RAG-based AI support agent for technical customer inquiries at a scientific instrument manufacturer.
The project focuses on requirements analysis, solution design, and architecture for an AI-assisted support system serving a small manufacturer of spectroscopy instruments used in research, industry, education, and applied science.
Customer support depends heavily on a single domain expert who also manages sales, procurement, marketing, website operations, and product R&D.
A significant share of expert time is spent processing repetitive technical inquiries, searching historical emails and operational records, locating support materials, and preparing responses.
Key challenges:
- high dependency on a single subject matter expert;
- repetitive support requests and long response preparation time;
- knowledge concentrated in individual experience;
- fragmented and incomplete support materials;
- limited scalability and risk of inconsistent responses.
The project designs a Retrieval-Augmented Generation (RAG) support agent that:
- classifies incoming customer inquiries;
- retrieves relevant information from approved knowledge sources;
- searches historical support emails and customer-specific order records;
- drafts grounded customer responses;
- locates drivers, calibration files, manuals, and technical documentation;
- routes complex or low-confidence cases to the domain expert;
- preserves source traceability and human oversight.
Customer Email → Classification → Retrieval → Response Drafting → Human Review / Automated Action
Primary knowledge sources:
- Historical sent emails
- Order register
- Product manuals and technical documentation
- business process analysis;
- AS-IS and TO-BE process modelling;
- support inquiry classification design;
- knowledge base and retrieval requirements;
- RAG workflow design;
- functional and non-functional requirements;
- data model design;
- integration requirements;
- AI governance and escalation rules;
- use cases and acceptance criteria.
- full production implementation;
- enterprise CRM implementation;
- complete knowledge base migration;
- product development and R&D processes;
- sales, procurement, and marketing automation.
The business owner:
- provides product and technical support expertise;
- defines business priorities;
- validates requirements and proposed solutions;
- reviews generated responses when expert involvement is required;
- acts as the primary source of domain knowledge.
The initial version supports a single primary request per email.
Additional customer requests may be identified and stored as contextual information but do not trigger independent workflow execution. Routing is determined by the primary category and next action.
Multi-intent workflow orchestration is intentionally excluded from the MVP to reduce solution complexity.
The MVP does not introduce a new CRM or order management system. It uses existing operational data sources:
- email inboxes and historical sent emails;
- Google Sheets order records;
- Google Drive folders;
- product manuals and support documentation.
The data model defines the information objects required for classification, retrieval, response drafting, workflow execution, and source traceability.
- AS-IS and TO-BE process models
- Functional and non-functional requirements
- AI inquiry classification model
- RAG and retrieval design
- Data model
- Integration and sequence diagrams
- Use cases and acceptance criteria
- AI governance and escalation rules
- n8n workflow design
The current phase focuses on requirements analysis, solution design, and architecture definition.
An end-to-end n8n workflow may be implemented as a limited proof of concept in a future phase.