Prentis is quietly becoming one of the most closely watched new labs in the AI agent space, and its planned one hundred million dollar funding round at a one billion dollar valuation shows just how quickly on screen automation is moving from experiment to serious business. The pitch is simple but ambitious: turn the way office workers already use computers into data, then train agents that can reliably take over those workflows at scale.
The rise of computer use agents
Over the past few years, large language models shifted from pure text prediction to tools that can call APIs, browse the web, and interact with applications through structured interfaces. Benchmarks such as OSWorld and WebArena emerged to measure whether models can actually control operating systems and browsers rather than just talk about doing so. AI’s role in job transformation has been pivotal in reshaping how these models are applied in real-world scenarios.
Large language models have evolved from text prediction to API-aware, screen-controlling computer use agents.
At the same time, specialized evaluations like ScreenSpot v2 and ScreenSpot Pro began to focus on precise user interface targeting, where even a small error in clicking a button or selecting a field can break an entire workflow. That broader evolution set the stage for a new generation of what researchers often call computer use agents. Instead of simply suggesting actions, these systems attempt to drive the mouse and keyboard in real applications, much closer to how a human assistant works.
Frontier models such as GPT 5.4 and Claude Opus 4.6 now top many of the standard computer use benchmarks, including variants of OSWorld and WebArena, which makes them the default choice for early adopters testing autonomous desktop agents. Against that backdrop, a lab that claims it can outperform these leaders on real screen control while cutting costs by an order of magnitude is going to attract investor attention.
Who is Prentis and what is it building
Prentis is a new AI research lab co-founded by Ritankar Das, alongside well-known technology figures Reid Hoffman and Marc Pincus. The lab was established in April twenty twenty six with a focus on computer use models that can automate high volume office work, especially repetitive tasks such as insurance claims processing and customs refund handling.
According to reporting on its investor materials, Prentis is in discussions to raise around one hundred million dollars at a valuation of roughly one billion dollars. The commercial traction behind that valuation is not purely speculative. Investor documents cited in recent coverage indicate that Prentis has already signed contracts worth up to fifty million dollars with customers across healthcare and manufacturing, and it is projecting an annualized run rate of seventy five million dollars by the third quarter of this year.
Those numbers matter because they show that enterprises are not just experimenting with agents in pilot programs; they are committing to multi-year deals that assume these systems will reliably carry a share of administrative workload.
A user interface first strategy
The core of the Prentis product strategy is to train AI agents that learn how office workers actually navigate documents, forms, and business software, then reproduce those behaviors autonomously on screen. Rather than insisting that clients refactor their systems around modern APIs or shared back ends, these agents operate directly on existing user interfaces, controlling applications through clicks, text entry, and navigation sequences much like a human staff member would.
This design targets organizations that still rely heavily on legacy software stacks and manual data entry, which is common in healthcare administration, industrial manufacturing, and regulated back office functions where workflows are complex but highly standardized. Technically, this approach tries to solve a long-standing pain point in enterprise automation.
Robotic process automation promised something similar years ago, but traditional scripts were brittle and expensive to maintain. By pairing modern vision and language models with explicit training on how screens change over time, Prentis is arguing that its agents can adapt more gracefully when layouts shift or new fields appear. If this pans out, it would let customers layer intelligent automation on top of their existing tools without a costly migration project, which helps explain the willingness of conservative sectors like healthcare to sign sizable contracts early.
Inside the Hive 32B model and the benchmark claims
Prentis describes Hive 32B as its flagship model, optimized specifically for on screen task execution rather than general chat or code generation. Investor documents and coverage around the round state that Hive 32B surpasses leading frontier models such as GPT 5.4 and Claude Opus 4.6 on benchmarks designed to measure real application control and user interface targeting, including WindowsAgentArena and ScreenSpot version 2.
These evaluations test whether an agent can follow multi-step instructions in a live environment, locate the correct region of the screen, and complete the intended workflow without human correction. There is useful context here. Independent benchmark tracking shows that frontier models dominate many of the current computer use leaderboards, with Claude Opus 4.6 leading OSWorld main, GPT 5.4 topping OSWorld Verified, and Gemini 2.5 Pro strong on WebArena.
