All work
AI · Automation

Vantage — Support AI Assistant

RAG-powered assistant resolving 41% of support tickets automatically, integrated with their existing helpdesk.

LLMsRAGPython
Va
Timeline
4 months, live in production
Team
2 AI/backend · 1 frontend · 1 support lead (client)
Volume
~2,850 tickets per quarter
Integration
Existing helpdesk, no migration
01The challenge

Vantage's support team was not short of tooling. They were short of hours. The same questions arrived every day, and every one of them was answered by a person from scratch.

Sector
Logistics
Region
Europe
Engagement
AI assistant + helpdesk integration

A quarter's worth of tickets — around 2,850 — came in through email and chat, and roughly 38% of them were billing questions with a knowable answer. Invoice queries, proration, why a surcharge appeared. Answering them consumed the team's capacity, which meant genuinely complex logistics problems sat in the queue behind them.

The support lead had a firm and correct objection to AI in support: a confidently wrong answer about someone's invoice is worse than a slow correct one. Any assistant had to be able to say it did not know, and a human had to be able to see exactly what it based an answer on.

There was also a hard constraint on scope. They had years of history and workflow in their existing helpdesk and were not migrating off it. Whatever we built had to work inside it.

02What we built

How the work was actually sequenced

  1. Retrieval over their own documentation, not a general model

    Answers are grounded in Vantage's help centre, billing policies and resolved-ticket history, chunked and embedded so the assistant retrieves the relevant passages before it writes anything. Every reply carries the sources it used, which is what made the support lead willing to turn it on.

  2. An explicit 'I don't know' path

    When retrieval confidence is low, or a question touches anything the assistant is not allowed to answer — refunds, account closures, disputes — it does not improvise. It hands off to a human with the conversation and the retrieved context attached. Escalation is a designed feature, not a failure state.

  3. Inside the existing helpdesk, not beside it

    The assistant works as an agent in their current tool. Drafts appear for agent review; only categories that passed a review period auto-send. No data migration, no second inbox, no retraining the team on new software.

  4. Shadow mode before it ever spoke to a customer

    For the first six weeks it answered every ticket silently and the support lead graded the drafts. That produced a concrete list of what it was reliably good at, which became the auto-resolve allowlist. Automation was earned category by category rather than switched on and hoped for.

03Results
41%Tickets resolved automaticallyallowlisted categories only
38%Of volume was billingthe first category automated
0Helpdesk migrationsbuilt into their existing tool
6 weeksShadow mode before go-liveevery draft human-graded

Their AI assistant now handles 40% of our support volume. The RAG setup is genuinely accurate, and they were transparent about trade-offs the whole way.

Marcus Bell
Head of CX, Vantage Logistics
04Stack

AI

  • LLM APIs
  • Retrieval-augmented generation
  • Embeddings
  • Evaluation harness

Backend

  • Python
  • FastAPI
  • PostgreSQL
  • pgvector
  • Celery

Integration

  • Helpdesk API
  • Webhooks
  • Audit logging
Next case study

Helios — Energy Analytics Platform

Web · Data

Got something like this in mind?

Tell us what you're building. We'll come back within one business day with a clear plan and an honest estimate.