support studio

The friendliest sentence in the queue is the most expensive one

Air Canada's website chatbot told a grieving customer he could book a full-fare flight and claim the bereavement discount afterward. The policy said the opposite. In 2024 a British Columbia tribunal (Moffatt v. Air Canada, 2024 BCCRT 149 — small-claims, persuasive rather than binding, and still the case everyone in support now knows by name) ordered the airline to pay up, holding it responsible for everything on its site, "whether from a static page or a chatbot." The bot didn't get fired. The company got the bill.

That's the job now. You're not racing the drafter — you're the reason its drafts are safe to send. AI writes a warm, confident, plausible reply in two seconds. Whether that reply matches the actual return window, cites a KB article somebody verified this quarter, and stops short of promising a refund you have no authority to give — that's yours. This chapter builds that checking as code you run, not vibes you hold in your head on ticket forty of a Friday shift.

Why it matters

• The evidence for AI-assisted support is the best in any occupation — and it's evidence for assist, not replace. The Quarterly Journal of Economics study of roughly 5,200 support agents (Brynjolfsson, Li & Raymond, May 2025) found a 15% average productivity gain from an AI assistant, with the newest agents improving in both speed and quality — while the most experienced agents saw small speed gains and small quality declines. The review step is where seniors put their edge back in.
• The cautionary tale already happened, publicly. Klarna launched its AI assistant in February 2024 claiming 2.3 million conversations in month one — the work of some 700 agents. By May 2025 it was rehiring humans, its CEO admitting "we focused too much on cost. The result was lower quality." The failures clustered exactly where you'd guess: emotional customers, multi-step problems, edge cases.
• Nobody teaches the checking. Vendor academies teach you to deploy AI answers in a few clicks. Almost nothing on the market teaches verifying an AI answer against policy before a customer sees it — the one skill the Air Canada case made mandatory.

What AI specifically gets wrong

• It promises what you can't deliver. "Guaranteed by Friday," "full refund," "this won't happen again" — a drafter has no refund authority, no carrier control, and no crystal ball, and it writes all three fluently. The banned-promises list exists because the friendliest sentence is the one that binds the company.
• It cites from memory, including stale memory. A draft that leans on a KB article nobody has verified since the policy changed is a confident wrong answer with a footnote. An unverified article is not a citation target.
• It answers the ticket in front of it, not the customer behind it. Third email this week, "I'm done with you people," a chargeback mention — a macro sees "where is my order" and fires. Those are exactly the tickets that must leave the queue for a human, and the triggers have to be conditions in code, not judgment calls under load.
• It scales a mistake to every matching ticket. One agent's bad answer is one apology. A macro's bad answer runs until someone pulls it — which is why every macro earns autonomy against golden cases before it's trusted, and loses it the moment an incident proves the eval set had a hole.

What you'll be able to do at the end

• Encode your support pack — tone rules, banned promises, a KB index with last-verified dates — as data a script can hold every draft against.
• Run a draft gate: banned-phrase lint plus citation-freshness check, SEND or HOLD with named reasons, on every suggested reply before it goes out under your name.
• Build macros as intent + approved answer + never-touch boundaries, and run each one against its own golden eval set — pass every case or stay review-required, no exceptions, not even for the macro with the impressive demo.
• Route tickets across the three lists — AI-may-draft, draft-plus-named-approver, human-only — with escalation triggers a new hire could apply on day one, and run the incident drill: pull the macro, send the correction as a human, and feed the failure back into the eval set so the gate is stronger than it was yesterday.

Every drill runs in plain Python, no helpdesk login required. The logic is the point: it's the same gate whether it lives in a script, a Zendesk trigger, or the agent-assist tool your company buys next quarter.