What to build yourself
Build agents yourself where you are the only user or an employee sees the mistake at once. The tools for this are open or can be installed on your own computer, and step-by-step guides are in our section.
An assistant for yourself: an instruction and your files, no code. It writes emails, sorts documents, and prepares draft reports.
Answers from internal documents: an agent with a knowledge base of rules and instructions for employees, with no customer personal data.
A Telegram bot for your team: reminders, summaries, search in the knowledge base.
Automation in a builder: an email arrives, the agent reads it and creates a task. This is a week of work in n8n or a similar tool.
The sign of a task for yourself: if the agent makes a mistake, you notice it at once and fix it in a minute, and the customer sees nothing. You will need a computer, a builder or Ollama for a local model, samples of your documents, and an evening or a week of time.
Anthropic advises starting with a simple solution and adding complexity only when needed; sometimes you do not need an agent at all. Building a simple agent in-house is a good way to check this before you talk to a contractor.
How to tell an agent for an evening from an agent for a week is described on the page about complexity levels.
What to hire for
Hire a team for an agent when a mistake costs money, customers, or a fine. Here you need a team that is responsible for the result under a contract.
An agent in the CRM and 1C (a Russian accounting system): it reads and changes deals, stock, and invoices. It needs a separate account with minimal rights, an action log, and testing on a test copy.
A voice agent: speech recognition, telephony, and a fast reply in a conversation where the customer will not wait.
An agent with customer personal data: names, phone numbers, messages, order history.
A team of agents: several agents share data and pass tasks to each other, and one agent’s mistake moves on to the next.
One more sign: the agent works around the clock without supervision. Then you need a log, monitoring, and a person who reviews errors every week, and that is a team’s job, not an evening project.
Personal data is a separate reason to hire a team. Under 152-FZ, Russia’s personal data law, data of Russian citizens is collected and stored in databases inside Russia, and before any cross-border transfer the operator files a notice with Roskomnadzor, the Russian data regulator. How we set this up is described on the page about data security.
Edge cases
Some tasks sit between the lists. The decision depends on who sees the mistake first: you or the customer.
An agent answers customers on the website from a knowledge base. Build the draft yourself in a builder and test it on your own questions; release it to live customers only after testing and with handover to a person.
An agent reads the CRM but changes nothing in it. You can build this one in-house if access is read-only and the database is not exported; writing to the CRM is already a job to hire for.
An agent on an open model on your own server. You can launch the model yourself, but server setup, updates, and data protection need an engineer.
An agent for the sales department prepares a draft reply, and a manager sends it. You can build it in-house: an employee sees the mistake before the customer does. If the agent starts sending replies by itself, the task moves to the “hire” list.
The MIT NANDA 2025 report says that buying from specialist vendors and partnering succeed in about 67% of cases, and in-house builds three times less often. This is about company deployments, not a personal assistant, but for an edge case with customers it is a strong argument.
A good compromise: build a prototype yourself to understand the task, and give it to a contractor as the brief. This way you pay for the build, not for the search for what you need.
What each path costs
Building yourself costs your time and the model expenses. Most tools can be installed in-house without a license fee, but each has its own terms: read the license before you build a product for customers on a tool.
Compare the paths by full price. Your employee’s hours for building, support, and fixes are money too; they just do not come as a separate invoice.
Model costs depend on the number of dialogs and the length of the instruction; how to count them is covered in the article about the running costs of an agent.
Our hired path: a pilot of one agent from $1,499 for 14 days, an AI department of 3–5 agents from $4,999 for 4–6 weeks, and support from $480 a month. The price and the metric are fixed in the contract before the start.
A typical agent is assembled from open blocks. If someone charges for it as for unique development, ask what makes up the price, and pay for a result by the metric, not for hours.
A full price breakdown for both paths is in the article about the cost of an AI agent.