Drafting authority is not spending authority
Everything your AI-assisted queue produces belongs on exactly one of three lists. Writing the lists down — with an owner and a date — is the escalation policy. Not writing them down means the policy gets decided ad hoc, per ticket, by whoever is most tired.
- AI may draft. Replies, case summaries, tags, macro suggestions. Draft, never send: everything here still passes the lesson-1 gate and a human review before a customer sees it. This list is long, and it's where the QJE study's 15% lives.
- Named approver required. Every customer-facing reply until its macro is eval-passed and the policy grants it — you built that ladder last lesson. And one category that never climbs the ladder: every refund and every adjustment, always. The AI drafts the refund reply; a human with refund authority executes the refund. Those are different verbs.
- Human only. Escalations, compensation decisions, anything touching legal or a regulator. The AI doesn't draft here — a model-written first draft of a chargeback response or a regulator letter anchors the human who edits it, and this is precisely where anchoring bias costs real money.
Why refunds never graduate
The temptation is obvious: the refund-status macro just went eval-passed, refunds under $20 are basically noise, why not let the macro close the loop? Because the two authorities are different in kind, not degree.
Drafting authority is about words being correct — checkable against a KB article, lintable against a banned list, evaluable against golden cases. You can earn it with evidence, which is what lesson 2 was.
Spending authority is about moving the company's money and making exceptions to its policy. Your company doesn't grant that by track record of good sentences — your team lead has a refund limit, and she writes well. An eval-passed macro has exactly as much spending authority as a really good template: none. The standing rule, from the same research base as the rest of this path: AI drafts, a human with authority executes. No eval score converts one authority into the other.
This maps onto the legal reality from lesson 1, too. Moffatt was about a statement — and the company paid for it. Now imagine the bot doesn't just misstate the refund policy but executes the refund it invented. The blast radius goes from one tribunal claim to a ledger.
The boundary is a routing decision, made per ticket
The three lists are static policy. The live question, forty times an hour, is: which list does this ticket fall on? That's escalation — and here the market teaches it wrong. The helpdesk courses on offer treat escalation as a routing feature: pick the tier-2 queue from a dropdown. What it actually is, on an AI-assisted team, is a decision discipline: a small set of conditions, written so precisely that a script — or a new hire on day one — reaches the same answer as your most experienced agent.
Klarna's failure territories tell you what the conditions must catch: emotional, multi-step, high-value. Next step encodes them.