More documents won't help. AI needs a procedure, access, a library and experience.

The first instinct with AI is to give it everything: manuals, spreadsheets, client history. And then wonder why it gets lost. There are four different ways AI gets knowledge about your company, and each solves a different problem.

A pile of documents turns AI into a confused newcomer.

An AI model knows a lot about the world but nothing about your company. It doesn't know how you handle complaints, where your orders are or what happened the last time a client complained. You have to supply this knowledge.

The easiest option seems to be pasting everything into the window and letting AI figure it out. But it often doesn't work: AI gets lost, wanders into dead ends or answers in general terms, not the way things work at your company.

Let's go through them with one example. An e-⁠shop customer writes that the goods arrived damaged and wants a refund.

A skill is a guide for one type of task. It describes the steps and the decisions: when to proceed and when to hand the task to a person. AI loads it only when the task actually arrives, so it doesn't take up attention needlessly.

Without the skill, AI wouldn't know this procedure. It isn't in any data it learned from. It exists only in your company. A skill can tell AI ‘check the order’, but it can't get into the order system on its own.

MCP (Model Context Protocol) is a common way for AI to connect to tools. Each system gets its own MCP server, which knows how to talk to it and offers that to AI as an action it can call.

Anthropic introduced MCP in November 2024 as an open standard. Today OpenAI, Google and Microsoft support it too, and since December 2025 it has been governed by the Agentic AI Foundation under the Linux Foundation. One connection therefore works with several AI tools.

An MCP server can offer AI three things: tools, meaning actions such as finding an order or opening a ticket, resources, meaning data to read such as a file or a record, and prompts, meaning templates for common tasks.

Why isn't a plain API enough? Every system has a different API. Without MCP, you need a separate connection for every pair of AI tool and system. With MCP, a system is connected once and any AI tool that supports MCP can use it. For the company, that means less custom code and an easier switch of models.

Just as important is what the MCP server allows AI to do. A good server offers only the actions the task needs and respects the rights of the person logged in. A complaints agent sees orders and shipments, not payroll. For more on how systems connect through APIs, see the article Your systems can talk to each other.

So it has the procedure and sees the data. But it still doesn't know which rules apply at your company.

RAG (retrieval-⁠augmented generation: searching documents before answering) means AI pulls out only the part of your documents it needs right now. You don't dump everything on it in advance. Someone put the documents in the library on purpose: terms and conditions, price list, internal rules.

How it works. Documents are first split into smaller chunks, for example by paragraph. Each chunk is converted into a vector: a long series of numbers that captures the meaning of the text. Texts with similar meanings have similar numbers. The vectors are stored in a vector database.

When a question comes in, it is converted into a vector the same way. The database finds the chunks whose meaning is closest to the question, even if they don't contain the same words. ‘Damaged goods’ will also find a paragraph about a ‘defective product’. AI gets only these chunks and answers from them, with the source.

What is used today. Searching by meaning sometimes misses exact terms, such as a contract number or a product code. That is why it is combined with keyword search, which is called hybrid search. Another model then ranks the chunks found by how well they actually answer the question. For questions that connect many documents, a graph of relationships between concepts helps (GraphRAG). And in agents, RAG becomes a tool: AI decides itself whether and where to search, and if the first results aren't enough, it searches again.

RAG is only as good as the documents. If the library holds two versions of the price list, AI may cite the old one. That is why every document in the library has an owner who keeps it up to date. How vectors work and what works better today is covered in the article Vector databases and RAG.

Memory is similar to RAG but differs in one thing: where the knowledge comes from. A person puts documents into the library. AI creates memory itself from previous cases.

Short-⁠term and long-⁠term memory. Short-⁠term memory is what AI holds during one conversation or task. When it ends, it disappears. Long-⁠term memory is notes AI stores outside the conversation and reads next time: what worked, what the client prefers, where something went wrong.

It can be stored in different ways: as text notes, as records in a database or, as with RAG, in a vector database, so AI can find a similar experience by meaning.

The risk of memory is that AI also remembers a mistake. A wrong conclusion from one case then repeats in the next ones. Memory therefore needs rules: what AI may write down, who checks and corrects the entries, and when an old note is deleted. For customer data, GDPR applies too.

In practice: start with memory a person can see. For example, a file of the agent's notes that someone reads and corrects once a week.

Signpost in one sentence

It all comes down to a simple rule:

Most companies need a combination, not just one. But you don't need all four at once. Start with what is missing most in the work that repeats most often.

Jaro Rais

Founder of itashu. Builds agents and connects AI to company systems.