Beyond the Code: The Push for Agentic Office Automation

For the last year, the AI conversation has been dominated by the idea of LLMs writing code. But from a practitioner's perspective, the more consequential shift isn't about generating syntax—it's about automating the tedious operational glue that holds a company together. We are moving toward agentic workflows where AI doesn't just suggest text, but actually controls the computer to execute routine office tasks.

This shift is the central bet of Prentis, a new AI research lab co-founded by Ritankar Das, Reid Hoffman, and Mark Pincus. As TechCrunch first reported, the company is training models specifically to navigate the fragmented systems and documents that office workers use daily. The goal is to build agents capable of handling high-friction tasks—such as processing insurance claims or managing customs duty refund exceptions—without requiring a human to manually track down paperwork.

***

**Opinion: The Efficiency Play**

In my view, the real value here isn't in the 'intelligence' of the model, but in the cost of the execution. Prentis claims its Hive-32B model is more economical to deploy across everyday workflows than frontier APIs, asserting a cost per task that is roughly 10 times lower. If you are automating a million routine administrative tasks, the efficiency of a smaller, specialized model beats a massive, general-purpose one every time.

***

Prentis is seeking to raise $100 million at a $1 billion valuation, TechCrunch reported. The company has already secured contracts worth up to $50 million with a variety of clients, including a manufacturer, goods and clothing manufacturers, and a healthcare management service organization. Investor materials obtained by TechCrunch project an estimated annualized run rate of $75 million by the third quarter of this year, though the company notes these figures are performance-dependent and based on a contracted fee of 20% of realized savings rather than recognized revenue.

To prove its efficacy, Prentis points to two computer-use benchmarks: ScreenSpot-v2, which tests the ability to locate on-screen controls, and WindowsAgentArena, which measures end-to-end task completion in real Windows applications. Prentis claims Hive-32B outperforms rivals such as Anthropic’s Claude Opus 4.6 and OpenAI’s GPT-5.4 on these metrics, though TechCrunch noted it has not independently verified these results.

This is a crowded race. A source told TechCrunch that OpenAI, Anthropic, and Mira Murati’s Thinking Machines Lab are all developing computer-use agents. Anthropic has already moved aggressively in this space, acquiring the Seattle-based startup Vercept earlier this year to absorb its talent and shut down its product.

Ritankar Das, the CEO of Prentis, brings a history of building AI companies through his holding company, Titan. Among companies launched through Titan are Tala Health, a virtual care provider that closed a $100 million seed round last year; Forta Health, an autism care startup that raised $55 million in a round led by Insight Partners in 2024; and Dascena, a disease prediction firm bought by CirrusDx in 2022. Prentis has already scaled its team to over 25 employees, drawing researchers from Meta, Google DeepMind, Alibaba, Tencent, and OpenAI.