AI Servicing Intelligence

Clusters complaints, ranks themes by opportunity, and exports a PRD draft instead of a dashboard.

The problem

How many PMs have been handed an Excel file and told "the insights are all in there"? You open it. Five hundred rows. Customer-agent conversations, one per row, each one a wall of text. You scroll for ten minutes, read maybe fifteen of them, and close it.

Or you're sitting in a QBR and someone puts up a slide with a handful of angry reviews on it and calls it customer feedback. You've been there. And the frustrating part isn't even that they're wrong — it's that nobody has anything better. Roadmaps get built on whoever escalated loudest last quarter, or whoever had a strong opinion in the room. The feedback sitting inside thousands of support conversations? It just sits there.

That's what this project is about. I work in payments. I've been that PM — sitting on a pile of servicing data that's technically full of product problems but practically useless as a planning input.

What would you actually want?

So what would you ideally want instead? A way to make sense of that CSV quickly — not by spending a week reading it yourself, and not by raising a ticket and waiting seven business days for an analyst to come back with buckets. You'd want those categories fast. You'd want to see what's actually in them. And then, for the themes that matter, you'd want a way to turn them into something high-impact — a story, a PRD draft, something you could put in front of your team, talk through, and get a decision out of. Not a data dump. An artifact.

So, what is ASI — AI Servicing Intelligence?

Upload a CSV of customer-agent conversations, configure what you're looking for, and get back a ranked list of complaint themes, the evidence behind each one, and artifacts ready for a planning session — PRD briefs, Jira stories. Not a chatbot. Not a dashboard. A workflow that ends with something you can hand to engineering.

How it works

Each conversation gets parsed into structured pieces: the core issue, what the customer asked for, how it resolved, and any flags worth noting — regulatory language, credit concern, fraud mention. The raw text isn't used again after this step. Raw conversations are full of greetings, hold music references, and scripted agent lines that have nothing to do with the customer's problem. Parsing first gives the clustering step something worth working with.

Those pieces get embedded and grouped by meaning. Because clustering runs on the extracted issue — not raw text — "couldn't cancel autopay" and "autopay charged me again after I cancelled" land as separate themes. Which is the right answer.

Then there's a review step before anything gets exported. You see representative conversations from each theme: the evidence, sentiment breakdown, risk signals. AI grouping isn't perfect — examples drift, some themes are too broad. This step is there so you're reviewing evidence before you commit to it, not after. It takes a couple of minutes, and it keeps the output trustworthy.

After you sign off on a theme, an insight summary gets generated: what customers are experiencing, the supporting data, a root-cause hypothesis, and what you'd want to validate before building. Then export — a PRD brief in Markdown or Jira stories in JSON, with Gherkin acceptance criteria and technical context included.

One thing worth knowing about severity

It's not just volume. Two hundred complaints about a minor UI annoyance ranks below fifteen conversations where customers mentioned filing a complaint with a regulator. The severity score weighs six things: complaint volume, negative sentiment, financial and regulatory risk signals, unresolved rate, escalation rate, and low helpfulness scores. Each piece is visible so you can understand — and push back on — the ranking.

Where it stands

This is also something I keep working on and refining — which, honestly, is just how AI products go. The model improves, the pipeline evolves, and there's always a next iteration worth building.