Teaching a pre-LLM system how customers actually talk
I ran the field research behind LogMeIn's issue categories and wrote the ontology that tied how customers describe a problem to how agents actually diagnose it.
A customer who says the screen is blank but nothing crashed has a Mac permission problem, and nothing in that sentence says so.
The problem
LogMeIn had bought Clarabridge to categorize support call transcripts. Its NLP was deterministic and pre-LLM. It couldn’t infer that two different phrasings pointed at the same issue, so every link between them had to be written by hand.
The existing categories didn’t help. Agents tagged each call in a 30-second window after hanging up, from a long list that rarely matched what they’d worked on. 75% of issues landed in three catch-all buckets: Account Set Up, Technical Issue, and Other.
What I did
- Listened to 350 customer calls and grouped callers by role: administrators, billing administrators, event organizers, attendees, and prospects.
- Spent three weeks at the Costa Rica support center leading three researchers. We shadowed agents across support tiers and ran workshops where agents documented every recurring issue, its severity, and how it got resolved.
- Ran card sorts to capture how customers described each issue in their own words, then checked that vocabulary across support groups.
- Wrote the ontology: customer language on one side, agent diagnosis on the other, and the ties between them.
The tie only field research could write
A customer says: “My screen isn’t sharing. It’s just blank, but the software isn’t crashing.” An agent hears a permissions or firewall problem and asks one question first: Windows or Mac? On a Mac, it’s almost always a recent OS update that reset the screen-recording permission.
Nothing in the customer’s sentence says permission, update, or firewall. Without someone writing that tie, the category couldn’t exist and the issue couldn’t be counted.
What else we found
Calls averaged 12+ minutes against an expected 5 to 10. Agents at the partner site couldn’t download files or approve simple refunds, a required morning manager meeting left refund approvals waiting, and constant security checks built distrust. We recommended renegotiating the contract or bringing care back in-house.
Outcomes
- The ontology became what Clarabridge categorized against, and the issue maps fed the content plan for the product chatbots.
- Customer care began moving back in-house, and agents got the same permissions as supervisors for refunds and similar calls.
- It became the foundation of the Voice of Customer program.