Case study 03
AI Support Copilot
Deflected 41% of support tickets with grounded, cited answers.
Client
Perlin Software
Industry
SaaS
Platform
AI Product
Services
AI Solutions, Web Application Development
Timeline
14 weeks
01 / Challenge
Support volume grew faster than headcount. Earlier chatbot attempts produced confident, wrong answers and were switched off within a month.
02 / Our approach
We built a retrieval system grounded strictly in product documentation and resolved tickets, with citations, refusal behaviour and a measured evaluation set from the first week.
03 / UX / UI
The assistant lives inside the existing help surface rather than as a floating bubble, with visible sources and a one-tap path to a human.
04 / Technology architecture
Python ingestion pipelines keep a pgvector index fresh, model routing sends simple queries to cheaper models, and every response is logged for evaluation.
05 / Development
A labelled evaluation set of 800 real questions gated every release. No prompt or model change shipped without a regression run.
Key features
What We Shipped.
Results
Business Impact.
41%
Tickets deflected
-33%
First response time
94%
Answer accuracy
$0.011
Cost per resolution
Technology stack
Gallery
Inside The Product.
Client testimonial
“It is the first AI feature we have shipped that support actually trusts. The evaluation discipline is why.”
Next project
Clinical Care CompanionWant Results Like These?
Tell us what you're building. We'll help you figure out the smartest way to bring it to life.