Case 01 · SaaS · AI
An AI agent that resolvessupport requests — not just answers them.
A first-line support agent wired into a SaaS company’s internal systems. It looks up real customer data, takes safe actions and hands everything else to people — with the full context.
- Industry
- SaaS · B2B software
- Region
- Europe & North America
- Engagement
- Product Team · Discovery to support
- Services
- AI & ML · Product Engineering
The challenge
A fast-growing SaaS company was answering the same questions every day: order and account status, billing, how-to requests. Support headcount grew with the customer base, and response times grew with it. A generic chatbot had already been tried and dropped: it answered from outdated FAQs and couldn’t see anything inside the product.
What we built
- An LLM agent that understands the request, looks up the customer’s real data and acts through internal APIs — checking an order, updating account details, resending an invoice.
- A knowledge layer (RAG) over help-center articles, internal docs and resolved tickets, with sources attached to every answer.
- Clear limits: actions the agent can take on its own, actions that need the customer’s confirmation, and topics it always hands to a person.
- A handoff to human agents with a summary of the conversation, so customers never have to repeat themselves.
- An evaluation suite and dashboards for answer quality, cost per conversation and escalations.
How we did it
We started by analyzing past tickets to find the categories worth automating first. The agent’s tools were built as narrow, typed functions over existing APIs, so it could only do what it was allowed to do.
- An evaluation set built from real, anonymized tickets — every change to prompts or models is tested against it before release
- Rollout by category and channel, starting with a share of traffic and expanding as quality held
- Weekly reviews of failed conversations together with the support team
The outcome
- ✓Routine requests resolved end to end, without waiting in a queue
- ✓The support team focused on complex cases, with context ready at handoff
- ✓Answers grounded in current data and documentation, with a trail for every action
- ✓A measurable quality baseline before every change
Stack
- Python
- FastAPI
- LangGraph
- OpenAI
- Anthropic Claude
- PostgreSQL
- pgvector
- Redis
- React
- Docker
[ Contact ]
Have a complex product in mind?Let’s talk.
> start_project --team neurovista
✓ reply within one business day · NDA on request
or write to {{CONTACT_EMAIL}}