Aria
Support ticket volume down 35% in 60 days.
A RAG-powered support assistant trained on the client's documentation, integrated directly into their existing product.
Navigate to Settings → Integrations and click "Connect Stripe" to link your account...
A B2B SaaS company with 800 customers and a three-person support team was spending 60% of their support hours answering questions that were already answered in their documentation — pricing questions, feature how-tos, and integration guides.
They had tried two off-the-shelf AI chat products. Both returned inaccurate answers frequently enough that customers stopped trusting them, which increased ticket volume rather than reducing it. They needed something trained specifically on their documentation and capable of citing the source of each answer.
The work
RAG pipeline on proprietary documentation
We built a retrieval-augmented generation pipeline that chunks, embeds, and indexes the client's entire documentation into Pinecone. At query time, the four most semantically relevant chunks are retrieved and passed as context to GPT-4. Every answer includes a direct citation link — removing the trust problem that caused the previous tools to fail.
Confidence scoring and human escalation
Responses below a calibrated confidence threshold are flagged and routed to the human support queue rather than delivered to the user. This maintains accuracy at the cost of some volume reduction — the client preferred a smaller improvement in tickets over an inaccurate answer.
Embedded directly in the product UI
Aria surfaces as a slide-over panel within the existing Next.js app rather than a separate chat widget. It has access to the authenticated user's account context — plan type, active integrations, recent activity — which allows answers to be personalised rather than generic.
The result
Ticket volume dropped 35% in the first 60 days. The support team's average response time on remaining tickets fell from 4 hours to 90 minutes — because the tickets that reached them were now the genuinely complex ones that needed human judgment. Customer satisfaction scores rose from 87% to 94%.
Related projects
Ready to build something like this?
Tell us what you are building. You will get a direct, honest assessment on a 30-minute call — no sales pitch.