The Seven Building Blocks Behind Modern AI Systems in 2026

A chatbot that only talks is easy to build. An AI system that actually gets work done, safely and every day, needs a whole set of parts working together.

Today is October 22, 2026, and this matters right now because most companies have moved past the "let us try a chatbot" stage. They now want AI that answers from their own documents, takes real actions in their own tools, and does not embarrass them. A popular infographic going around this month sums up how that works in seven building blocks, each compared to part of a factory. This post walks through all seven in detail, in plain language, and shows how they fit together.

The big picture: think of a factory

Picture a small factory that turns requests into finished work.

• There is a machine that does the actual making.
• There is a raw material room where the right materials are fetched before the machine starts.
• There is a storage warehouse where all materials are kept, organised so they are easy to find.
• There is a floor manager who decides what happens next and which tools to use.
• There is a standard power outlet so any tool can be plugged in without custom wiring.
• There is a safety system that stops dangerous actions.
• There is a quality check at the end that tests whether the output is any good.

Each of these maps to one building block of a modern AI system. Let us take them one at a time.

1. The LLM: the machine

An LLM, or large language model, is the part most people already know. It is the model behind tools like ChatGPT, Claude and Gemini. It has been trained on a huge amount of text, so it can write, summarise, explain, translate and reason through a problem step by step.

In the factory picture, the LLM is the machine. It is powerful, but on its own it has two big limits:

• It only knows what it learned during training. It does not know your company's leave policy, yesterday's sales figures or the ticket a customer opened this morning.
• It can sound confident while being wrong. When it lacks the facts, it may fill the gap with something that sounds right. This is often called a hallucination.

That is why the LLM is only the first block, not the whole system.

Real-world example: Klarna. In February 2024 the payments company Klarna announced that its customer-service assistant, built on OpenAI's models, had handled 2.3 million conversations in its first month, about two-thirds of its customer-service chats. Klarna said customers sorted out their issues in under two minutes instead of eleven. These are the company's own launch figures, not an independent audit, and Klarna later stressed that customers can always reach a human. Even so, it shows how much routine writing and explaining a well-placed LLM can take on.

2. RAG: the raw material room

RAG stands for retrieval-augmented generation. The idea is simple: before the model answers, the system first looks up the relevant information and hands it to the model along with the question.

So instead of asking the model "What is our refund policy?" and hoping it guesses, the system:

1. searches your documents for the refund policy,
2. pastes the most relevant paragraphs into the request,
3. asks the model to answer using those paragraphs.

In the factory, this is the raw material room. The machine works far better when it is given the right materials first. RAG is the main reason an AI assistant can answer correctly about your own business, and it also makes it possible to show which document an answer came from.

Real-world example: Morgan Stanley. In September 2023 Morgan Stanley rolled out an internal assistant for its financial advisers, built on OpenAI's GPT-4. Instead of relying on the model's general knowledge, it answers from the firm's own library of about 100,000 research reports and documents. That is RAG in practice: find the right internal documents first, then let the model write the answer from them.

3. The vector database: the storage warehouse

RAG needs somewhere to search. That is the job of the vector database.

Here is the key idea. Each piece of text, such as a paragraph from a manual, is turned into a long list of numbers called an embedding (or vector). Texts with similar meaning end up with similar numbers. When a question comes in, it is turned into numbers too, and the database finds the stored pieces whose numbers are closest.

This is why the infographic says it "retrieves by meaning, not keywords". A question like "Can I get my money back?" can find a paragraph titled "Refunds and returns", even though none of the words match.

In the factory, the vector database is the storage warehouse: everything is stored in an order that makes the right item quick to find. Popular options in 2026 include dedicated vector databases and vector add-ons for ordinary databases such as PostgreSQL.

Real-world example: Spotify. Spotify has long used this "find the nearest neighbours" idea for music recommendations. In 2023 it open-sourced a library called Voyager, which searches large collections of vectors to find the closest matches, and Spotify says it is queried hundreds of millions of times a day inside its products. The same idea that finds "songs that feel like this one" also finds "paragraphs that mean the same as this question".

