SaaS · 10 weeks to production

Deflecting 62% of support tickets with a grounded AI assistant

A retrieval-grounded AI support assistant that resolves 62% of inbound tickets with answers cited from the client's own documentation.

of tickets resolved without an agent: 62%; of tickets resolved without an agent
median first response time: < 5 s; median first response time
customer satisfaction on AI answers: 4.6/5; customer satisfaction on AI answers

The challenge

Client: B2B software company
Services: AI & Machine Learning, Custom Software
Timeline: 10 weeks to production

A 6,000-article knowledge base and a support team drowning in repetitive tickets, but a previous chatbot attempt had been scrapped after it confidently invented answers and damaged customer trust.

The bar was explicit: the assistant must cite real documentation, admit when it doesn't know, and hand off to humans smoothly. Measured, not vibes.

What we did

The approach, decision by decision

01
Built evaluation before the assistant
Turned 400 historical tickets into a graded test set. Every retrieval and prompt change ran against it, so quality moved by measurement instead of anecdote.
02
Retrieval engineered for citations
Hybrid search over chunked documentation with reranking, strict grounding prompts, and inline citations, where answers link to the exact article section they came from.
03
Designed the handoff, not just the bot
Low-confidence answers escalate to agents with full conversation context and suggested replies, turning the AI into an assist for humans rather than a wall in front of them.

The results

62% of inbound tickets fully deflected within three months of launch.
Median first response time fell from 3.2 hours to under 5 seconds.
Hallucination rate below 1% on the eval suite; every answer carries citations.
Support team redeployed to onboarding and success work, and attrition stopped.

Built with

Claude API
pgvector
Next.js
Python
LangGraph
AWS

Keep reading

More case studies

Fintech · Consumer lending fintech

Cutting loan decisions from 4 days to 7 minutes

A digital lending platform with automated underwriting that took loan decisions from a 4-day manual review to a 7-minute automated flow.

Read case study

Logistics · Freight operations startup

Scaling a logistics SaaS from zero to 50,000 shipments a month

A multi-tenant freight management platform launched in 12 weeks, now processing 50,000+ shipments a month across three markets.

Read case study

Your project could be the next case study

Tell us the metric you need to move. We'll scope the build that moves it: fixed price, weekly demos, results you can publish.

Book my free consultation Send a project inquiry

We reply within 1 business day.