Prentis, new AI lab co-founded by Reid Hoffman, Mark Pincus in talks to raise $100M

Prentis Aims to Automate Office Workflows with Leaner Computer-Use Models

A new AI research lab co-founded by LinkedIn’s Reid Hoffman, Zynga’s Mark Pincus, and serial entrepreneur Ritankar Das is betting that automating routine office tasks will soon eclipse coding as AI’s most impactful use case. Prentis, launched in April, is in talks to raise $100 million at a $1 billion valuation, according to two people familiar with the discussions. The startup is training models specifically for “computer use” — the ability to navigate documents, forms, and enterprise software just as a human would — and has already signed customer contracts worth up to $50 million.

For developers and builders evaluating where to place their bets in the agentic AI landscape, Prentis represents a notable pivot: instead of competing head-on in code generation, it is targeting the vast, messy workflows that dominate enterprise operations. The company’s pitch hinges on smaller, cheaper models that claim to outperform frontier APIs on key benchmarks — a strategy that could reshape how teams think about deployment costs for agentic workloads.

What Prentis Actually Builds

Prentis trains models to learn how office workers navigate routine workflows across documents and systems. The goal is AI agents that can control a computer to automate tasks like handling insurance claims or processing customs duty refund exceptions without a human needing to hunt down paperwork. The startup markets its model, called Hive-32B, as a purpose-built computer-use agent.

According to its own benchmarks, Hive-32B outperforms OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6 on two tests: WindowsAgentArena, which measures end-to-end task completion in real Windows applications, and ScreenSpot-v2, which tests a model’s ability to locate the correct on-screen control. TechCrunch has not independently verified these results, but the company’s pitch deck claims roughly 10 times lower cost per task than frontier APIs. The reasoning: a smaller, more focused model is more economical to deploy across everyday workflows at scale.

That cost efficiency argument should resonate with any team that has watched agentic API bills spiral. If Prentis’s numbers hold, it could make computer-use automation viable for processes that would otherwise be too expensive to automate with general-purpose large language models.

Early Customer Traction and Revenue Signals

Prentis has already signed contracts worth up to $50 million with several customers, including a healthcare management service organization, a manufacturer, and goods and clothing manufacturers. Investor materials obtained by TechCrunch project an estimated $75 million annualized run rate by the third quarter of this year. However, the pitch deck notes that these figures reflect “estimated annualized value based on a contracted fee equal to 20% of savings realized, not recognized revenue,” and are “performance-dependent and subject to final execution.” In other words, revenue is tied directly to the savings Prentis delivers — a high-stakes business model that aligns incentives but also introduces risk if models underperform in production.

The structure suggests Prentis is taking a results-oriented approach to monetization, which could accelerate adoption in risk-averse enterprises but also puts pressure on the model’s real-world accuracy.

The Competitive Landscape: Crowded, but Not Homogeneous

The computer-use category is far from empty. Anthropic, OpenAI, and Mira Murati’s Thinking Machines Lab are all developing AI agents for similar tasks. Anthropic has already been acquiring talent in the space — it bought Seattle-based computer-use startup Vercept earlier this year, folding in its founders and shutting down the product. For developers, this means multiple platforms will likely offer computer-use APIs, each with different cost profiles, reliability, and supported environments.

Prentis’s bet is that specialization and cost efficiency will win out over general-purpose frontier models. If you are building an internal automation tool for document-heavy processes, the choice may come down to whether a purpose-built 32B model can consistently beat a larger model on specific tasks while staying cheaper per invocation.

This also highlights a broader trend: the most powerful AI use cases for enterprises may not involve writing code at all. While best AI coding agents continue to improve for developers, the addressable market for automating back-office workflows is enormous — and arguably less saturated. Prentis is betting that non-coding automation will outpace coding as AI’s biggest use case, a view worth watching closely.

Founding Team and Backing

CEO Ritankar Das, now 31, has an unusual background. He was UC Berkeley’s youngest University Medalist in over a century, graduating at 18 with a double major in bioengineering and chemical biology before earning a master’s at Oxford. He founded Titan in 2014 after dropping out of an AI PhD program at Cambridge, where he had been a Gates Cambridge Scholar. Titan operates as a holding company modeled on Berkshire Hathaway — funded by its own exits rather than outside limited partners — and has launched several AI-adjacent companies, including Tala Health (virtual care, raised $100M seed) and Forta Health (autism care, $55M led by Insight Partners). Prentis is the latest addition to the Titan stable.

For Hoffman and Pincus, Prentis is a side project. Hoffman recently stepped down from Microsoft’s board to focus on Manas AI, a drug-discovery startup, and famously co-founded Inflection AI before Microsoft absorbed most of its team. Pincus runs a venture, but both are lending their networks and credibility to Das’s vision.

Implications for Developers and Builders

Prentis’s strategy carries several practical takeaways for teams building AI-powered automation:

• Cost matters as much as accuracy. A model that is 10x cheaper per task can unlock automations that would otherwise break the budget. When evaluating computer-use solutions, developers should benchmark cost-per-task alongside accuracy — especially for high-volume, low-complexity workflows.
• Domain-specific benchmarks are increasingly important. Generic leaderboards may not predict real-world performance on Windows applications or screen navigation. Prentis’s focus on WindowsAgentArena and ScreenSpot-v2 signals a shift toward evaluating agents in the environments they will actually operate in.
• The “coding vs. office automation” divide is artificial but useful. Most developers already use tools like Cursor vs Claude Code to boost their own productivity. Now, the same agentic muscle is being applied to back-office tasks that have never seen automation. The architectural lessons from coding agents — context management, error recovery, cost optimization — transfer directly to computer-use agents.
• Revenue tied to savings creates alignment but risk. If you adopt a solution like Prentis, your vendor’s success depends on measurable outcomes. That can be positive, but it also means the model must operate reliably in your specific environment. Due diligence on real-world accuracy is essential.

The Bottom Line

Prentis is entering a competitive but rapidly expanding market with a focused thesis: small, cheap, specialized models can beat large general ones for computer-use tasks. Early customer traction and a roster of heavyweight co-founders give it credibility, but the space is moving fast — Anthropic and OpenAI are not standing still. For developers evaluating the next wave of agentic AI, Prentis’s approach offers a compelling data point: sometimes the best tool for a job is not the biggest, but the most efficient.

Source: TechCrunch. Details as reported; verify specifics at the source.