The awkward line in any rollout plan is the one that names which traffic gets inspected and which does not. Lasso's account of why that line exists is a costing argument: guardrails built on a second large language model need accelerated hardware behind every request, whether or not the request turns out to be dangerous [6], and the company says the bill pushes most enterprises into inspecting a portion of their AI traffic and accepting the risk on the rest [7].
Consider the arithmetic behind that claim. Lasso says its software covers tens of thousands of agents and billions of requests and actions a month [12]. Call it one billion across a 30-day month and that averages about 386 a second [2]. A verdict in under five milliseconds [2] means a single serialized inspection lane tops out near 200 a second [3], so even the cheap engine needs at least two lanes working concurrently just to keep up [4]. That is the arithmetic before anyone prices the same 386 a second through GPU inference, and the per-request GPU figure is the one number a buyer has to compute from their own invoice rather than a vendor's.
Teams tell themselves they have guardrails, when what they actually have is a coverage percentage that never appears on the slide. Ophir Dror, Lasso's co-founder and chief product officer, made the same point from the selling side, saying the idea of agents overseeing other agents is great "unless you are the one paying the bill" and is not feasible in deployments with millions of prompts and actions [10].
The catch sits in the routing layer, which decides which engine sees what so that most traffic never touches a graphics processor [9]. LEAP works as a classifier, screening requests automatically. Requests that need a judgment call against a plain-language policy go to a second engine, RAPID, running a self-hosted model that Lasso says costs a small fraction of the same request through a commercial provider [8]. Lasso calls that share small and does not put a number on it [8]. The split is the entire economic case: a classifier that auto-clears most traffic on someone else's benchmark and a lot less on your mix brings the accelerated hardware straight back, along with the sampling decision it was supposed to retire. The throughput claim of thousands of times existing guardrails is Lasso's own measurement [2], which is precisely what a pilot is for.
Two axes are enough to place your own situation. Across the top, the share of prompts and agent actions that passes through any inspection today, taken from logs rather than from belief. Down the side, whether the un-inspected remainder was chosen (read-only internal tooling, agents with no write path) or inherited from a hardware ceiling. High coverage that was deliberately scoped needs no new engine. Low coverage that was inherited is the buying case, and the metric worth watching in a pilot is the deferral rate on your traffic and the size of the coverage gap after a month, rather than requests processed. Elad Schulman, Lasso's chief executive, frames the requirement as inspecting not only what agents say but what they do [11], and actions are where the volume, and the cost, actually sit. The round itself is roughly 4.3 times everything Lasso had raised before it [5], which buys time to find out whether the split holds.
