What was the result?
self-reportedMarker.io CEO Gary Gaspar says replacing Intercom's Fin with a custom AI support agent built with Claude Code "saved $9,000 a year", taking the cost "from $800/mo to $50/mo." He says the savings were a bonus: "We simply wanted a support agent that works."
Gary Gaspar is CEO and co-founder of Marker.io, a software company that makes a bug reporting tool. Their off the shelf support bot kept making things up: features they didn't have, support articles that didn't exist, prices in the wrong currency, and no idea which bugs were already fixed. So his team spent 3 months building their own support agent with Claude Code. It reads screenshots users send, tells users when a bug they report is already fixed, checks the public roadmap when someone asks for a feature, and is trained on their docs so it sounds like them.
Who did this?
How do I try it?
- List the 20 questions your support inbox gets most and the wrong answers your current bot gives.
- Collect your help docs, pricing page, public roadmap and a list of recently fixed bugs in one folder.
- Build a support agent (Claude Code or a similar tool) that answers only from those files and says so when it doesn't know.
- Have it check the fixed bugs list and the roadmap before answering bug reports and feature requests.
- Run it beside your current bot for a few weeks, read every answer, then switch over.
Editorial adaptation, not Marker.io's setup. You answer support emails for [product]. Use only the files in this folder: help docs, pricing, roadmap and fixed bugs list. If a user reports a bug that is on the fixed list, tell them it is fixed and how to get the update. If they ask for a feature, check the roadmap and say whether it is planned. Never invent a feature, a price or a link. If the answer is not in the files, say a person will reply and flag the ticket.
Make it run every week
A play works once. Turn it into a routine in three small steps:
- Run it once and check the result yourself.
- Give your AI a schedule. Tell it when to run this play (every week, or every month) and to message you with what it found before you act.
- Keep a log. One line per run: what was checked, anything flagged, what you did.
What should I watch out for?
Self-reported. It took his team 3 months to build, so this is not a weekend project. Under his post, someone from Intercom questioned the $800 a month figure and said it looked closer to $80, so treat the exact savings as Gary's own number. Running costs for an AI model can change with use. Safer version: have a person review a sample of answers every week and keep an easy way to reach a human.
Did it work for you?
PageMotor keeps this site free.