Prentis, a four-month-old startup teaching AI systems to click, type and move through the same document workflows a human office worker would, is negotiating a $100 million round that would value the company at $1 billion. Two sources with direct knowledge of the negotiations described the terms to TechCrunch. The number that stands out is not the valuation. It is what backs it: contracted deals worth up to $50 million, signed before Prentis has shipped a mature product to the open market.
That gap matters. A $50 million contract book against a $100 million raise is a rich ratio for a company built around unproven agent software, and it signals investors are pricing Prentis less on current traction than on how fast performance-based deals could convert into recurring revenue once the agents run reliably at scale. Reid Hoffman, who co-founded LinkedIn and sits as a partner at Greylock, and Mark Pincus, who built Zynga and now leads Reinvent Capital as an investor, joined serial entrepreneur Ritankar Das, the company’s chief executive, in founding Prentis.
Prentis builds models meant to operate the software office employees already use rather than replace it. The pitch: agents that handle insurance claims processing or chase paperwork behind customs duty refund exceptions, tuned to whatever repetitive task a client wants off a human’s desk. Das has framed the company as a bet that routine desk work, not software engineering, becomes AI’s largest deployment surface.
Investor materials obtained by TechCrunch put the annualized value of those contracts near $75 million once the current quarter closes, a figure spanning a healthcare management services firm, a manufacturer and several clothing makers. Prentis’s own pitch deck adds a caveat worth repeating: those numbers reflect estimated value tied to a fee pegged at one fifth of client savings, not recognized revenue, and are explicitly described as dependent on execution still to come.
Internal testing, according to the company, has Hive-32B clearing a pair of computer-use evaluations, WindowsAgentArena and ScreenSpot-v2. The first grades whether an agent finishes a task from start to end inside real Windows applications. The second grades whether a model clicks the correct on-screen control. Prentis claims Hive-32B beats OpenAI’s GPT-5.4 on that pairing. It also claims an edge over Anthropic’s Claude Opus 4.6, all while running at roughly a tenth of the cost per task. Those are the company’s own results. TechCrunch has not independently verified them, and no outside lab has published a replication.
The market Prentis is entering is already crowded with better-funded rivals. Anthropic and OpenAI are both building computer-use agents alongside their frontier chat products. Mira Murati’s Thinking Machines Lab is chasing the same target, according to one of the people TechCrunch spoke with. Anthropic has been buying its way into the category directly: it acquired Vercept, a Seattle computer-use startup, earlier this year and shut down Vercept’s own product after absorbing its founders.
Das’s track record gives the round some cover. He runs Titan, a Berkshire Hathaway-style holding company founded in 2014 that has spun out three prior ventures. Tala Health closed a $100 million seed round, Forta Health took in $55 million from Insight Partners, and Dascena was sold to CirrusDx in 2022. Hoffman, who stepped back from Microsoft’s board last month to focus on the drug-discovery startup Manas AI, and Pincus both treat Prentis as a side project rather than a primary commitment.
For enterprise buyers evaluating computer-use vendors over the next quarter, the question is not whether Prentis’s benchmarks hold up under independent testing, though that matters too. It is whether a pricing model built on a share of client savings, still untested against Anthropic and OpenAI’s bundled agent offerings, survives contact with procurement teams used to fixed software licensing.
TechCrunch, reporting by Marina Temkin, published July 24, 2026.
