IT services · New Zealand
AI quality checks on every support call
An IT support desk gets every call transcribed, summarized and scored, in one dashboard, plus a daily email of calls that went badly.
52%
lower hosting cost after moving to external APIs
Daily
email of negative-sentiment calls
Every call
searchable, with sources shown in chat
The problem
The service desk recorded every call, but nobody had time to listen to them. Managers checked quality by sampling a few calls by hand, so problems surfaced late, usually as complaints.
What I built
A pipeline that picks up each call recording, transcribes it with speaker labels, then has an AI model summarize and score it for a quality audit. Results land in one dashboard with search. A daily email lists calls with negative sentiment, and a weekly CSV compares the numbers.
Every call also goes into a knowledge base. Staff can chat with past calls to find how a problem was solved before, and each answer names the calls it came from, so nobody has to trust a guess.
The first version ran its own speech and embedding models on a large server. I moved it to external APIs on a smaller one, which cut hosting cost by 52%.
The result
Managers read a short daily list instead of sampling calls. Coaching starts from real calls, the same day.