SYSTEM CASE STUDY / WAYFIX / 2024–2025
WayFix
Support Copilot
The right procedure, inside the agent’s workflow.
We built a RAG-based knowledge assistant inside WayFix’s proprietary CRM, integrated with Salesforce. Agents could search fragmented documentation in everyday language, read a concise answer with source references and approve suggested actions without leaving their workspace.
From a customer’s words to a supported next step.
UNDERSTAND THE QUESTION
“The light is flashing red.”
An agent enters the issue in natural language.
↓ semantic retrieval + BM25
FIND THE EVIDENCE
Relevant procedure & source links
The response is grounded in retrieved documentation.
↓ agent reviews the proposed step
SUPPORT THE ACTION
An answer, an action or an escalation
The agent remains responsible for executing suggested commands.
Illustrative interaction flow, not a screenshot of the production interface.
THE OPERATIONAL CHALLENGE
The answer existed. Finding it took too long.
WayFix handled more than 15,000 technical-support interactions a day. Resolution procedures were spread across legacy systems, hardware manuals, Confluence pages and intranet notes.
FRAGMENTED KNOWLEDGE
Multiple sources. One live call.
Agents spent an average of 3.5 minutes navigating documentation to find the right procedure. That search time became part of a longer customer interaction.
A LANGUAGE MISMATCH
Symptoms are not error codes.
Customers described what they saw, while documentation used product terminology and technical identifiers. Keyword-only search often missed the connection.
TWO STAGES OF THE WAYFIX PLATFORM
From rules and exact matches to contextual assistance.
The 2024–2025 Copilot was a later implementation. Its RAG and LLM capabilities are separate from the original WayFix work of 2011–2013.
WAYFIX / 2011–2013
Rules and exact-match search
The earlier platform used decision trees, relational SQL data and exact-match keyword search. Agents had to navigate predefined paths or find the terminology used in the documentation. Changes to the knowledge workflow required maintaining those rules.
WAYFIX COPILOT / 2024–2025
Retrieval, synthesis and suggested actions
The later implementation combined Pinecone retrieval and Claude 3.5 Sonnet through Amazon Bedrock. It connected questions to relevant documentation, synthesised cited guidance and exposed API actions for the agent to approve.
WHAT WE BUILT
A connected knowledge and action workflow.
The implementation covered knowledge ingestion, structure-aware chunking, embeddings and indexing, hybrid retrieval, LLM response generation, CRM integration and suggested actions through WayFix’s API.
Knowledge pipeline
01 / SYNC
Refresh source material every 24 hours
Automated extraction from Confluence, SharePoint and PDF repositories.
↓
02 / STRUCTURE
Keep procedures together
Document headings, lists and steps informed chunk boundaries, preserving procedural context.
↓
03 / INDEX
Prepare meaning-based retrieval
Embeddings indexed in Pinecone made related terminology discoverable.
During the support interaction
04 / RETRIEVE
Combine semantic search with BM25
The agent’s question retrieved relevant passages through meaning and lexical matches.
↓
05 / SYNTHESISE
Return a concise, cited answer
Claude 3.5 Sonnet, accessed through Amazon Bedrock, used retrieved passages to propose an answer linked to the original sources.
↓
06 / ACT
Keep execution with the agent
Eligible procedures exposed an agent-operated action connected to WayFix’s API.
Knowledge path: source documents → structured chunks → searchable index → retrieved context → cited assistance → agent decision.
Technology: WayFix proprietary CRM · Salesforce integration · Claude 3.5 Sonnet via Amazon Bedrock · Pinecone · semantic retrieval + BM25 · Confluence, SharePoint and PDF ingestion.
SOURCE GROUNDING & HUMAN CONTROL
Make the basis of an answer inspectable.
EVIDENCE
Citations beside the answer
Generated guidance included direct references to the original documentation so an agent could inspect the source before following the procedure.
ESCALATION
A defined no-answer path
The system was instructed to answer from retrieved material. When a supported procedure was unavailable, it returned a knowledge-base fallback for Level 2 escalation.
ACTION
Suggested, then approved
A procedure could surface a control for an API action, such as a remote restart. The agent chose whether to execute it from the support workspace.
FEEDBACK LOOP
Turn weak answers into a retrieval improvement signal.
Agents could rate each answer positively or negatively. That feedback fed improvements to retrieval relevance and helped identify answers that needed attention.
PROJECT RESULTS
Less searching. More resolution.
Before-and-after results from the WayFix implementation.
Documentation search time
94.3% shorter
Before · 210 seconds (3.5 minutes)
After · 12 seconds
Time per search; both bars use the same zero-based scale.
Average handling time
27.7% lower
Before · 11.2 minutes
After · 8.1 minutes
AHT per interaction; both bars use the same zero-based scale.
First contact resolution
+14 percentage points
Before · 68%
After · 82%
Share resolved at first contact; both bars use a 0–100% scale.
Agent onboarding
62.5% shorter
Before · 4 weeks
After · 1.5 weeks
Onboarding duration; both bars use the same zero-based scale.
Changes are calculated from the before-and-after values shown. Percentage-point change is used for first contact resolution.
APPLIED AI IN A WORKING SYSTEM
Connect knowledge, generation and execution.
Retrieval
Bring fragmented source material into the context of the current support problem.
Integration
Deliver assistance inside the CRM and connect suggested steps to an existing API.
Agent judgement
Expose sources, retain action approval and provide an escalation path when knowledge is missing.
LET’S BUILD THE WORKFLOW
Where do your teams lose time looking for answers?
Bring your documentation sources, a typical support question and the workspace your agents use. We can define the retrieval, integration and delivery scope together.