Starting point
The customer operated a conventional customer service team with eight full-time positions. The problems were:
- A high proportion of recurring standard questions about delivery times, terms and product use
- Response times of more than 24 hours during peak periods
- Knowledge silos, with valuable expertise held by only a few experienced employees
- Onboarding new service staff took 6–8 weeks
Solution
A RAG chatbot connected to the central knowledge base, including product manuals, FAQs, historical tickets and terms and conditions:
- Foundation model: GPT-4o via Azure OpenAI in an EU region
- Vector store: pgvector + BM25 hybrid search
- A source citation was mandatory for every answer
- Confidence below 70% → automatic escalation to a person
- Embedded in the web front end, Microsoft Teams and the telephone hotline using voice
Approach
- Prepared the knowledge base data in weeks 1–3—by far the most time-consuming step
- Built the bot and evaluated it against 200 test questions in weeks 4–5
- Soft-launched it for internal employee use in weeks 6–7
- Launched it externally in week 8
- Ran an optimisation iteration after every 4 weeks in production
Results after 6 months
~60%of standard enquiries resolved by the bot
< 5 minaverage response time, compared with 24 hours previously
+8 NPSpoints in customer satisfaction
~3 weeksshorter onboarding for new service staff, with the bot acting as a knowledge resource
Critical success factors
- Data quality beats model power. Three weeks of data preparation delivered more value than any model upgrade.
- Source citations are non-negotiable. Without them, the internal compliance team would never have approved the system.
- Escalation must be seamless. If the bot makes people feel dismissed, trust disappears immediately.
- Gain employee buy-in first. Service staff were introduced to the bot as a tool, not a threat.
Do you handle a similar volume of enquiries?
Book a 30-minute consultation about your customer service setup. We will assess whether a RAG bot is justified—or whether a simple automated FAQ responder would be enough.