Stop asking if it's AI. Ask which part.

When a buyer asks “is there AI in it?” the honest answer is almost always: which part? “AI” has stopped being a useful word. The four things people actually mean — LLMs, RAG, agents, MCP — each do one job well, each have real limitations, and only one or two of them are usually the right tool for any given workflow. We wrote a plain-English primer on all four. This post is the headline version, plus why we think the distinction matters more than the marketing copy suggests.

The four layers, in one paragraph each

LLMs — the brain. A Large Language Model reads and writes language. Text in, text out. Summarize a 200-page document, draft an email, translate, classify, explain. On top of pattern-matching from training data, modern LLMs can reason about, plan, and analyze problems expressed in words. What they can’t do, by themselves: look anything up in your data, take actions, or persist knowledge across sessions without an external memory system.

RAG — the brain with a reference library. Retrieval-Augmented Generation lets the LLM look things up before answering. It searches a private knowledge base (your incident history, your event archive, your API docs), pulls back the relevant snippets, and hands them to the LLM along with the user’s question. The answer is grounded in your facts plus the model’s reasoning — not general internet knowledge.

Agents — the brain that takes actions. An agent is an AI system that can use tools, make decisions, and execute multi-step workflows toward a goal. It can call APIs, send emails, route tickets, query databases, post to Slack — then read the result and decide what to do next. Modern agents often combine a planner, a memory system, a workflow engine, and one or more models working together. The LLM is one component, not the whole thing.

MCP — how the brain plugs into the world. Model Context Protocol is an emerging open standard for connecting AI systems to tools and data sources. Before MCP, every product wrote its own integration glue for every tool the AI needed to use. MCP is the common plug-and-socket: a tool exposes itself once, and any compliant AI client can use it with minimal custom work. Think USB for AI.

Why this matters when you’re buying software

Most pitches that say “powered by AI” are talking about LLMs — layer 01. That’s a fine layer to be powered by, but it’s also the layer with the most well-known failure modes: it’s confident even when wrong, it doesn’t know your data, and it forgets between sessions.

If the workflow you’re trying to fix needs the system to know your facts — your historical incidents, your customer list, your inventory — what you actually need is RAG. If the workflow has a chain of steps that today takes a human ten clicks, what you need is an agent. And if you have a stack of existing tools the system is supposed to work with, what gets the integration cost down to something realistic is MCP.

“Does it have AI in it?” is the wrong question. The right question is: which of these four layers does it use, where, and why?

How the four layers show up across our products

We use a different mix in every product. The same four building blocks; different recipes.

• QORA — LLMs write the one-sentence incident summary on the operator console. RAG pulls related historical incidents so the model knows “has this person tried this gate before?” An agent classifies severity and proposes a dispatch the operator approves. MCP increasingly connects to camera, NVR, and intercom vendors as those vendors expose compatible interfaces.
• VenueFuze — An LLM drafts the post-event thank-you emails the organizer reviews. An agent composes the donor follow-up, summarizes the night for the board chair, and generates tax-receipt PDFs. RAG recognizes “Sarah, our top donor” across years of historical attendee records.
• Next-Gen Chess — The LLM translates chess-engine output into plain English a 4th-grader can read. An agent watches a game, notices a recurring weakness, and queues a personalized puzzle for tomorrow.
• Zafronix APIs — RAG over our own structured sports archive plus human-reviewed facts produces the AI-enriched storyline field on every match. We expose this for developers building on top.

None of these products is “an AI product.” They’re products that solve workflows. We picked the building blocks that earned their place.

The punch line

If you stop reading here, this is the most important sentence: none of the four layers is “AI” in the marketing–poster sense. They’re building blocks. Real products are mostly old software with carefully–chosen pieces of these layers swapped in at the right spots. That’s the job. That’s what how we build means at Zafronix.

The full version of this — with the trade-offs, the failure modes, the “what AI is and isn’t” bits we left out for length — lives on the “What is AI?” section of our How We Build page. Worth a read if any of the above raised more questions than it answered.