I sell AI customer support software and my own second-line support was still manual. Fixed that this week.
Mildly embarrassing admission: I run SiteSpeak, an AI chatbot platform for customer support, and until this week my own second-line support was fully manual. The chatbot on our site answers most visitor questions, but everything it escalates (bug reports, billing, prospects who want a human) landed in an inbox I worked through by hand.
Background that made this easy: for the last ten weeks I've had a launchd job run Claude Code at 6am with a "morning report" instruction file. It pulls traffic, new Sentry issues and uptime through MCP servers and emails me one report. 67 runs. This week I added a support pass to it:
- Read every support-inbox conversation since yesterday (via our MCP server) and classify it: urgent, prospect, or noise.
- Investigate before drafting. Bug reports get checked against production errors in Sentry and against the actual code. How-to questions get verified against the live docs. Rule in the instruction file: anything unverified stays out of the draft.
- Check my email client for an existing thread from that person. Visitors email support minutes after chatting with the bot, and replying twice looks terrible.
- Draft the reply in my email client. It is not allowed to send. I review and send everything.
- Archive the thread, log it to a state file so runs never double-process.
First run was this morning: 5 threads, 3 drafts, 2 correct skips (one no-email, one already answered by email). It also caught our own bot giving a customer a wrong answer the day before and flagged it for a same-day human correction, which honestly stung and delighted me in equal measure.
Volume reality check: a handful of escalated threads a day. The value isn't hours saved, it's that every thread gets investigated properly, prospects hear from a founder quickly, and support stopped being the thing I do worst at 8am.
We built an MCP server for SiteSpeak, which is what made me finally try this, but nothing in the recipe is ours-specific. Any support tool + error tracker + email client with MCP servers gives you the same loop, driven by one markdown file.
Happy to share the instruction file structure if anyone wants to copy it.
Archive the thread only after the reviewed reply is sent, not when a draft exists. Otherwise an interrupted review can turn into a silent drop. A drafted / reviewed / sent state machine is probably more important than the classifier.
The interesting improvement isn't drafting support replies with AI—it's requiring the AI to earn the right to respond by investigating first. I'd keep validating whether customers ultimately value faster responses or the higher accuracy that comes from connecting support, code, documentation, and production data before an answer is written. That's a much harder capability to replicate.