AI workflows handling recurring customer-facing internal ops without manual touch.

Two production workflows shipped at a US-based staffing agency: marketing extraction from meeting transcripts, and bilingual draft generation for customer service-pause requests.

The problem

The management team was hand-writing two recurring tasks every week. One: pulling marketing-worthy quotes, client wins, and culture moments out of Fireflies meeting transcripts. Two: replying to customer emails asking to pause service, while also notifying the bilingual operations team in Spanish so they could handle the schedule change. Both tasks were predictable, repetitive, and getting deprioritized whenever the week got busy.

What I built

1. Meeting summarizer

   Pulls Fireflies transcripts and produces a marketing-ready extraction (client wins, testimonial-grade quotes, culture moments) as both markdown for human review and JSON for downstream automation. The marketing team gets a usable artifact 5 minutes after a meeting ends.

2. Service-pause handler

   When a customer emails asking to pause service, drafts a reply on the existing Gmail thread (CC'd to management) and a separate Spanish-language notification to the operations Slack channel so the bilingual ops team can handle the schedule change.

3. Human-in-the-loop on every customer artifact

   Drafts ship to a human for approval before anything sends. The AI handles the 90% of the work that's predictable. The human handles the 10% of the work that needs judgment.

4. Reusable pattern

   Both workflows share the same skeleton: trigger (transcript or email) to LLM extraction to draft generation to human review to send. The agency now uses this pattern for any future "customer wrote in about X" automation.

What it ships

Stack

Claude CodePythonFireflies APIGmail APISlack APIAnthropic API