A $1,600/month support team, replaced by one agent

The problem

An online education business that has served more than 10,000 clients produces a constant stream of the same work: onboarding questions, "where do I find this", payment and access problems, homework that needs feedback. Two curators handled it for a combined $1,600 a month. They were good at it. They were also asleep at 11pm on a Sunday, which is exactly when people study.

The salary was not the real cost. The queue was. Answers took anywhere from one to eight hours depending on when a question landed, students lost momentum while they waited, and every decision the curators could not make on their own turned into an escalation to the founder.

What I built

A small set of agents that own the student side of the business end to end, rather than one chatbot that "helps with support":

• Front-line agent answers using the actual course material and the student's own progress, not a generic FAQ.
• Homework agent reviews submitted work and gives structured feedback against the criteria of that specific lesson.
• Access agent verifies the payment, grants the right product for the right duration, sends the welcome email, and handles the "I paid but nothing happened" case without a human.
• Escalation rules decide what a person must see, and that list is deliberately short and strict.

What it refuses to do

This part matters more than the feature list. The agent hands over, every single time:

• Refund requests
• Complaints
• Anything about money that looks unusual
• Anything it cannot verify against real data

Results

• Both curator positions closed permanently, removing $1,600 a month from the cost line.
• Response time went from 1–8 hours to seconds, at any hour.
• 40 to 60 conversations a month handled without a human touching them.
• Running continuously since February 2026.
• The founder stopped being the escalation point for routine questions.

What I'd tell you before you build one

If your support queue is mostly questions that have a correct, verifiable answer somewhere in your own material, this works and it works fast. If your queue is mostly judgement calls, negotiation or emotional situations, an agent will make things worse and you should not build it.

The honest test: take last week's conversations and mark each one as "a correct answer exists" or "someone had to decide". If the first group is under half, don't automate yet. Fix the process first.