glossary
AI implementation, in plain English.
Every term you'll hear on a discovery call, defined for operators, not researchers. Bookmark this. Send it to the team before the next vendor pitch.
AI agent
An AI system that takes a goal, picks actions, runs them, observes the result, and iterates until the goal is complete. Different from a chatbot: a chatbot replies, an agent does work.
An inbound lead agent reads an email, looks up the contact in your CRM, drafts a reply in your voice, schedules a viewing, and updates the deal stage — all without a human in the loop.
RAG (Retrieval Augmented Generation)
Pattern where an AI model retrieves relevant context from your private documents before answering. Lets the model give answers grounded in your company knowledge instead of hallucinating from training data.
A property manager asks 'what's the deposit policy for unit 14?' The system retrieves the relevant lease + house manual, then the model writes the answer citing both.
Voice AI
AI that holds a real-time spoken conversation: speech-to-text + LLM + text-to-speech, stitched together with sub-300ms latency so the interaction feels human.
An inbound buyer calls a brokerage at 11pm. A voice agent qualifies the budget, checks availability against the calendar, books a viewing, sends a confirmation SMS — all in one call.
LLM (Large Language Model)
The underlying model that reads and writes text. Claude, GPT-4, Gemini, Llama, and Mistral are all LLMs. Picking the right one per task balances cost, latency, and reasoning quality.
Claude Sonnet for production agents that need judgment. Haiku for high-volume classification tasks. GPT-4o for vision tasks. Llama running locally for sensitive data that can't leave your cloud.
Embeddings
Numeric representations of text that capture meaning. Two sentences with similar meaning end up with similar embeddings. The math behind semantic search and RAG.
A buyer asks 'something near the beach with parking.' The system converts that to an embedding, finds listings whose embeddings are closest in meaning, then returns them — even if the listings never used the word 'beach.'
Vector database
A database optimized for storing embeddings and finding nearest-neighbour matches fast. The substrate underneath any production RAG system.
Pinecone, Weaviate, pgvector (Postgres extension), and Supabase Vector are common picks. We default to pgvector when the data is already in Postgres.
Fine-tuning
Continuing to train an existing LLM on your specific data so it adopts your tone, vocabulary, or task. Usually overkill — RAG + good prompting beats fine-tuning for 90% of business use cases.
Fine-tune when you need consistent brand voice at scale (10k+ outputs/month) or domain-specific reasoning. Skip when you just want the model to answer questions about your docs (use RAG).
MCP (Model Context Protocol)
Open standard for connecting AI models to external tools, data, and services. Lets one agent talk to dozens of systems (CRM, calendar, database, file storage) through a unified interface.
Instead of writing custom integration code for every tool, we ship MCP servers that expose your CRM, calendar, and inbox to any agent. Swap models or vendors without rewriting integrations.
Agentic workflow
A multi-step process executed by one or more AI agents, with handoffs between steps and tools. Replaces multi-person manual workflows with one orchestrator running 24/7.
Lead intake: agent reads email → enriches contact → scores fit → assigns owner → drafts personalized reply → schedules follow-up reminder. What used to take an SDR 20 minutes now takes 12 seconds.
Hallucination
When an LLM produces confident output that's factually wrong. Mitigated by grounding the model in your real data (RAG), running validation checks, and routing low-confidence outputs to a human.
Production systems always run a verification pass: 'does this answer cite a real document?' or 'does this generated SQL parse?' Failing answers escalate instead of ship.
Prompt engineering
Crafting the instructions you send to an LLM to get reliable output. Less art, more discipline: clear role, explicit format, examples, edge cases, escape hatches for uncertainty.
A prompt that says 'reply professionally' is unreliable. A prompt that says 'reply in 3 sentences using the tone of the attached email examples, ending with one call-to-action question' produces consistent output.
Self-healing system
AI system that detects its own failures, retries with adjusted parameters, and escalates to a human only when it can't resolve. Reduces operator load to exception handling instead of monitoring.
If an agent fails to extract a date from a PDF, it retries with a different OCR model, then a different prompt, then routes to the human queue with the specific failure logged.
Want one of these built into your business?
Free discovery call. Honest read on whether AI moves the needle.