Ticket and satisfaction analysis for customer service

The leak

The customer-service team had the data: exports of every ticket, and a separate export of customer-satisfaction scores. What it did not have was the analysis. Each month someone rebuilt the same pivots by hand, compared them to last month by eye, and wrote a deck. Which categories were rising, which channels were slowing down, where turnaround times had slipped, how the regulatory complaint channels were trending, which issues drove the lowest satisfaction: all of it was answerable from the files and none of it was answered consistently, because the work of answering it was a day of someone's month.

The constraint

The team is not technical and the data is sensitive, so the tool had to run on their own machines with no server, no account and no upload. It had to accept the exports exactly as the systems produce them, including the quirks. It had to handle the channels that matter to a regulated lender, including the regulator's complaint portal, the government grievance channel and the nodal officer's mailbox, as first-class categories. And it had to produce the deck, not a dashboard someone then screenshots into a deck.

The system

A desktop application, signed and distributed for both Windows and Mac, with its own release process so updates reach the team cleanly. The team opens it, points it at the ticket and satisfaction exports for the months they want, and the application does the rest: cleans and maps the channels, computes volumes, categories, turnaround times and satisfaction by channel and by issue, compares every month to the ones before it, and surfaces what moved.

A language model then writes the findings: which categories went up and why that matters, where turnaround slipped, which issues the lowest-scoring customers raised. The output is a presentation and a PDF report, in the team's own format, produced in minutes from files that used to take a day.

Versioned, built and signed through an automated release pipeline, so a new version is a download rather than a support call.

The number

A month's analysis in minutes instead of a day, every month, in the same shape, with the narrative written. The team runs it whenever it wants a view, not only when the monthly deck is due.

What I would do differently

Add the customer-satisfaction verbatims from the start. The scores told the team where satisfaction was low; the comments would have told them why. Reading those through the same language model was an obvious second step that should have been the first version.