Warmwind publicly launched its autonomous AI workers on August 26 after three years of development and a multi-month closed beta. Each worker runs on its own isolated cloud computer in Germany and operates ordinary software visually -- moving a virtual mouse, typing on a virtual keyboard -- rather than calling APIs, which lets it drive interface-heavy systems like SAP that were never built to be automated. You teach it by demonstrating the job once, and it repeats it on a schedule while your own machine is off.
Key facts
- Pricing averages EUR 1.00 to EUR 1.50 per hour of active execution, depending on task complexity and the AI model selected.
- Pre-launch registrations grew from 12,000 to nearly 90,000 across more than 30 countries; roughly EUR 450,000 in annual recurring revenue in the two weeks before launch.
- 15 employees in Jena, Germany; customer infrastructure hosted in Falkenstein; a EUR 1.5M seed round already closed, plus a new multi-million-euro strategic investment from Hetzner.
- Launched August 26, 2026. Primary source: Warmwind launch release.
The hook. The pitch is not "write a better prompt." It is: take over the cursor, do the job once while narrating why you are clicking, and stop. The demonstration becomes the workflow.
Background. Automating office work has historically meant one of two things: integrating systems through APIs, which is expensive and impossible when the vendor never built one, or robotic process automation, which records clicks and breaks the moment a button moves. GUI agents -- models that look at a screen and decide where to click -- are the third path, and they have a long history of working beautifully in demos and failing on the second unfamiliar screen. Our explainer on AI agents covers the general shape.
What they built. Each worker gets a dedicated, isolated cloud computer and "interacts visually with software using a virtual mouse and keyboard -- just as a human operator would," in the company's words. Once a workflow is configured, workers execute independently, repeat on schedule, and keep running when the user is offline. Multiple workers run in parallel, which is how the company frames capacity expansion: more workers, not more integrations. Named use cases are invoice processing, customer support, ERP data management, portal monitoring and lead generation. The workers operate across Windows, Linux, Android and web applications, with browser, desktop and mobile clients for monitoring and control.
How it works. Think about training a temp. You do not hand them API documentation. You sit down, open the three programs, and walk them through the sequence: pull the invoice from email, look up the order number in the ERP, paste it into the spreadsheet, reply to the customer. They watch, they take notes, and next Tuesday they do it alone. That is the interaction model, and it is why "vision-first" matters: the agent's interface to your software is the same one you use, so nothing has to be integrated. Warmwind's own architecture writeup says it can delegate sub-steps to reasoning models when a task needs research or judgment.
Why it matters. Maximilian Schilling, CEO and technical founder, framed the market this way: "Many organizations lose a significant portion of their workday to repetitive screen work across disconnected systems. With Warmwind, we are building an execution platform where AI workers take over these structured workflows." His co-founder Richard Wieduwilt drew the boundary: "Our goal isn't to replace humans in processes. It's about taking the burden of repetitive screen tasks off human teams." The commercial signals are modest but real -- roughly 450,000 euros in recurring revenue before the public launch, live deployments with mid-sized companies and DAX-listed corporations, and a Hetzner partnership sized to scale to 10,000 paying users.
Warmwind also published its own benchmark, Pointer Bench (repo, dataset): 1,500 synthetic tasks across spreadsheets, text and professional applications, built because general benchmark scores do not transfer to office software. The interesting signal there is not the leaderboard position but the unevenness -- GUI grounding stays brittle exactly where you would predict, in dense text, text cursors, and spreadsheet structure.
The honest caveat. The security story is not disclosed. An autonomous worker that opens your email and your ERP has to hold credentials somewhere, and Warmwind's public material does not say where they are stored, whether the agent uses your own logins, or how secrets are scoped. Google Play's listing for its companion app says data is encrypted in transit and not shared with third parties, and the launch emphasises isolated German cloud instances with technical separation between customers -- but that is infrastructure, not credential architecture. A commenter on the launch thread asked exactly these questions and got no public answer. We have written before about agents shipping with standing logins to email and CRM; it is the same risk surface.
Beyond that: no independent evaluation of the product exists. The entire evidence base is company-authored material and the company's own benchmark. For technical users, OpenClaw is the open-source, runs-on-your-own-machine alternative, and the coverage itself names it as the comparison. Warmwind is the managed version for people who do not want to build the agent.
Originally published on Ground Truth, where every claim is checked against the primary source.
