One support rep, 35 hours a week on "where's my order"
You run a store, 3,000+ orders a month. 80% of support messages are variations of one line: "where's my package," "how do I return." One rep burns 35 hours a week on replies, near burnout. What if you swap in AI? A Canadian store at that scale actually tried it — the result is worth dissecting.
How they deployed it
The store put a Chat + email agent on top of the order system, handling order status, returns, refunds (auto-approved up to $100), escalating anything complex to a human.
Build cost $14,500 CAD, $320 CAD/month to run, shipped in 5 weeks.
The numbers, translated
Year one: 67% of tickets fully auto-resolved, support team cut from 2 FTE to 1, saving $52,000 CAD in salary a year. Payback in 4.3 months, ROI 184%.
The feel: like hiring someone who earns back the recruiting fee by month five, then hands you a free headcount every month after. For high-volume, support-heavy stores, that math is clean.
The catch: 0.3% hallucination nearly bit them
Not all smooth. In 0.3% of chats, the AI invented return-policy details — not from the knowledge base, but off the top of its head. Small rate, but one slip is a complaint or a dispute.
They had to add a hard rule: the agent must quote policy verbatim from the knowledge base, no paraphrasing, no summarizing. That wall killed the hallucinations. Also, auto-resolve landed at 67%, under the 75% projection — below forecast, but still strong.
Lesson: AI takes the repeat work, but lock the edges
- High-frequency, rule-clear support (tracking, returns) → AI genuinely saves a headcount, ship it
- Critical policy (refunds, payouts) → lock to "verbatim quote," never let the model freelance
- Low-volume stores → hold off: the $14.5k build won't amortize, payback stretches
Why this matters to you
If you run a Taobao store, a Shopify site, any support-heavy business — AI support isn't a gimmick, it really absorbs that 67% of repeat Q&A. But don't go fully hands-off: split "auto-answerable" from "needs a human," lock policy tight, escalate the messy ones. The headcount you free up only pays off if you point it at revenue, not firefighting.