AI agents in a company: what they are and when they help

An AI agent is a program you give a task to in plain language, and it works through the steps itself. One example is Claude Code, which reads code, edits files and runs commands. A similar tool is Codex from OpenAI. We use an agent like this every day, so here is where it helps and where to be careful.

How an agent differs from a chatbot

A chatbot answers a question, and you carry the result into your own work. An agent uses the tools you give it and does the work step by step.

An example shows the difference. Ask a chatbot how to check whether a website is on Google, and you get instructions. Ask an agent to "check which addresses in the sitemap are not on Google". It fetches the sitemap, queries Search Console for each address and returns a list. That is how we checked 38 addresses when our new website launched.

An agent versus ordinary automation

Ordinary automation follows fixed steps. When something new turns up, it stops or makes an error. An agent can adjust its next steps to what it finds along the way, but it can also head in the wrong direction. So a fixed task that always looks the same is better left to a simple script, which is cheaper and predictable. An agent helps where the data is inconsistent and the task takes reading and judgement.

What the agent does for us

• The website. It changes code and content, runs the tests and publishes new versions after approval.
• Search Console. It checks indexing and gathers search data.
• Facebook. It drafts and schedules posts, then checks them without logging in once they are live.
• Google Play. It helps with releases, the store listing and screenshots. Every change needs the owner's approval.
• Reviews. An independent agent checks the first one's work. On one article it found 10 issues in the code and 9 in the text.

This is how we publish an article. The owner picks the topic, and the agent reads official sources and writes the Polish and English versions. It then runs the tests and shows the page on the owner's computer. An independent agent looks for mistakes in the text and facts. The owner reads it all, gives feedback and only then agrees to deploy. After that, the agent checks the page in production.

Where it could help your company

These are examples that can be built if a system has an API or a data export:

• Reports. Each week the agent gathers data from several places into one summary.
• Content. From data you already have, it drafts posts, product descriptions or answers to questions you often get.
• Document housekeeping. It compares files, finds gaps and prepares a list to check.
• Moving data. It helps move data from an old system to a new one and checks that nothing is lost.
• Software work. It writes and fixes code, and a developer checks the result.

Where to be careful

• The agent makes mistakes and sounds sure of itself. Our review caught a wrong claim about Meta's requirements that the agent had written shortly before. That is why a person or an independent review checks the result.
• Access. Claude Code has several permission modes. In Manual mode it asks for approval before sensitive operations, but the documentation notes that reviewing commands before approving them is up to you. Check which mode your agent runs in. Give only the access a task needs.
• Passwords and keys. Do not paste them into a chat or a repository. A password manager helps with that, but it does not protect against the agent itself. If the agent fetches a key on its own, it can also read it and send it elsewhere. So give it a separate key with narrow permissions that you can withdraw at any time. Do not share your main passwords.
• Hidden instructions. An agent reads websites and documents, and someone can hide instructions for it inside them. This is called prompt injection. Claude Code has safeguards, but its documentation notes that no system is completely immune. Command isolation is off by default there. Even when it is on, commands can still read most of the computer unless you limit that. For important access, run the agent in an isolated environment.
• Customer data. Before you send it to an AI provider, check their terms and your legal basis. For sensitive data, talk to a lawyer.
• Cost. The tools and models are paid and prices change. Work out the cost on a specific task before you decide.

When it pays off

An agent makes sense when a task repeats, the result can be checked and someone at the company is responsible for it. A weekly report from inconsistent data or moving data between systems are good candidates.

It does not make sense for a one-off task that is quicker by hand. It does not replace legal or financial decisions or a conversation with a customer. Do not use it where nobody will check the result.

Where to start

Pick one repeating task and describe what a good result looks like. Start by only reading data, without saving or publishing anything. Once the results have been good for a few weeks, add more permissions and approve every important change by hand.

This describes the position on 4 November 2026. If your company has a task that comes up every week, contact us. We will tell you whether an agent makes sense for it.