Superhuman is expanding its Go AI assistant across its entire productivity suite, turning the tool once known mainly for email into a personal AI assistant that shows up wherever people work.

The company announced the rollout this week, bringing Go to Docs, Grammarly, GPTZero, Fathom and Mail, alongside several new features across the suite. The pitch is simple: instead of asking users to open a separate chatbot, Superhuman wants its AI to work inside the apps where work already happens.

The scale behind that bet is striking. Because its AI runs continuously across more than a million apps and websites where people write and work, Superhuman says its systems now make over 100 billion large language model calls a week — thousands of calls per user per day.

From spell-checker to AI infrastructure

Grammarly, the product at the heart of the suite, serves about 40 million daily active users and generates more than $700 million in revenue, according to its chief executive. The company has rebranded around the Superhuman name, with Grammarly now one product inside a four-product AI suite.

Chief executive Shishir Mehrotra has argued that Grammarly is widely misunderstood as a spell-checker. In his framing, it is infrastructure for AI agents: an assistant that observes what people are doing across desktop, web and mobile apps, annotates it unobtrusively and acts on their behalf. Until now that action was mostly limited to grammar. The Go expansion extends it to everything else — research, scheduling, writing and managing tasks.

The latest product, Superhuman Go, lets enterprises deploy AI agents directly where employees work, rather than in a standalone tool they have to remember to open.

Scaling to 100,000 requests a second

Running an always-on assistant at that volume is an engineering challenge of its own. Grammarly's grammatical error correction model runs continuously in the background, so suggestions have to arrive fast enough to act on without interrupting the flow of writing.

To meet demand, the company consolidated its correction pipeline into a single model of more than a billion parameters and adopted a hybrid serving strategy, combining its own infrastructure with external providers. Its ML team runs about 50 services handling up to 200,000 LLM requests per second, and a recent deployment of the correction model on CoreWeave's dedicated inference platform reached a stable 100,000 requests per second after engineers resolved ingress and autoscaling bottlenecks.

The numbers put Superhuman among the largest production users of large language models anywhere — a quiet giant of the AI era, built not on a chatbot but on the red underline.

The agent race moves into the workflow

The expansion lands as every major AI lab races to put agents inside everyday software. OpenAI, Anthropic and Google are all pushing assistants that can take actions across apps, and Superhuman is betting its advantage is distribution: it is already embedded in the workflow of tens of millions of people.

Whether users want an assistant watching everything they type is a separate question — and one the company will have to keep answering as Go spreads across the suite. For now, the bet is that the most-used AI assistant will be the one you never have to open.

By Hamza Faisal