Trending:
AI & Machine Learning

Phone repair shop automated 90% of customer service with Claude agents, Airtable integration

Santifer iRepair built an AI agent system handling WhatsApp, calls, and bookings by routing requests to specialized sub-agents querying Airtable directly. The approach: tool calling over prompt stuffing, early escalation over degradation, existing databases as agent memory.

The Implementation

A 16-year-old phone repair shop automated customer service by building an orchestration layer, not a chatbot. The system routes incoming WhatsApp and voice requests to specialized sub-agents: one handles appointments, another discounts, a third manages orders. Each queries Airtable databases directly for inventory, customer history, and available slots.

The architecture matters. Tool calling replaced prompt engineering. Instead of cramming business logic into system prompts, each sub-agent calls explicit functions: check_available_slots(), create_booking(), calculate_discount(). The model reasons about intent. The tools execute actions. When confidence drops, the system escalates to humans with full conversation context.

What Worked

The stack: Claude API for reasoning, n8n for orchestration, WATI for WhatsApp Business API, Airtable as source of truth. No separate AI database. No sync issues. Appointment booking alone represented 40% of inquiries, the obvious automation target.

The result: 90% autonomous handling, sub-2-minute human response times for escalations, and a sellable asset when the business changed hands in early 2025.

The Pattern

This maps to broader enterprise implementations. WhatsApp Business API integrations with Claude are handling high-volume customer service across APAC, particularly for SMBs unable to staff 24/7 support. The technical constraint is rate limits: Claude API tiers matter for production chatbots processing thousands of daily messages.

The lesson isn't about repair shops. It's about router architectures, tool calling over prompt complexity, and treating existing databases as agent memory. The 10% requiring human escalation got better service than before because context transferred.

What to Watch

Structured outputs from Claude API improve reliability for business automation. Token limits and rate limit tiers dictate which use cases work at scale. Role-based access control becomes relevant for enterprise WhatsApp integrations touching customer data across departments.

The question for CTOs: which repetitive workflow is your appointment booking?