In Russian we gave this process an ugly word, iiizatsiya, built the same way as computerisation, and we know how it sounds. But it says straight away what it is about: AI comes into a company the way computerisation once did. Not as a separate project with a slide deck and a contractor, but as a change in how people do their work. In English we will call it AI adoption. What AI native means and why it is not the same as saying we use AI, we covered in the first article. This is the practical part, for the owner of a small company who has no IT person but does have the question we hear word for word: where do we even start?

We think that within three or four years the answer to that question becomes a market of its own. A telling story: a friend of ours, one of the founders of a large online school, sold it a few years ago and now runs a small AI agency of one or two people, telling companies what to adopt and where to begin. The demand is there because almost nobody knows the first step. Below is our version of that first step, and we begin with a mistake we have seen many times.

The mistake everything starts with

Picture a small company where an accountant comes in two days a month. The owner decides: the operations are routine, I will build a bot and it will replace the accountant. He builds it, and the bot does not work, and not because it is bad: the owner simply does not do this job every day and does not know what it actually consists of, which documents arrive mangled, what the bank is really asking, where the numbers disagree. The accountant knows, but nobody invited her to the project. And a product nobody corrects every day does not improve.

Hence the first rule, more important than all the others: the assistant is corrected and improved by the person who does the work, not by a contractor, not by the IT person, and not even by the owner unless the owner does that process personally. An employee like that can only be replaced by another employee like that. But that employee can build an assistant for themselves, if someone nearby walks them through the first step and takes over the part nobody likes doing.

Step 0. Who does what

The owner picks a process and gives the person responsible for it two hours a week to tinker with the agent. Without those two hours nothing happens, and this is the most common reason pilots die.

The person who does the process builds the agent. It sounds frightening until you see that the first agent is a chat with a role, an instruction and materials. One of us walked a psychologist with no technical background through it in a three-hour call, and she kept using it afterwards.

A consultant is needed for three things: to go through the first step together, so the person sees that it works and does not quit on day two; to take over everything manual and unloved, registrations, tokens, access, connecting to Telegram or the website; and to be reachable in the first week, when something breaks. Further down we say plainly what that costs.

Step 1. One process that repeats every day

There are three criteria: the process repeats every day, not twice a month; the input materials already exist, documents, correspondence, standard questions, a price list; and a person can check the result without getting into the technology.

Good candidates: answering the same customer questions, first-pass sorting of incoming enquiries, preparing estimates and quotes from a price list, answering from internal documents. Bad candidates: come up with our marketing, analyse our business, and generally anything that cannot be checked in five minutes.

Our own example looks like this: the School of Great Books has transcripts of several open meetings. People arrive in the Telegram channel and ask the same questions: what is this, how does it work, who is it for, where do I start. The process repeats every day, the materials are there, and the school's founder checks whether an answer is right in a minute. That is the first agent we are building now.

If you cannot pick a process, there is an AI readiness audit for exactly that: one or two days of review, and at the end two or three candidates with an estimate, and an honest answer that AI is not needed here yet, if that is the case.

Step 2. The first agent in a chat, in one evening

Open Claude or ChatGPT. Create a project, which is a folder. Put the materials in it: transcripts, the price list, instructions, examples of good answers. Start a chat inside the project and write it an instruction in ordinary words: you answer questions about the school, only from these materials, and if the answer is not in them you say so and offer to put the person in touch with a human.

Then the most important part: ask it ten real questions that customers ask, not invented ones, taken from your correspondence. Where it answers wrong, fix the instruction or add material. That is all there is to improving an assistant: not training a model, but editing the instruction and the materials. The project and the chat with a role take three minutes to create, the ten questions take an evening.

Do not start with servers, your own models and a Mac mini. We managed to look up Mac mini prices before we had built a first agent, and that was a mistake in the order of operations. Your own server and your own model are the next question. The first one is: does the agent answer ten real questions correctly.

Step 3. Put the agent where the people are

While the agent lives in your chat it helps only you. For it to answer customers it has to be where the customers are: in a Telegram bot, on the website, in the CRM.

This is where the one part that really is manual shows up. For Telegram it means going to BotFather, answering two questions (the bot's name and its address), copying the token and handing it to whoever will assemble the bot. Registrations, email accounts, passwords, tokens and access are things nobody likes doing, with AI or without it, and they stay manual in every project. We consider this the biggest headache of every rollout, and it is exactly the part that makes sense to hand to a consultant.

After that the bot can be assembled in one conversation with Claude Code or Codex, if you have a subscription, by voice or by text. GitHub is not required for this, and that is a frequent question. What is required is the token and a description of what the bot should do, the same thing you already wrote in the instruction to the agent.

Step 4. Correct it every day

The agent does not learn by itself, and this is worth understanding before you start. The model it runs on does not change from your conversations: tomorrow it is exactly the same as today, and it does not remember yesterday's mistakes unless you wrote them into the instruction. Learning happens when a neural network is trained or fine-tuned, and none of that is happening to your agent. The only things that change in an agent are the things you change: the instruction, the materials in the project, the examples of good and bad answers.

So the agent gets more accurate exactly as fast as the person who does the process looks at its answers and edits the instruction. Two weeks of such editing, in our experience, give more than a month of development without a user, because the mistakes are found on real questions rather than invented ones.

What to count after two weeks to know whether it works: how many manual steps disappeared, how much time goes into checking answers, which mistakes repeat, and whether the person can stop the agent without a technician. More on evaluating a pilot in the article on AI agents for business.

The mistakes we have seen most often

The first two are described above: replacing a person with a bot instead of helping the person build an assistant, and building something nobody uses every day. The second one kills a project in about a month, and it is not noticed right away: the bot is up, it seems to work, it is just that nobody opens it.

The third is subtler, and it is about the consultant rather than the client. It feels as though once you have handed over the knowledge the person will manage alone and you are out of work. In practice it is the other way round: once you have taught ten people to launch agents, they stay on the first-level questions while you move on to the second level, and they keep coming back with new questions. So after the handover the consultant's work does not end, it changes: instead of the first launch he handles the second-level questions. Until recently one of us was afraid of exactly the opposite.

Two more we see in others: starting with infrastructure, a server, a model, parameters, when the first question is quite different, does the agent answer ten questions from the correspondence correctly; and waiting for the perfect version, although a rough one that is already in use will get further in a month than a perfect one sitting in a folder.

What it costs

Subscriptions: if you already pay for ChatGPT, you already have Codex, it is included, you only need to install the app on your computer. One of us had paid for ChatGPT for a long time and found this out only recently. Claude with a subscription works too. A Telegram bot is free in itself, hosting for it costs a few euros a month.

Time: one evening for the first agent in a chat, then two hours a week from the person who does the process.

The consultant: with one of the authors, Maksim Ilin, it works like this: the first step is free, a twenty-minute call on which you pick the process and agree what the first working version looks like. There is a button for a company and a button for a person starting from zero on the home page of ilinmaks.com. After that the prices are public: a readiness audit from 450 euros, an agent for one process from 1,500 euros with a first working version in two weeks.

What comes next

The first bot built this way is for the School of Great Books, and it is being built by the school's founder, not by a developer. When the bot has worked its first week and answered its first real questions, a link to a case study with numbers will appear here: how many questions, how many edits to the instruction, what had to be done by hand.

Dmitry Merkulov, co-founder of the School of Great Books. Maksim Ilin, AI engineer and AI consultant.

Originally published at ilinmaks.com.