads) — owner ___ / AI prepares: ___

## Enforcement (how the line holds under pressure)
- Technical: ___ (no credential / approval gate / blocked action — Lessons 88.2–88.3)
- Procedural: ___ (review queue SLA, second signer for ___)
- Log: every red-line decision records ___ + decider + date

## Violation drill
- If automation crosses a line: stop via ___ / notify ___ / correct by ___

Worked mini-example — local service business: red lines are (1) refunds over $50 — owner reviews the job history AI assembles; (2) public replies to negative reviews — AI drafts two options, owner posts; (3) changing a booking's price after confirmation — AI shows the contract and margin, owner decides. Enforcement: the workflow's credential literally cannot issue refunds (scope denial, not policy hope), and review replies sit in an approval gate with a 4-hour SLA.

Finish line: a HUMAN-OWNERSHIP.md with three named red lines, an owner and AI-prepare role for each, and enforcement that's technical (not just promised).

Verify quickly: for each red line, ask "could the workflow physically do this by accident tonight?" If yes — credential too broad, gate missing — the line is decorative. Make crossing it impossible, not impolite.

Common failure mode: the ambitious autopilot — "AI handles refunds and replies; I just check weekly." Weekly review of irreversible actions is archaeology, not ownership. Red lines need gates before the action, not reports after it.

Check your understanding

1. What do all six human territories share that makes speed harmful there? 2. What's the difference between AI preparing and AI deciding — and how does the log show it? 3. Why must enforcement be technical (scopes, gates) rather than a written promise?

Next

Boundaries drawn. One ritual keeps everything — quality, cost, runbooks, red lines — improving instead of drifting: the weekly operating review. Lesson 89.5 builds it.