ScreenSpot Pro scores tell a more sobering story, where specialized graphical user interface models outperform general purpose language models and frontier systems often score below twenty percent. That split illustrates how hard it remains for general models to translate textual understanding into robust low-level control. Prentis is effectively claiming that Hive 32B closes this gap.
The lab further asserts that Hive 32B runs at roughly one tenth the cost of top frontier models on a per task basis, which would allow enterprises to deploy agents across many everyday workflows rather than reserving them for only the highest value processes. If both the performance and cost numbers hold under independent scrutiny, Hive 32B would represent a meaningful shift from general language models toward specialized computer use engines tuned for reliability and economics at the screen level.
Funding, valuation, and what it signals
A one hundred million dollar round at a one billion dollar valuation places Prentis near the top tier of early agent companies in terms of investor confidence. Similar momentum is visible in life sciences, where $100m Series C funding for Apprentice is being deployed to strengthen pharma supply chains and adapt vaccines for COVID-19 variants. Part of that confidence clearly comes from the founding team. Reid Hoffman and Marc Pincus bring both capital networks and deep experience in scaling software businesses, which lowers perceived execution risk for a technically ambitious project.
Another part comes from the revenue story: contracts up to fifty million dollars and a projected seventy five million dollar run rate suggest that the lab is already converting its research into paying deployments rather than living on pure promise. The timing also matters. Enterprises are starting to move autonomous agents into production environments, and new security and governance startups have begun to raise sizable rounds to manage the risks.
NeuralTrust, for example, recently raised twenty million dollars to secure AI agents operating inside enterprises, highlighting how quickly the ecosystem around agent deployment is maturing. In that context, Prentis is positioning itself not as a general AI platform but as a focused provider of office automation agents, with benchmarks and contracts as proof points.
Implications for technology, businesses, and workers
If Prentis delivers on its claims, the immediate impact will be felt in back office operations where repetitive tasks consume large portions of staff time. Claims processing, invoice reconciliation, regulatory document preparation, and similar workflows are natural candidates for agents that can watch how humans do the work and then replicate it reliably.
For businesses, the value proposition is straightforward: fewer manual hours spent on low judgment work, faster turnaround times, and the ability to scale operations without linear increases in headcount. However, a user interface first strategy comes with trade-offs. Agents that depend on pixel-level understanding and learned navigation patterns can be fragile when software vendors roll out updates or redesign interfaces.
Success will hinge on whether models like Hive 32B can adapt quickly to those changes, or whether customers end up reintroducing the maintenance burdens that plagued earlier automation tools. There are also non-technical considerations. Enterprises will need clear controls over what agents can access, how data is logged, and how errors are detected, particularly in regulated fields such as healthcare.
For workers, the story is complex but familiar. In the near term, many deployments will likely focus on offloading the most tedious portions of jobs rather than eliminating roles outright. Over time, though, widespread adoption of reliable computer use agents could reshape entry-level administrative paths, pushing more work toward exception handling, oversight, and process design.
That shift will reward organizations that invest early in retraining and in clear human in the loop structures, rather than assuming agents can be dropped into workflows without careful redesign.
What to watch next
Several questions will determine whether Prentis becomes a foundational player in the agent ecosystem or a niche specialist. Independent evaluations of Hive 32B on public benchmarks such as OSWorld, WebArena, and ScreenSpot v2 will be critical to validate the performance claims relative to GPT 5.4, Claude Opus 4.6, and other frontier models.
Cost comparisons will also need to be tested in real deployments, where infrastructure, monitoring, and integration overhead often erode headline savings. On the business side, watch how quickly Prentis can move from a few large contracts in healthcare and manufacturing to a broader base of customers across insurance, logistics, and government services.
Revenue concentration in a small number of deals would make the story more fragile, while diversified adoption would strengthen the case for both the technology and the valuation. Just as importantly, pay attention to how enterprises blend these agents with security, compliance, and workforce strategies, an area where companies like NeuralTrust are already emerging as key partners.
The planned funding round is ultimately a signal that computer use agents are entering a new phase. The era of demo videos and pilot projects is giving way to hard questions about reliability, economics, and impact on real organizations. Whether Prentis can meet those expectations will depend not only on the strength of Hive 32B but on the lab’s ability to combine technical performance with the kind of operational trust that enterprises demand.
