Glamnetic didn't have a broken support org to fix. They had a 4.8/5 CSAT, strong handle times, and what Gorgias CEO Romain Lapeyre calls a top-5% customer experience. This made automating harder, not easier. Every conversation Gina, their AI agent, took over had to land five stars, because a single four-star reply moves an average that good. In roughly ten weeks, the team went from a 25% automation rate to 55%, with 60% in sight, by wiring the AI into their actual stack, running a daily AI-on-AI improvement loop, and killing failed experiments inside 24 hours. This episode is the walk-through: the skills, the nested actions, the QA process, and the one thing that still doesn't work.

Key takeaways

  • Automation went from 25% to 55% in about ten weeks, with a 4.8/5 CSAT held flat. The target is 60%+. No single lever did it: Skills, guidances, and actions had to move together.
  • Connect the agent to the stack, then nest the actions. Glamnetic wired Gina into Shopify, ShipMonk, Inveterate, and Extend. A reshipment isn't one action: check inventory, pick the right warehouse, confirm the missing item, then ship.
  • Put the improvement loop on a daily cadence. Gaia audits the last 24 hours of tickets and posts suggestions to Slack each morning. On one call the team ran 13 improvement loops in 25 minutes: work that normally takes a week.

The hard part wasn't fixing support. It was automating something already good.

Most automation stories start with a support team that can’t keep up. Glamnetic's didn't. CSAT sat at 4.8 out of 5, handle times were strong across chat and email, and the team's stats were, in Mia's words, already amazing. The problem was the opposite of broken: how do you hand work to an AI agent when the bar is a five-star reply every time?

Romain puts the constraint plainly: at a top-5% experience, even a four-star conversation drags the average. So the goal was never raw deflection. It was moving low-value, repetitive work off the team without spending a single CSAT point to do it.

The 10 week automation challenge: 25% to 55% automation

A few months before recording, Glamnetic was automating about 25% of conversations. At the time of the episode, they were at 55%, pacing toward a 60%+ goal Jackie calls a huge win.

Jackie explains that no single lever led to this. Skills, guidances, and actions had to move together, each QA'd for whether it automated cleanly or handed over correctly. As Jackie puts it, “Sticking to only one thing won't get you where you need to get.”

Connect AI Agent to your entire stack so it can act, not just reply

Romain's advice for an operator stuck at 30% is two-part. First, cover the common intents with the built-in Skills: returns, subscriptions, order issues. Second, connect the AI agent to the rest of the stack. Glamnetic wired Gina into Shopify, ShipMonk, Inveterate, and Extend, which upgrades the agent from identifying a problem to actually fixing it.

This better equips you to handle the edge cases. What do you tell a customer whose order has been stuck at the Canadian border for five days? Jackie walks us through how Glamnetic handles this with its shipping address edit skill: required fields, boundaries, merged-order detection, and a linked action that edits the order in Shopify only after the customer confirms.

Gaia: an AI that audits the AI, every morning

The loop that moved the number fastest is an AI reviewing an AI. Gaia audits the previous 24 hours: Skills, flows, handovers, human-handled tickets and posts findings to Slack each morning. Mia and Jackie review, accept or reject, and the changes go in.

Romain counted 13 improvement loops on a single 25-minute call, against the week that pace normally takes. It also makes tedious work worth doing: a long-tail action that would take two hours to build by hand and only serve five tickets never got built before. When an AI drafts it for near-free and a human just approves, it does. Glamnetic now runs close to 20 actions.

Experiment fast, cut fast, and be honest about what fails

Glamnetic's willingness to test is exactly what you'd expect a 4.85 CSAT to prevent. Their answer is speed: run the experiment, learn inside 24 hours, and if it failed, cut it. This keeps damage to 24 hours and wins compound. Mia frames the license to try as earned: a consistently high CSAT buys the leeway to drop a few points and recover.

QA is the job now

Jackie's day starts in the ticket queue. Views separate chat from email and good from bad; she QAs the last 24 hours, flags what's off, and leaves written feedback that routes back to Gorgias. The process repeats for guidances, then actions, checking for errors and whether each fired successfully. Tedious, she says, and worth it, because QA is the only way to know whether any of it is actually working.

The rating loop is the part worth stealing: rate the reply, write the reasoning, then thumbs up or down each source the agent used so good knowledge gets prioritized and bad recommendations don't resurface. Mia nearly hired a dedicated AI QA person before Gaia made the role unnecessary.

Actions that paid for themselves, and where this goes next

When asked what action to turn on first, Jackie says cancellations then reshipments. These branch by scenario: damaged, defective, lost in transit, partial or full. The one she keeps coming back to is editing shipping addresses during Black Friday / Cyber Monday, when address-change requests spiked, and Gina absorbed them without glitching.