Not a manifesto about AI transforming everything — a concrete list of shop chores that stop existing after a day or two of engineering each, with honest hour estimates and one rule for what to never automate.
Every small shop runs on invisible chores. Nobody budgeted for them, nobody loves them, and they expand quietly until someone is spending Thursday afternoons copy-pasting order numbers into emails. We build products for a living, and still catch these chores growing in our own operation — routine is a weed, and the only lasting herbicide is engineering.
This is not a manifesto about AI transforming retail. It is a list of seven specific jobs that stop existing after a day or two of honest work each — with effort estimates from doing them, and one rule at the end about what should never be automated.
The pattern that matters is the outbox: the shop writes “email owed” into a database table in the same transaction as the order change, and a separate worker sends and retries. The naive version — fire the email during checkout and hope — silently drops messages the day your email provider hiccups, and you find out from an angry customer. The outbox version survives provider outages and never blocks a sale. We run every transactional email this way; the retry log has already paid for the half day several times.
Google Merchant Center, marketplaces, price aggregators — each wants a feed, and each hand-exported spreadsheet is wrong within a week. Generate feeds from the same database that renders your catalog and they cannot drift: price changes propagate everywhere on the next fetch. The day of work replaces a permanent monthly chore plus the occasional “why does Google show the old price” investigation, which always lands during a sale.
A nightly database dump with rotation is two lines of cron. The part everyone skips is the drill: actually restoring the dump somewhere, monthly, on the calendar. An untested backup is a hope, not a plan — the difference surfaces at the worst possible moment, and we say this as a team whose habit of daily dumps once turned a dead hosting provider from a catastrophe into a lost weekend.
A health endpoint that checks the database and mail queue, an external monitor that pings it, and an alert that reaches a phone. Two hours, total. The alternative monitoring system is “a customer emails you”, which has excellent coverage and catastrophic latency. The shop should tell you it is sick before Twitter does.
Sort last quarter's support inbox by frequency and the top five questions are probably licensing, delivery, refunds, compatibility, invoices. Answer each once, publicly, on the product page — then link, not retype. We wired public product Q&A into our own store for exactly this reason, with a bonus we did not fully anticipate: the answers get indexed, and AI assistants quote them. Support content is marketing content wearing overalls.
Orders, revenue, top products, failed payments — one query each, one scheduled email every Monday morning. Not a dashboard; dashboards are where curiosity goes to feel busy. A pushed email gets read, and its real function is anomaly detection: the Monday you glance at it and something feels off, it has done its job for the year.
If a human still SSHes into the server to release, automate that first, before anything on this list. Push to the main branch → tests run → server updates itself → smoke check confirms the homepage answers. Ours even refuses to deploy when tests fail, which has saved us from ourselves more than once. Releases stop being events and become punctuation.
Doing something once is discovery. Twice is coincidence. The third time, stop and script it — before the fourth, while the details are still fresh and the annoyance is still motivating. And keep one category human forever: the apology, the edge-case refund, the reply to a customer who is genuinely upset. Those messages are where a small shop out-competes a big one, precisely because a human wrote them. Automate the routine so you have time to write those well — that was the point all along.