A Claude-built support inbox that drafts replies, pulls orders and learns from the owner's edits
A small online seller had Claude build Relay, an internal web app that turns email, Facebook and Instagram messages into tickets every 10 minutes, drafts replies with order data, and handles cancellations and returns after his approval.
Evidence: The author reports this. We have not checked it beyond reading the source.
The business problem
After acquiring a store with its own website, the owner started receiving direct customer messages across email and social channels, and answering them took up his mornings.
What was tried
1. Connect the inboxes the AI can reach (Gmail, Facebook, Instagram). Check each platform's API limits first: here Instagram direct messages can be sent by the app, but Messenger replies still need copy and paste. 2. Ask Claude to build a small progressive web app that creates one ticket per email, message or comment, with a job pulling new items every 10 minutes. 3. For each ticket, Claude identifies the request, looks up the order in the store or marketplace when needed, and drafts a reply. 4. Review every ticket: edit the reply, dismiss it, or leave feedback. A checkbox decides whether your edited reply is saved to the assistant's knowledge base, so one-off answers for unusual cases do not become standard replies. 5. Add tools for repeat jobs: order cancellations with refund after verification, address changes (cancel and reship), comment moderation, asking customers for permission to use their photos, and returns (photograph the return label, the app reads the tracking number, finds the order, and refunds the items you tick as received). 6. Next step the author describes: log whether each proposed action was accepted or edited, and let ticket types you always approve run automatically, starting with deleting offensive comments.
What was reported (positive)
The author says the app saved him hours of time in under two weeks. No ticket counts, response times, customer satisfaction figures or costs are reported.
Limitations
Self-reported, under two weeks of use, no numbers. The business sells physical products (binocular harnesses) through its own Shopify site and Amazon, so the order, return and shipping steps are specific to physical goods; the inbox, drafting and feedback-loop parts transfer to online services and digital products. The Claude model version is not named and no code is shared. The author notes that the assistant began reusing his one-off replies as general answers until he added the feedback controls. Some Meta channels cannot be fully automated.
What you need
Access to Claude and Claude Code, API access to your email, social accounts and store, and willingness to review every reply at first.
Keep a human approval step on refunds and cancellations until the acceptance log shows the assistant is reliable for that ticket type.
Sources
- The Automated Operator (Alex Willen) ↗ Firsthand write-up, published October 15, 2026
Source published: October 15, 2026. Last reviewed here: October 26, 2026. Spot a mistake? Tell us.
Tools in this workflow
- Claude: Drafts replies and runs ticket actionsTry Claude
- Claude Code: Builds and changes the appTry Claude Code
- Shopify: Order data for the storeTry Shopify
- Shippo: Return labelsTry Shippo
Get new workflows like this by email
A short selection of the most useful new workflows, including the ones that did not work. Free, and you can unsubscribe at any time.
Related workflows
Five-part checklist for setting up scheduled AI jobs that you can trust to run unattended
A Build to Launch guide maps the scheduled-task options in Claude, ChatGPT, Perplexity, n8n and cron, and gives a checklist for each job: define a good result, set limits, keep run receipts, and name an owner and off switch.
Automate monthly refreshes of data-driven articles with an agent that fetches, filters, rebuilds and drafts for human review
An Ahrefs marketer describes a monthly pipeline that pulls fresh keyword data, filters it with written rules and examples, rebuilds tables, updates WordPress drafts and notifies the editor, who approves every change before it goes live.
Build a five-email onboarding sequence with stop rules and a dry-run cron job, using a copy-paste prompt for your AI coding tool
A beginner guide for first-time SaaS builders: pick one first-win action, log signups and sends, map five plain emails over 14 days with stop rules, send them from a scheduled job after a dry run, and measure activation instead of opens.
Cut LLM costs in a small SaaS with flex-tier fallback, cache-friendly prompt order, side-by-side model migrations and backend circuit breakers
A bootstrapped founder's four habits for the AI calls inside his product: try the cheaper flex tier first and fall back on rate limits, order prompts so the long repeated part is cached, run old and new models side by side before switching, and cap spend in the backend.