Conclusion
Prentis is aiming to raise 100 million dollars at a one billion dollar valuation for a focused bet on office automation using computer controlling AI agents, a move that signals how quickly this new wave of workflow automation is becoming strategic rather than experimental for large organizations. The combination of a sizable early funding target, marquee founders, and concrete customer contracts suggests investors see more than hype in this category and are starting to treat computer use agents as core infrastructure for knowledge work.
Why this funding round matters now
For more than a decade, companies have tried to automate routine office work using tools such as robotic process automation, rules based workflows, and traditional business process outsourcing. Those approaches typically required brittle scripts, long integration projects, and constant maintenance whenever user interfaces or policies changed.
Over the past two years, large language models and copilots have shifted the focus toward agents that can understand natural language and interact with applications more flexibly, from office suites to cloud dashboards. Prentis is part of a new generation of labs betting that the next step is agents that can drive a computer almost like a human operator, across many different applications, but with far higher reliability and far lower cost per task.
The decision to target a one billion dollar valuation at such an early stage reflects a belief that whoever builds the trusted infrastructure for this kind of automation could sit in the critical path of everyday office work, much as cloud providers did for computing and storage in the last decade.
Who is behind Prentis and what it is building
Prentis was launched in April 2026 by Ritankar Das together with Reid Hoffman and Mark Pincus, two founders with long track records in consumer internet and enterprise technology. The lab is focused on what it calls computer use models, systems trained to watch how office workers navigate documents, enterprise applications, and internal tools to complete routine workflows.
The goal is to produce AI agents that can operate computers on behalf of staff, handling tasks such as processing insurance claims or managing customs duty refund exceptions by collecting information across multiple systems and filling in the required forms. Rather than integrating deeply into each piece of business software, these agents watch the screen, reason about what needs to be done, and click, type, and scroll much as a human would.
Prentis reports that it already employs more than twenty five people, including researchers who previously worked at organizations such as Google DeepMind and Meta, which strengthens its claim to technical depth and experience with cutting edge model training. Building computer controlling agents that are safe, reliable, and auditable is a hard problem, and the presence of experienced researchers matters as much as capital at this stage.
Inside the Hive 32B model and claimed performance
At the core of the Prentis story is its Hive 32B model, which the lab positions as a specialized system tuned for computer use tasks rather than a general purpose chatbot. According to its materials, Hive 32B outperforms models such as OpenAI GPT 5 point 4 and Anthropic Claude Opus 4 point 6 on two benchmarks that matter for practical automation.
On WindowsAgentArena, a test that measures end to end task completion using real Windows applications, Hive 32B is said to complete more tasks correctly than those larger frontier models. On ScreenSpot v2, which evaluates how well a model can locate and interact with the correct on screen control, it again reportedly leads.
Prentis attributes this edge not to raw size but to specialization and efficiency, arguing that Hive 32B is significantly smaller and cheaper to run than general frontier models while still delivering higher success rates on the relevant computer use benchmarks. In its pitch materials, the company claims roughly ten times lower cost per task compared with frontier application programming interfaces, a gap that would matter greatly for high volume workflows if it holds up in production.
Another important aspect is pricing. Prentis emphasizes performance linked pricing and savings share arrangements, where customers pay in relation to realized cost reductions and productivity gains rather than purely on usage volume. That structure aligns the incentives of the lab and its clients and forces the company to prove that its agents can consistently deliver measurable business value.
Early commercial traction and the funding bet
Prentis is not raising on vision alone. It has already signed customer contracts worth up to fifty million dollars across sectors such as healthcare and manufacturing, focused on automating routine paperwork heavy workflows. These are domains where mistakes are expensive, regulation is strict, and the work is often repetitive but nuanced, making them useful testbeds for computer controlling agents.
The planned 100 million dollar raise at a one billion dollar valuation would give Prentis significant runway to expand compute capacity, refine Hive 32B, and scale deployment and support for these early customers. For a research oriented lab, this level of capital also makes it possible to run extensive experiments, gather long term reliability data, and invest in safety and evaluation, which are crucial for systems trusted to act on critical business processes.
From an investor perspective, the combination of early revenue commitments, strong founder pedigree, and a clearly differentiated technical focus helps justify the valuation target, even though the wider market for office automation agents is still immature and competitive.
