Writing emails is tedious, and replying is even worse. Superhuman, the premium email client designed to save professionals time, has just introduced an auto-draft feature for responses that, in internal testing, delivered near-final drafts with minimal editing required. The feeling, those who tried it report, is of an assistant that grasps the gist of the message and returns it with the right tone, leaving little to refine.

On the surface, it looks like a pure productivity win. But pop the hood and you realize the processing happens entirely in Superhuman's cloud, meaning every email—loaded with personal details, business strategies, or sensitive data—passes through third-party servers. At a time when data sovereignty has become a binding requirement (GDPR, sector regulations for healthcare and finance), such tools are simply off-limits for any organization that cannot outsource email management without violating internal policies or legal frameworks.

The winners are clear: the individual professional, the small business with light compliance burdens, and Superhuman itself monetizing immediacy. The losers are law firms, healthcare providers, public administrations, and large corporations under audit: they remain cut off from AI assistance that elsewhere becomes standard, widening the productivity gap between those who can delegate to the cloud and those who cannot.

The second-order effect is equally concrete: features like this raise the bar for all communication tools. A user accustomed to receiving near-final drafts will hardly go back to typing every reply from scratch. That pressure inevitably spills over onto platforms that remain on-premise for compliance reasons, accelerating the development of open-source email clients with local AI assistance. Projects like Ollama or libraries for embedding LLMs in desktop clients already show that running inference on your own hardware is not science fiction; the polish of Superhuman is still missing, but the gap is closing faster than many expect.

At an even deeper level, the concentration of email data in the hands of a cloud provider—even if there is no evidence today that Superhuman trains models on customer emails—represents a structural risk that regulators are beginning to notice. If AI becomes a standard component of email, the battle for control over information flows will turn strategic. An on-premise AI assistant then becomes not just a technical choice but a pillar of digital sovereignty: the only way to guarantee that no conversation ever leaves the corporate perimeter.

For those evaluating the deployment of AI assistants for communication, the question is not whether to adopt them, but on which infrastructure they should run. Local inference, using quantized models on consumer GPUs or enterprise servers, offers a path to replicate these capabilities without surrendering control. The trade-offs in latency, generation quality, and total cost of ownership must be analyzed case by case, and at AI-RADAR we have developed frameworks to navigate these decisions. The direction is clear: AI in email is here to stay, and those who build the right infrastructure today will determine whether the assistant of the future is a trusted employee or a stranger in the cloud.