Across AI

There is a peculiar thing about the AI tools that promise to remember everything for you: remembering everything is not actually the hard part. Storage is cheap. Retrieval is a solved problem, more or less. The hard part - the part that a human salesperson does effortlessly and a database does terribly - is deciding what is worth remembering, and quietly letting the rest go. Across AI, a San Francisco company founded in 2024, has built a business around that second, harder idea. Its pitch, stripped of the agentic-AI vocabulary, is roughly: what if the software forgot the way a good employee forgets?

This is a more interesting bet than it sounds. The company calls its product "The Operator for Enterprise AI," and it starts with a workflow that is famously resistant to automation - business-to-business enterprise sales. Enterprise sales runs on institutional memory that lives in people's heads: which objection a customer raised eight months ago, which product nuance closed the last three deals like this one, which stakeholder actually signs. That knowledge is enormously valuable and enormously fragile. It walks out the door when the salesperson leaves. Across AI's thesis is that this memory can become a system - not a wiki, not a CRM field, but something that continuously adapts, retaining what is relevant, discarding what is stale, and surfacing the right thing at the right moment.

The company describes three capabilities stacked on top of each other. There is memory - persistent context that spans systems, unstructured data, and the institutional knowledge that usually goes undocumented. There is reasoning - what Across calls patent-pending multi-layered reasoning graphs, which are a way of encoding how a particular business actually operates, rather than treating every company as generic. And there is execution - the ability to do long-horizon work, with permissions and full lineage tracking, so a human can see why the system did what it did. That last part, the lineage tracking, is the tell. It signals that Across is aiming at enterprises who need to trust and audit an AI's decisions, not consumers who want a clever chatbot.

The distinction the company keeps returning to is between an AI that retrieves and an AI that operates. A retrieval system hands you a document and wishes you luck. An operator advances the deal. It is a marketing line, but it is also a genuine architectural claim, and it is the thing you would want to poke at if you were evaluating the product. Does it actually move work forward, or does it just produce very good-looking summaries of work that a human still has to do? The company's answer is a set of customer numbers - a 36% increase in bookings without added headcount, 26% more deals closed, 2.6 times the annual contract value recovered from aging pipeline. These are self-reported, which is worth saying plainly, but they are at least the right kind of numbers: outcomes, not engagement metrics.

In December 2024, Across AI announced an oversubscribed seed round co-led by Village Global and Cota Capital. Village Global is the venture firm whose limited partners include Mark Zuckerberg and Eric Schmidt, chaired by Reid Hoffman; Cota Capital's contribution was led by early Dropbox investor Bobby Yazdani. It is the kind of cap table that opens doors, which for an enterprise-sales company is not a trivial asset. Third-party trackers later estimated the company at roughly $2.1M in annual recurring revenue with a small headcount - early numbers for an early company.