Hoffman-backed Prentis, co-founded by Zynga's Pincus, bets $1B that office paperwork, not coding, is AI's next big market

Prentis is in talks to raise $100 million at a $1 billion valuation four months after launch, and says it already holds up to $50 million in signed customer contracts for software that automates paperwork like insurance claims and customs duty refund forms (TechCrunch).

Coding assistants grabbed the first wave of agent revenue. Prentis's thesis is that the larger category is office work that sits between documents, legacy systems, and human handoffs, and that a smaller, task-specialized model can win that category at a fraction of what frontier labs charge for general-purpose APIs. The $100 million talks and the $1 billion price tag are anchored to that bet.

The lab was co-founded by serial entrepreneur Ritankar Das, and Zynga founder Mark Pincus, and launched in April 2026 (TechCrunch). It trains computer-use models that watch how office workers navigate applications and then run those workflows themselves: an agent that drives a real operating system, opens applications, clicks through forms, and copies values between windows instead of waiting for a human to feed it the next step. Frontier labs are publicly pursuing the same class of agent.

The technical pitch is a 32-billion-parameter model called Hive-32B that Prentis says outperforms OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on WindowsAgentArena, a benchmark that scores end-to-end task completion inside real Windows applications, and on ScreenSpot-v2, which measures whether a model can locate the right on-screen control. Prentis says Hive-32B does this at roughly one-tenth the per-task cost of frontier APIs (TechCrunch).

TechCrunch has not independently verified the Hive-32B results, and the 10x cost claim is self-reported. The headline numbers read like a pitch deck because they are: Prentis investor materials project an estimated $75 million annualized run rate by Q3 this year, but the deck itself describes that figure as a contracted fee equal to 20% of savings realized, not recognized revenue, and explicitly performance-dependent and subject to final execution (TechCrunch).

Prentis's commercial structure aligns the vendor with the customer. The lab does not bill per seat or per API call. The deck describes a fee equal to 20% of the savings a customer realizes from the agent doing the work, a structure that explains how a 32-billion-parameter model can compete on price against frontier labs at the same time it gives the lab a path to a $75 million run rate. Revenue only lands if the work actually saves the customer money: if the work pays off, the lab collects 20%; if it does not, the lab collects nothing. That is a load-bearing assumption for a billion-dollar valuation.

Prentis says it has signed contracts worth up to $50 million with a healthcare management services organization, a manufacturer, and goods and clothing manufacturers. No customer is named, and the dollar figure is an estimated annualized value, not booked revenue (Prentis).

Anthropic shut down its Vercept computer-use team on March 25, 2026, three months after acquiring the Seattle-based startup for an undisclosed sum. Vercept had raised $50 million total, including a $16 million seed, and co-founder Matt Deitke reportedly left for Meta on a $250 million pay package after the deal. Anthropic, OpenAI, and Mira Murati's Thinking Machines are all publicly building computer-use agents alongside Prentis. Category consolidation is already underway, which is part of why a $1 billion valuation on a four-month-old lab is itself a signal about how the market is pricing the bet (TechCrunch).

Prentis has 25-plus researchers drawn from OpenAI, Google DeepMind, Meta, Tencent, and Alibaba (Prentis). The team is small enough that every researcher ships and large enough to train a 32-billion-parameter model in-house rather than rent one.

Office workflow agents outpace coding agents as a market, and a small specialized model priced at 20% of realized savings is the right shape for that market. A $1 billion valuation tests that specific claim. If the bet holds, a four-month-old lab clearing a frontier-tier price will look early. If it does not, the same numbers will read as a pitch deck that priced a thesis. The next data points are the closed funding round and a benchmark run by someone other than the company.