Case study · CX Insights / Knowledge Hub
Turning support tickets into decisions.
Support knowledge tends to scatter across tickets, macros, spreadsheets and people’s heads. I built an operational intelligence layer that turns it into structured, reviewed knowledge — and into the reporting the business runs on.
Can I pause my subscription instead of cancelling?
My parcel says delivered but it isn’t here
Do refills fit the original version?
Can I change my address after ordering?
Why was I charged before my box shipped?
01 · The problem
Everything we knew was somewhere.
In any support operation that has grown over the years, the same questions get asked thousands of times. The knowledge to answer them exists, but it tends to live in four places at once:
- TicketsYears of solved conversations: rich in answers, hard to search.
- MacrosWritten at different times, by different people, for different situations.
- SpreadsheetsUseful tracking that lives apart from the conversations behind it.
- People’s headsThe senior agents who “just know” the answer.
That’s normal for a growing operation, but it makes some simple questions hard to answer. What do customers ask about each product? Why is compensation issued? What changed this week? And any AI put in front of customers is only as good as the knowledge it draws from.
02 · The idea
Treat every ticket as data.
Instead of asking people to write documentation, extract it from the conversations the team is already having — then put a human in charge of what becomes true.
- 01 · sourceTickets
Solved conversations from the helpdesk, across every brand.
- 02 · extractAI extraction
Questions, answers, products, reasons and sentiment pulled into structure.
- 03 · reviewHuman review
Approve, edit, merge or reject. Nothing is trusted until a person says so.
- 04 · reconcileApproved library
Near-duplicates reconcile into families with one canonical answer.
- Agents & macros
- Product questions
- Compensation reporting
- Summaries & voice of customer
03 · What I built
Capture. Curate. Use.
Not a feature list: three moves that turn a pile of tickets into something the business can make decisions with.
AI proposes. A person approves.
Candidates extracted from tickets land in a review queue with their source conversations attached. Reviewers approve, edit, merge or reject.
The queue is where CX judgment lives: the model does the reading, a person decides what the company actually says.
Can I pause my subscription instead of cancelling?
One question, one answer.
Customers ask the same thing a dozen different ways, across brands. Instead of a library that grows forever, near-duplicates reconcile into a family with one canonical, approved answer.
So the library gets sharper over time instead of just bigger.
From anecdotes to numbers.
Once tickets are structured, the same layer answers questions that are otherwise hard to pin down: what customers ask about each product before and after they buy, why refunds and credits are issued, and what changed this week.
Summaries and voice-of-customer reporting go to product and ops on a schedule. Nobody has to build a spreadsheet first.
04 · Decisions
Choices I’d make again.
Most of the value came from a few unglamorous calls, not from the model.
-
01
Review before truth
Extraction is automatic; approval never is. It’s slower on day one, and it’s what makes the library worth trusting.
Humans where it matters -
02
Families over folders
Reconciling phrasings into one canonical answer keeps the library small and correct, instead of large and contradictory.
Measure the experience, not just the speed -
03
Boring infrastructure
A relational database and scheduled workflows the team can inspect. Every step is visible, re-runnable and easy to switch off.
Boring beats clever
05 · Outcome
From fragments to one layer.
What this project took
- CX domain knowledge
- Product thinking
- Workflow design
- AI & LLMs
- Data modelling
- UX
- QA
- Software engineering
Contact
Sitting on a pile of tickets?
Your support data probably knows more than your dashboards do. I’m happy to talk through what it could tell you.