4. The AI agent: the floor manager

An agent is what turns an AI that talks into an AI that does things.

An agent is a loop built around the LLM. It looks at the goal, decides the next step, uses a tool, checks the result, and repeats until the job is done. For example, given "Reschedule my 3 pm meeting and tell the attendees", an agent might:

1. read the calendar,
2. find a free slot,
3. move the meeting,
4. send a short email to each attendee,
5. report back what it did.

In the factory, the agent is the floor manager. It does not do every job by hand; it decides what happens next and sends work to the right tool. This is powerful, and it is also where the risks grow, because an agent can now change real things.

Real-world example: GitHub Copilot coding agent. In May 2025 GitHub put a coding agent into public preview. A developer assigns an issue to Copilot, and the agent sets up its own working environment, makes code changes, and pushes them to a draft pull request that people review before anything is merged. It plans, uses tools and reports back, which is the floor-manager pattern, with a human still approving the final result.

5. MCP: the power outlet standard

MCP stands for Model Context Protocol. It is an open standard, first published in late 2024, for connecting AI applications to tools and data: files, databases, calendars, ticket systems, code repositories and so on.

Before standards like this, every AI tool needed its own custom connection to every app. Ten AI tools and ten apps could mean a hundred separate integrations. With MCP, an app is wrapped once as an "MCP server", and any AI application that speaks MCP can plug into it.

That is why the infographic calls it the power outlet standard: one plug shape, many devices. You do not rewire the building every time you buy a new appliance.

Real-world example: Anthropic and OpenAI. Anthropic announced MCP on November 25, 2024, together with ready-made connectors for Google Drive, Slack, GitHub, Git and Postgres. On March 26, 2025, OpenAI added MCP support to its Agents SDK and said ChatGPT's desktop app would follow. When two competing AI companies support the same plug, tool makers can build one connector and have it work in both.

6. Guardrails: the safety system

Guardrails are the rules and checks that decide what the AI may and may not do. They sit around the model and the agent, not inside them.

Common guardrails include:

• Input checks that block harmful requests or attempts to trick the model into ignoring its instructions (often called prompt injection).
• Output checks that catch leaked personal data, unsafe content or answers outside the allowed topic.
• Action limits that stop an agent from doing high-risk things on its own, such as deleting data, sending money or emailing customers, without a human approving first.
• Access control, so the AI can only reach the data the person asking is allowed to see.

In the factory, this is the safety system. Machines are allowed to run fast because the guards are there.

Real-world examples: Air Canada and DPD. Two public cases from early 2024 show what happens when the guards are missing.

• Air Canada: its website chatbot told a passenger he could claim a bereavement discount after travelling, which went against the airline's own policy. In February 2024 a British Columbia tribunal ruled that the airline was responsible for what its chatbot said and ordered it to compensate him, rejecting the argument that the chatbot was a separate entity responsible for its own actions.
• DPD: in January 2024 a customer of the UK parcel company got its chatbot to swear and write a poem criticising the company. DPD said the problem followed a system update and switched off the AI part of the chat while it was fixed.

Neither case needed a smarter model. Both needed better checks on what the system was allowed to say.

7. Evals: the quality check

Evals, short for evaluations, are tests for AI output. Normal software tests check that the same input gives the same output. AI is less predictable, so evals check whether answers are correct, safe and useful across many sample questions.

A typical eval set might contain a few hundred real questions with known good answers. Each time you change the model, the prompt or the documents, you run the set again and compare. Teams usually measure:

• quality: is the answer right and does it use the given sources?
• safety: did it refuse what it should refuse?
• cost: how much did each answer cost to produce?
• latency: how long did the user wait?
• success: for agents, did the task actually get finished?

In the factory, this is the quality check at the end of the line. Without it, you only find out something broke when a customer complains.

