AI customer-service assistant for an express delivery company

An express delivery company handling more than 900,000 parcels a day receives around 45,000 support requests daily through its contact centre, chat and social media. Most questions are repetitive: where is my parcel, when will it arrive, can I change the address. The company wanted to test an AI assistant that could resolve these requests on its own, freeing agents to focus on complex cases. Zentrova is now running the pilot with 200,000 customers.

The challenge

During peak shopping seasons, contact centre waiting times sometimes exceeded 15 minutes and up to 30% of callers hung up. The previous chatbot relied on rigid scripts that recognised only a few dozen phrasings and constantly handed customers over to agents. Any new solution had to guarantee that it would never give wrong answers about fees, compensation policy or a recipient's personal details. It also had to cope with the informal, abbreviated Vietnamese customers actually type, including regional expressions and missing diacritics.

Our solution

• Large language model with retrieval-augmented generation (RAG): the assistant answers only from approved policies, price lists and procedures, citing its internal sources.
• Direct integration with operational systems via APIs to track shipments, reschedule deliveries and open claims.
• Identity verification before disclosure, with personal data masked in conversation logs.
• Safety guardrails that detect out-of-scope questions, frustrated customers or compensation requests and hand them straight to an agent, together with a conversation summary.
• Automated evaluation suite of more than 4,000 real customer questions, re-run whenever the model or documentation changes.
• Controlled pilot on web chat and the mobile app, measured against a control group of customers.

Results

After three months of the pilot, early indicators show:

• 58% of requests fully resolved by the assistant.
• First response time under five seconds.
• Answer accuracy of 96% in sampled quality reviews.
• Post-conversation satisfaction of 4.5/5.
• Agent handling time for escalated cases down 40%, thanks to ready-made summaries.

The next step is to bring the assistant to the voice channel of the contact centre and open it to all customers. Before that, the team will extend the evaluation suite to voice transcripts and agree clear service thresholds with the client that must be met before each expansion.

We want AI to do the repetitive work well, and people to do the work that needs empathy. This pilot shows that balance is achievable. — Head of Customer Experience, client company