A mid-size UK e-commerce business with 80 staff and €12M annual revenue was facing a worsening support backlog heading into peak season. Ticket volume had grown 40% year-on-year, with 12 agents handling ~2,100 tickets per week and an average first-response time of 22 hours. During the previous Black Friday, response times exceeded 60 hours and negatively impacted their Trustpilot score. Hiring four more agents at €180k/year still wouldn’t have solved peak load, so they needed a different approach.
Diagnosis
Over two weeks, we analysed their support data and found:
- 63% of tickets were repetitive, factual queries about:
- Order status
- Return eligibility
- Product availability
- Delivery estimates
These had clear, lookup-based answers and didn’t require human judgment.
- Phone was underused but highly effective:
- Only 12% of contacts were by phone
- Phone calls had much higher resolution and satisfaction than email
- Callers got answers in minutes vs. hours for email
- After-hours demand was high:
- 35% of tickets arrived between 6pm and 9am
- These waited 10–14 hours before anyone saw them
Conclusion: a 24/7 AI voice agent that could handle high-volume, factual queries and redirect email-style issues into a faster, higher-satisfaction phone experience would address the root cause more effectively than adding headcount.
Solution Architecture
We built an AI voice support system on Vapi with a bring-your-own-LLM setup:
- Vapi for voice agent infra, telephony, STT/TTS
- Claude 3.5 Haiku for standard queries
- Claude 3.5 Sonnet for complex cases and escalations
- Shopify API for:
- Real-time order status
- Return eligibility checks
- Stock availability
- Zendesk API to create human tickets with:
- Call summary
- Caller details
- Context of the issue
- Twilio for SMS follow-up (e.g. return references, booking confirmations, tracking details)
- n8n as the orchestration and business logic layer between Vapi and downstream APIs
Capabilities:
- Handle calls in English and Polish
- Look up orders by order number or email
- Determine return eligibility based on policy, including edge cases:
- Sale items
- Bundles
- Store-credit purchases
- Final-sale items
- Opened packaging rules
- Provide real-time stock availability
- Calculate delivery estimates via carrier API
- Escalate to humans for:
- Complaints
- Damaged goods
- Payment disputes
- Any account-level judgment calls
On escalation, the agent transferred the call and pushed a structured summary into Zendesk.
Rollout Strategy
We used a staged rollout to validate accuracy and customer experience:
- Days 1–10 (Pilot):
- AI handled ~20% of inbound calls
- Human team operated as usual
- Goal: validate accuracy and containment before scaling
- Week 2 (Controlled Scale-Up):
- 60% of calls routed to AI, 40% to humans as a control group
- Order lookup accuracy: 99.3% (3 incorrect out of 450 calls, all due to order merges)
- Resolution rate on the four core query types: 87%
- Week 3 (Full Rollout):
- AI became the primary front line
- Human agents focused on escalations and email
Results After 90 Days
Key outcomes:
- Email ticket volume: ↓ 75%
- Many customers who previously emailed now called and got instant answers
- Response time for AI-handled queries:
- From 22 hours → under 2 minutes
- Containment rate:
- 87% of calls fully resolved by AI without human escalation
- Human workload:
- 35% reduction in agent workload
- Evening/weekend staff redeployed to business hours
- Trustpilot score:
- From 3.9 → 4.4 over 90 days
- Multiple factors contributed, but improved support experience was the main driver
- Cost per interaction:
- From €8.40 → €1.20 for AI-handled contacts
During Black Friday, call volume spiked 340% above baseline. The AI absorbed the surge without capacity issues, avoiding the response-time collapse seen in previous years.
Key Learnings
- Return policy logic is deceptively complex
- Edge cases around sale items, bundles, store credit, final sale, and opened packaging required significant rule design and testing.
- This was the most time-consuming part of the build.
- Polish language support exceeded expectations
- Using Deepgram STT plus Claude’s multilingual reasoning delivered robust performance in both English and Polish.
- SMS follow-up reduced repeat contacts
- Sending an SMS with return references or tracking numbers led to fewer follow-up queries.
- A small addition with a clear impact on downstream ticket volume.
Cost and ROI
- Build cost: £18,500
- Roughly equivalent to four months of one additional support agent
- Ongoing infra cost: ~£650/month at current call volume
- Payback period: < 3 months
If you’re facing similar support volume and response-time issues, you can book a free scoping call. We’ll assess whether voice AI is the right lever for your support mix and outline realistic timelines and costs.