How Prentis fits into the broader evolution of office automation
To understand the significance of Prentis, it is useful to place it in the context of previous waves of automation. Robotic process automation tools tried to capture keystrokes and mouse movements to replay workflows but often struggled when user interfaces changed or exceptions appeared.
More recently, products such as Microsoft OneDrive agents and Microsoft 365 Copilot have introduced document aware assistants that can summarize changes, answer questions, and highlight important details without opening each file. There are also frameworks like OfficeCLI that give AI agents programmatic control over office applications such as Word, Excel, and PowerPoint, enabling more structured automation inside specific suites.
Prentis is pushing further by aiming for agents that can handle full workplace processes across many systems, not just within a single productivity suite or document repository. The emphasis on measured task completion on real applications and on cost per task aligns with the practical requirements of operations teams and chief information officers who must justify automation initiatives with hard metrics, not just user delight.
Opportunities for businesses and society
If Prentis can validate its performance claims at scale, computer use agents could change how organizations think about routine knowledge work. Tasks such as claims processing, invoice matching, account reconciliation, or compliance documentation involve structured rules but also frequent edge cases. Automating these with traditional scripts is expensive and rigid, while outsourcing them can lead to long feedback loops and less control.
Agents that can watch the screen, understand instructions in natural language, and act across many systems could reduce the cost of these workflows and free human staff to focus on exception handling, relationship driven work, and strategic design. This has clear productivity benefits but also raises important questions.
One concern is employment displacement in back office roles that currently rely on repetitive computer work. Organizations will need thoughtful transition strategies, including reskilling and clear communication, to avoid eroding trust among employees as automation spreads. Another challenge is governance. When an AI agent controls the computer, every click and field entry must be logged, auditable, and reversible where possible.
There are also sector specific risks. In healthcare and insurance, incorrect data entry or misinterpretation of policy rules can lead to real harm, regulatory penalties, or customer disputes. This makes robust evaluation, monitoring, and fail safe design essential. Prentis will have to demonstrate not only benchmark scores but also incident response processes and transparent reporting across its deployments.
Competitive landscape and durability of the Prentis approach
The market for AI agents that automate office tasks is increasingly crowded, with large platforms and startups experimenting with different approaches, from in app copilots to workflow builders to full computer controlling agents. Frontier model providers can respond by fine tuning their own systems for computer use, and application vendors can deepen integration with their own assistants, reducing the need for a separate layer.
Prentis is betting that a specialized, efficient model with strong computer interaction capabilities and a business model tied directly to customer savings can carve out a durable role as an infrastructure layer. The technical argument is that specialization plus lower cost per task will win in high volume environments. The business argument is that savings share pricing will embed the lab as a partner rather than a commodity vendor.
Durability will depend on several factors. First, whether Hive 32B remains significantly more effective than general models on real world computer use tasks as those frontier models improve. Second, whether the company can expand beyond early pilots into broad production deployments across many industries without unacceptable error rates or escalating support costs. Third, whether regulators and customers develop clear standards for agent behavior, logging, and oversight that Prentis can help shape rather than merely comply with.
Key takeaways and what to watch next
Prentis is trying to turn computer controlling AI agents from a promising prototype into a practical, scalable infrastructure for white collar automation, backed by a targeted 100 million dollar funding round at a one billion dollar valuation. The lab already has credible founders, experienced researchers, early contracts worth tens of millions of dollars, and a specialized model that reportedly beats leading frontier systems on real computer use benchmarks at much lower cost per task.
The opportunity is significant. If agents like Hive 32B can reliably handle routine office workflows across diverse applications, they could reshape how organizations structure back office work, cost centers, and roles. At the same time, the risks around reliability, governance, and labor impact are real, and the competitive environment will intensify as platforms and frontier model providers push their own automation solutions.
Over the next year, the most important signals to watch will be independent evaluations of Prentis performance, case studies of production deployments beyond initial pilots, the evolution of pricing and savings share arrangements, and any early regulatory guidance on computer controlling agents in highly regulated industries. Those developments will reveal whether Prentis can move from a well funded bet on office automation to a trusted infrastructure layer in the daily workflows of organizations around the world.
