Level 1: a ready-made service
You can build it yourself in an evening: an assistant with an instruction and files, with no code. You write the instruction (who it is, how it answers, what it never does), attach documents and ask questions in a chat window.
Where it works: email drafts, summaries and answers from the rules, for you and your colleagues. Where it does not work: the assistant changes nothing in your systems and does not answer customers on its own.
An example instruction: “You are a sales team assistant. Answer from the price list and delivery terms in the attached files. If the files have no answer, say so.” Such an assistant is built without a single line of code.
The ready-made service can be a cloud chat from a model developer or a chat with an open model on your computer through Ollama. In the second case, documents never leave the computer, which matters if they contain customer information.
For many tasks, this level is enough. In its guide to agents from December 2024, Anthropic advises starting with a simple solution and notes that sometimes you do not need to build an agent at all.
Level 2: a builder and a knowledge base
You can build it yourself in a week: an agent in a builder with tools and a knowledge base. Tools (connections to email, a messenger or a spreadsheet) let the agent act, and the knowledge base lets it answer from company documents.
Builders at this level are open: Flowise, n8n and Dify. You connect blocks with the mouse: incoming message, model, document search, reply.
The week goes not on building but on testing. The agent must answer a hundred real questions correctly before it is trusted with anything beyond an internal chat.
Many companies stop at this level. According to a McKinsey survey of November 2025, 62% of organizations are experimenting with AI agents, but only 23% are scaling them.
Examples of agents of this level and the first level are collected in the list of 10 agents you can build without a programmer.
Level 3: an agent in your systems
You need a developer: the agent works with the CRM, 1C, your server and personal data. It creates deals, checks stock and books customers, which means it changes the data the company runs on.
An example: a customer writes in a messenger, the agent finds them in the CRM, checks stock in 1C, offers a delivery date and creates a deal. Every step touches the company’s live data, and the customer will see any mistake.
Here come tasks that a builder does not have: access rights to systems, an action log, a handoff to a person in a disputed case and nonstop work at night and on weekends.
Personal data changes the requirements. Under 152-FZ, Russia’s personal data law, the data of Russian citizens is collected and stored in databases in Russia, and before data is transferred abroad, the operator files a notice with Roskomnadzor, the data protection regulator.
How we store and process customer data is described on the page about data security.
A typical agent of this level is still built from open blocks. If you are asked to pay for it as a unique development, ask what the price is made of, and pay for the result by a metric, not for hours.
Level 4: a team of agents
You need a team and an AI director: a team of agents works in several departments, and the result metric is written in the contract. For example, one agent takes requests, a second checks payment and a third prepares a daily report for the manager.
The difficulty here is not technical. You need to decide which processes to give to agents, in what order and how to measure the result, or the team of agents will automate what brings no money.
At this level, agents pass work to each other, so one agent’s error becomes the next one’s input. You need shared rules, a shared log and a person who answers for the result of the whole chain.
In June 2025, Gartner named the reasons why more than 40% of agentic AI projects will be canceled by the end of 2027: cost, unclear value and weak risk control. It also said that out of thousands of “agent” vendors, only about 130 are real.
By the same forecast, agents will be in 33% of enterprise software by 2028. How systems of several agents work is covered on the page about multi-agent systems.
How to find your level
Answer four questions in order. The first “yes” is your level.
1. Are there several agents working in different departments? Then it is level 4: you need a team and an AI director.
2. Does the agent write to the CRM or 1C, work with customers’ personal data or run on your server? Level 3, you need a developer.
3. Does the agent need tools (email, messenger, spreadsheet) or a base of many documents? Level 2, you can build it yourself in a week.
4. Are an instruction and a few files enough? Level 1, you can build it yourself in an evening.
If you are unsure between two levels, start with the lower one and test it on a real task. Anthropic’s advice is the same: add complexity only when the simple solution falls short.
Where the line between doing it yourself and ordering it lies, and how to count which is cheaper, is covered on a separate page.