Real-world example: Morgan Stanley again. Public accounts of Morgan Stanley's adviser assistant highlight the testing as much as the model: the firm spent months preparing its documents and had human experts check the assistant's answers, and OpenAI's published case study describes an evaluation framework as central to the project. The result reported in that case study is that over 98 percent of adviser teams use the assistant. Careful evals are a large part of why people trusted it enough to use it every day.

How the seven work together on one request

Here is one request flowing through all seven, for a made-up company called Acme:

Step 1: The question arrives

An employee asks: "How many leave days do I have left, and can you book Friday off?" Guardrails first check the request is allowed and that the person is signed in.

Step 2: The agent plans

The agent decides it needs two things: the leave policy and the employee's leave balance, and then a booking.

Step 3: Retrieval

RAG searches the vector database for the leave policy and passes the relevant paragraphs to the LLM.

Step 4: Tools through MCP

The agent calls the HR system through its MCP server to read the balance and draft the leave request.

Step 5: Safety before action

A guardrail asks the employee to confirm before the request is submitted.

Step 6: The answer

The LLM writes a short, clear reply quoting the policy and confirming the booking.

Step 7: Measuring it

The exchange is logged, and the team's evals regularly test questions like this one to make sure answers stay correct as policies change.

Conclusion

The LLM gets the attention, but it is only one part of the machine room. Real-world results come from the full set: the LLM to think and write, RAG and a vector database to bring in the right facts, an agent to act, MCP to connect tools in a standard way, guardrails to keep it safe, and evals to prove it works. If you are planning an AI project, check which of the seven you already have, and which ones you are quietly hoping you will not need.

Merits

• Answers come from your own, current documents instead of the model's memory.
• Agents and MCP let AI finish real tasks, not just suggest them.
• A shared standard like MCP cuts the number of custom integrations.
• Guardrails make it possible to give AI more responsibility safely.
• Evals turn "it seems fine" into numbers you can track over time.

Demerits

• Seven parts mean more moving pieces to build, host and maintain.
• Retrieval is only as good as the documents behind it; outdated files give outdated answers.
• Agents that can act can also make costly mistakes if limits are loose.
• Guardrails can be too strict and block legitimate requests.
• Good eval sets take real effort to write and keep up to date.

Caution

This article is educational and simplifies a fast-moving field. The company name Acme and the walk-through workflow are placeholders, not a real system. The named examples (Klarna, Morgan Stanley, Spotify, GitHub, Anthropic, OpenAI, Air Canada and DPD) are summarised from public announcements and news reports at the time they were made; figures are as the companies reported them, and products and policies may have changed since. Tools, standards and best practices change quickly, so verify the current documentation and test carefully, especially guardrails and agent permissions, before relying on any AI system for real decisions or real customer data.

Frequently asked questions

• What is the difference between an LLM and an AI agent? — The LLM is the model that reads and writes text; an agent is a loop around the model that plans steps and uses tools to complete a task.
• What does RAG mean in AI? — Retrieval-augmented generation: the system finds relevant documents first and gives them to the model, so answers are grounded in real, current information.
• Why do AI systems need a vector database? — It stores text as embeddings so the system can find passages by meaning, which is what makes RAG search work well.
• What is MCP in AI? — Model Context Protocol is an open standard for connecting AI applications to tools and data sources through one common interface.
• What are AI guardrails? — They are rules and checks around the model that block harmful inputs, catch unsafe outputs and limit risky actions.
• What are evals in AI? — Evals are repeatable tests that measure whether AI answers are correct, safe, fast and affordable across many sample questions.
• What happened in the Air Canada chatbot case? — A tribunal ruled in 2024 that the airline was responsible for wrong refund information its website chatbot gave a passenger, a clear reminder that companies own what their AI says.
• Do I need all seven building blocks? — A simple chatbot may need only the LLM and guardrails, but any assistant that answers from company data or takes actions benefits from all seven.

Tags

#artificialintelligence #llm #rag #vectordatabase #aiagents #mcp #aiguardrails #aievals #generativeai #ai2026