AI Support Console cuts support email ~40% in four months.
Athleticknit is a sporting apparel manufacturer serving B2B and B2B2C accounts across Canada and the United States. Custom team uniforms, performance wear, and branded merchandise, sold through resellers and direct corporate accounts. Named public reference — Daniel Sliwin has reviewed FYB publicly on Clutch.
THE PROBLEM
The support team was drowning in repetitive inquiries. Four patterns dominated the queue:
- Daily emails asking the same questions about product lines and ordering procedures.
- Customers unable to find local resellers without calling support first.
- Constant "Where is my order?" requests tying up staff time.
- Order misplacements caused by customers not understanding the B2B ordering process.
Volume was climbing; headcount couldn't. A basic chatbot wouldn't solve it because the biggest queue drivers — order-status and reseller-lookup — needed to actually talk to real systems.
OUR APPROACH
We designed and built a multi-agent AI Support Console — not a chatbot, a system. A supervisor agent routes each incoming message to one of four specialised sub-agents, each with a specific job and access to only the systems that job needs.
The FAQ Agent handles product knowledge and ordering guidance via RAG on the client's knowledge base. The Reseller Agent handles geolocation-aware reseller lookup with zone filtering across the US and Canada. The Order Agent connects directly to the client's OMS via a custom middleware layer and can look up real order status, tracking, and fulfillment timelines. The Handoff Agent kicks in when a case needs a human — it summarises the conversation, captures the context, and routes to support with the ticket already framed.
One customer-facing surface. Four specialised backends. Continuous improvement from conversation analytics.
KEY TECHNICAL DECISIONS
- Multi-agent architecture on LangGraph. Supervisor + 4 specialised agents. Modular by design — new capabilities added as new agents without touching the routing layer.
- Hybrid search (semantic + BM25) on the knowledge base. Semantic alone missed exact-match product codes; BM25 alone missed rephrased questions. Hybrid caught both.
- Function calling with token-based auth for OMS integration. PII stays inside the client's network; the model only sees what it needs to answer the question.
- Proactive deep-linking to static content. Unexpected side benefit: customers arrive at product spec and process pages they wouldn't have found via nav — end up more educated buyers.
- Embeddable widget + admin console + analytics dashboard. The client can update the knowledge base, monitor conversations, and export usage stats without us in the loop.
RESULTS
- ~40% reduction in support email volume on order-related queries (measured over first 4 months post-deployment).
- Increased traffic to static information pages via the chatbot's proactive deep-linking.
- Fewer order misplacements from customers now walked through the B2B process step-by-step.
- Higher fulfillment accuracy as a downstream effect.
Got a support queue bigger than your headcount?
— A few seconds to fill in a form and we'll do the rest!
Book a 45-minute discovery call. Bring your ticket volume and top ticket types. We'll tell you honestly which parts an AI Support Console can deflect and what an honest pilot would look like.