THE GUARDRAILS & KILL-SWITCH SHEET From "The One-Person AI Business — Intermediate", Chapter 8 ================================================================ You don't delegate "functions" — you delegate dozens of small ACTIONS, and they are not equally dangerous. Something fast, tireless, and confident-even-when- wrong will one day make a bad call with nobody watching. This sheet is the governance that means it can't do irreversible harm behind your back: a map of what needs a human before it goes live, and a kill switch for when something slips through anyway. ⚠️ Write this BEFORE you need it. The one certainty about the middle of a live incident is that you won't think clearly inside it. SORT EVERY ACTION BY THREE QUESTIONS ---------------------------------------------------------------- REVERSIBLE? Can it be undone? (a draft: yes. a public post, a sent email, a released payment: no, or only at a cost.) REACH? Who sees it if it's wrong? (just me / one customer / the public) COST? Money, trust, or legal exposure if wrong? (trust is the expensive one — it's the one you can't refund.) The more an action is irreversible × public × costly, the more it must pass a human BEFORE it goes live. THE RISK MAP (sort your actions into three buckets) ---------------------------------------------------------------- RUNS AUTOMATICALLY (low-risk, reversible, private): ______________________________________________________________ RUNS BUT I SAMPLE IT: ______________________________________________________________ HUMAN EVERY TIME, FOREVER (public claims · promises · exceptions · legal · the MOVEMENT of money): ______________________________________________________________ ⚠️ Trust is granted by RISK CLASS, not by FUNCTION. "All demand posts on automatic" is the mistake — a public factual claim must stay human-gated even when the rest of demand has earned autonomy. ⚠️ THE MONEY RULE: AI may CHECK and PREPARE a rule-qualifying refund up to your ceiling; a HUMAN approves AND executes every actual movement of money. AI never approves or moves money itself, so it never needs payment wiring. THE KILL SWITCH (a written procedure — four steps, in order) ---------------------------------------------------------------- 1. DISCONNECT — pause every autonomous action at once. Revoke posting/sending permissions or flip each function to draft-only. How, exactly, for me: ____________________________________________________________ 2. CONTAIN — get the correction out with a PRE-WRITTEN retraction template. Where my template lives: ___________________________________ 3. AUDIT — find the blast radius, clear the pending queue to zero: what went out, who saw it, who acted on it, what else is queued. 4. RESTART — bring functions back GATED, not automatic; re-earn autonomy only on evidence, one class at a time. LIABILITY (name it — this book does not advise on the law) ---------------------------------------------------------------- You are responsible for what your AI does. Depending on how your business is set up and where you live, the exposure may fall on you personally. When the stakes are real, ask one professional in your country one specific question. THE FORCING COMMITMENT (dated — before you need it) ---------------------------------------------------------------- "As of ______, these classes require my review before going live: __________. These may run automatically: __________. My kill switch: __________ (the exact steps to pause everything, correct, and restart). I wrote this down before I needed it." ================================================================ Yours to keep — free, and no strings attached. One tool from the ePORTS One-Person AI Business series. This one is from Intermediate — the governance that makes delegating to AI safe for a business only you are liable for. The full shelf, whenever you're ready: eports.global/one-person-ai-business ================================================================ ePORTS — your ecom source. eports.global.