AI Agent Controlling a Computer: Claude 3.5 Sonnet and n8n Open a New Era of Automation Without API

Most small and mid-sized business owners in Poland associate automation with building blocks: connecting a CRM system with Gmail, and an online store with an accounting program. One condition must be met – both tools must have a so-called API (application programming interface). Unfortunately, in the Polish business reality, "closed" systems still prevail – old warehouse programs, niche industry portals, or desktop applications without integration capabilities. Until recently, the only solution was manual data transcription or investing in expensive RPA (Robotic Process Automation) systems. This has just changed.

Claude 3.5 Sonnet with the "Computer Use" feature: AI that clicks for you

On October 22, 2024, the company Anthropic announced a breakthrough update to the Claude 3.5 Sonnet model, introducing a feature called "computer use". According to the manufacturer's official announcement, this model is the first on the market capable of interpreting computer screen content in real time, moving the cursor, clicking buttons, and typing text just like a human does. Instead of rigid integration code, the AI "looks" at the interface of any application and executes given instructions within it, which opens up entirely new possibilities in the ecosystem of tools such as n8n.

What does this actually change for a small business? The end of the no-API barrier

For the SME sector, this novelty primarily means a drastic reduction in the entry threshold for advanced automation. According to a report by McKinsey & Company, generative artificial intelligence has the potential to automate tasks that currently take up 60% to 70% of employees' working time. Previously, the obstacle was implementation costs – creating a dedicated "bridge" between systems cost tens of thousands of zlotys.

Real benefits for a business owner include:

• Reduction of legacy system maintenance costs: You can automate an old accounting program or CRM that hasn't been updated for years and lacks an API, without the need to replace it.
• Time savings: Processes that required manual logging into external portals (e.g., checking shipping statuses with multiple couriers at once) can now be performed by an AI agent in the background.
• Flexibility: The agent doesn't need instructions for every click – it understands the command "find the invoice from company X and download it to the folder," even if the website layout has changed slightly.
• Scalability: One n8n agent using the Claude model can handle processes for which you would previously need several full-time employees performing tedious copy-pasting.

It is worth paying attention to the numbers. According to Gartner, by 2028, as many as 33% of business applications will include agentic AI features that will make autonomous decisions on behalf of the user. The "computer use" feature means that a small company from Poland can implement these solutions today, without waiting for multi-million dollar corporate budgets.

How to use a computer-controlling agent in n8n? Scenarios and numbers

The n8n platform is ideally suited for the role of an orchestrator that manages the work of the AI agent. The "Computer Use" scenario differs from classic automation in that instead of sending a query to a database, n8n launches a container with a browser or desktop to which Claude has access via API. The table below presents a comparison of traditional automation, manual labor, and the new generation of AI agents:

Practical implementation example: Imagine a transport company that must log into three different transport exchanges every day that do not provide an API. In n8n, we create a scenario that:

1. Is triggered every 15 minutes (Trigger).
2. Runs an AI Agent node connected to the Claude 3.5 Sonnet model.
3. The agent receives instructions: "Go to the exchange X website, log in with data from the n8n vault, search for orders on the Warsaw-Berlin route for 3.5t trucks, copy the data to a Google Sheets spreadsheet, and notify me on Slack if the price exceeds 2000 zł."
4. The model takes screenshots, decides where to click, transcribes the data, and finishes the process.

The cost of such an operation? At current Anthropic API prices, a single check of the exchange by an agent costs around 0,20 - 0,80 PLN (depending on the number of steps/screenshots). On a monthly scale, the technology cost is a fraction of the cost of an employee's labor who would have to do it manually, distracting them from more important sales tasks.

However, one must remember the risks. This feature is currently in the public beta phase. Anthropic openly admits that the model may make mistakes when precisely clicking on small interface elements (e.g., sliders). Therefore, in n8n, it is crucial to use so-called "human-in-the-loop" – automation performs 90% of the work, and a human only approves critical actions, e.g., the final sending of a transfer or order confirmation.

Summary

The introduction of the "computer use" feature to Claude 3.5 Sonnet is a turning point. Automation stops being limited by programmers and API availability, and begins to depend solely on whether a given process can be logically described and performed on a monitor screen. For a small company, this means the possibility of digitizing processes that previously seemed "unautomatable."

My recommendation: If employees in your company spend more than 2 hours a day transcribing data between systems that "don't talk to each other," test an AI agent in n8n. Start with a small, repeatable process that is not critical for financial security to safely check the potential of the new Anthropic technology.