91% Resolved.
No Ticket Queue. No Agent Burnout.

The first stage of every query, classifying incoming support requests into intent categories (billing, how-to, technical, account management) and extracting key entities before any retrieval begins. When a customer asks "Why was I charged twice this month and how do I get a refund?" the engine simultaneously identifies the intent category (billing dispute), the action requested (refund), and the time reference (this month), all within 200ms of message submission. Intent classification matters because it determines which knowledge domains are searched and which response templates are applied: a billing question should retrieve from policy documents and payment FAQs; a technical question should retrieve from product docs and resolved tickets with matching error codes. Misclassified intent produces worse retrieval, which produces worse answers, so the classification layer is fine-tuned on Brainfish's actual ticket corpus, not generic customer service datasets, to match the specific vocabulary and phrasing patterns of their customer base.

A semantic search layer that queries product documentation, FAQ articles, and 24 months of resolved tickets simultaneously, finding the most relevant context for each specific query before any answer generation begins. The retrieval architecture uses dense vector embeddings rather than keyword matching, which means a customer asking "the sync keeps breaking" retrieves results for "integration sync errors" and "data synchronization failures" without needing to use the exact terms in the documentation. Each knowledge source is indexed separately and given a relevance weight based on query type: technical questions weight the resolved ticket corpus more heavily (because past resolutions contain step-by-step fix instructions that documentation often lacks); policy questions weight the policy document index. Retrieved context is assembled into a ranked package of the top-N passages, each tagged with its source document for citation. The retrieval stage is also where the system determines whether sufficient context exists to attempt autonomous resolution, low retrieval confidence triggers the escalation path before generation is attempted, not after a bad answer is delivered.

A GPT-4 class generative model that composes precise, contextual answers using only the retrieved knowledge as its source, eliminating hallucination by design. Unlike a general-purpose LLM that draws on its training weights to answer customer questions, the Brainfish generation layer is constrained: it can only use information present in the retrieved context package, and every claim in its output must trace back to a specific source document. This grounding mechanism is what makes the system trustworthy at enterprise scale, a customer asking about a specific feature limitation gets an answer sourced from the current product documentation, not from the model's training data memory of what the feature used to do six months ago. Answers are formatted to match the query type: how-to questions generate numbered step-by-step instructions; policy questions generate plain prose with the relevant policy language quoted; technical questions generate diagnostic steps followed by the most likely resolution path. Citations to source documents are included in every response, allowing customers to verify the answer and agents to audit it.

A quality gate that evaluates every generated answer against two independent signals, retrieval quality (how relevant and complete the retrieved context was) and answer coherence (how well the generated response addresses the original query), before deciding whether to deliver, route to human review, or escalate directly. The confidence gate is the mechanism that protects the 91% resolution rate from being achieved through overconfident delivery: it's not that the system resolves 91% of everything it receives, it's that it resolves 91% of the queries where it can do so with verifiable confidence. For the remaining 9%, the gate routes to the escalation handoff builder rather than delivering a likely-wrong answer that a customer would then need to correct through a second ticket. Confidence thresholds are calibrated per intent category: billing disputes require higher confidence before autonomous delivery than how-to questions, because the cost of a wrong billing answer is higher than the cost of a wrong navigation instruction. Thresholds are reviewed and adjusted monthly based on CSAT correlation data from the learning pipeline.

When escalation is needed, the system generates a complete case summary, conversation history, extracted entities, retrieved knowledge already searched, attempted resolution steps, and a plain-language description of what the customer needs, so the human agent can continue the resolution rather than starting a new intake conversation. The handoff builder is not an afterthought: it's engineered with the same rigor as the autonomous resolution path, because how the 9% of escalated tickets are handled determines whether customers trust the AI system or route around it. Agents receiving Brainfish handoffs resolved escalated tickets 45% faster than equivalent tickets arriving via cold transfer, because they arrived with context rather than needing to ask the customer to repeat everything they had already explained. The handoff also includes a suggested resolution based on the retrieved knowledge, which agents confirmed and executed in 67% of cases rather than composing a response from scratch. Better handoffs mean agents spend more time on judgment and less time on intake, which is the structural change that drove the 40% improvement in agent satisfaction scores.

A closed-loop improvement system that turns every resolved ticket, human or AI, into a training signal: tracking resolution outcomes, CSAT scores, and knowledge gap patterns, then feeding those signals back into the knowledge base and model retraining pipeline on a weekly cadence. Every ticket resolved by the AI adds to the corpus of examples the system learns from. CSAT scores on AI-handled interactions surface knowledge quality issues: if a category of questions consistently produces lower CSAT, the learning pipeline flags those answers for review and identifies whether the source documentation needs updating or the retrieval is selecting the wrong context. Knowledge gap detection identifies the questions the system could not answer with sufficient confidence, creates a prioritized list of documentation to write or update, and validates coverage once the new content is ingested. The system running at month six is materially better than at month one, not because it was retrained on a new dataset, but because 180 days of resolved tickets have continuously enriched the knowledge base with the specific questions, language, and resolution paths of Brainfish's actual customer population.