Nobody Had Time to Read the Support Inbox. So a Machine Did.

In 2024 the head of customer service at the company I work for asked a simple question: what are people actually writing to us about?

Nobody knew for sure. The support team had a feel for what came up most, but a feel isn’t a roadmap. So product fixed whatever got loud that week.

The answer was sitting in the support inbox. Two years of conversations. Nobody had time to read them.

So I built something that would.

What it did

I pulled every conversation from the last two years. Thousands of them.

I threw out the automated messages, leaving only what customers wrote.

Then I had an AI read each one and write a few plain sentences: what is this person’s problem? Most of them came out with a short summary.

Then I had it sort those summaries into piles by what they meant, not what words they used. If only a handful of people had written in about the same thing, it didn’t count as a pile.

Last, a report the team could click through, with real tickets under each problem.

Idea to findings in front of the team: about two weeks.

What came out

Six product problems were driving most of the tickets.

The biggest by a mile: getting text messaging set up. A verification step could take two weeks, customers didn’t understand why, and they wrote in. About a quarter of the sorted tickets. The next biggest, lessons not showing up when people expected them to, was about a tenth.

After that: video upload, scheduling and time zones, password and login trouble, and automations not firing as expected.

None of this surprised the support team. What was new was the ranking. When you can see one problem is more than twice the size of the next, it changes what you fix first.

The honest part: about a third of the conversations didn’t fit any pile. One pile that did form was just “not enough information to act on.” That’s a finding too. It says something about the support form.

What happened after

Product and design took it from there. The text-message verification messaging got rewritten, a better setup flow got built, and support got a real process for walking customers through it. Video upload got fixed. The rest went into the roadmap by size.

Support ticket volume is down year over year.

Here’s the limit. I can’t tell you this caused that drop. A lot changed that year. I can’t put a dollar figure on it either. What it produced was a ranked list of problems and a roadmap that got reordered because of it. That was the point.

What it cost

Two weeks of my time and a small AI bill for the summaries. No data team. Nothing you couldn’t run again next quarter.

What I’d do differently

Use a better model for the summaries. I used the cheap one in 2024. A third of conversations with no home tells me some meaning got lost. Better models are cheap now.

Run it every quarter, not once. Once tells you where you were. Every quarter tells you whether the fixes worked.

Keep score from day one. Count per problem before each fix, run it again after. Then I’d have a real before-and-after instead of a year-over-year guess.

What to take from this

If you run a business with a support inbox, the answer to “what’s driving our support volume” is already in there, in the tickets you already paid people to answer. Nobody on your team has time to read two years of them and keep score. A machine does, in about two weeks.