Prentis targets $100 million to automate work on PCs

Prentis enters the race for agents capable of acting on a computer

A new name is joining the already long list of companies seeking to turn large artificial intelligence models into software operators. Prentis, an AI lab co-founded by Reid Hoffman and Mark Pincus, is reportedly in discussions to raise $100 million, according to information reported by TechCrunch. The amount is not presented as a finalized deal: these are ongoing negotiations, in a sector where fundraising rounds, valuations and technology alliances can change rapidly.

The significance of this information goes beyond the amount being considered. Prentis would fit into a direction that has become central to the industry: moving generative AI from the role of conversational assistant to that of an agent capable of carrying out concrete tasks within a computing environment. The goal would therefore not simply be to answer a question, summarize a document or produce a draft. It would involve entrusting a system with sequences of repetitive actions performed from a workstation: navigating between applications, entering information, handling procedures or carrying out operations defined by a user.

This target marks an important shift from the first major market for AI agents, that of software development. Since the rise of programming assistants, demonstrations have multiplied: generating functions, explaining existing code, fixing bugs, writing tests or helping with documentation. Code offers several advantages to AI providers. It has a formal syntax, validation tools exist, and results can often be tested automatically. Automating administrative and operational work on a computer is far more complex: interfaces change, instructions are ambiguous, data are sometimes sensitive and the consequences of an error can be immediate.

The positioning attributed to Prentis by TechCrunch targets precisely this broader space. The company would be betting on the automation of repetitive computing tasks beyond software development. In other words, the promise is not to replace a code editor or add a chatbot to an office suite, but to take over part of the work involved in manipulating digital tools on a daily basis. This is a potentially vast field: most office jobs are structured around software, forms, emails, databases, spreadsheets, customer relationship systems and business portals.

The name Prentis also suggests a change in vocabulary. Companies are no longer selling only models, assistants or generation features. They are seeking to sell execution capability. From this perspective, an agent is useful only if it can understand an objective, break it down into steps, use interfaces, verify certain results and flag situations where human intervention remains necessary. The challenge is to perform these operations reliably enough for organizations to agree to delegate part of their processes.

For industry observers, the presence of Reid Hoffman and Mark Pincus brings immediate visibility to the project. Hoffman is notably known as a co-founder of LinkedIn and as an investor in numerous technology companies. Pincus founded Zynga, a company that became famous for its social games. The former is associated with the history of professional networks, platforms and technology investment; the latter with building consumer products at scale and software engagement mechanisms. Their partnership around an AI lab is therefore watched less as the simple creation of a startup than as a sign of potentially structural ambition.

Nevertheless, the published details do not make it possible to present Prentis as an already established market player. TechCrunch refers to a new lab and funding discussions. It is important to distinguish between the project, the reported intentions and the products actually available. No conclusion can be drawn at this stage about the technical performance of its future systems, their commercialization timeline, their degree of autonomy or the customers they could target.

A bet backed by two figures of American tech

Reid Hoffman occupies a particular place in the recent history of Silicon Valley. LinkedIn, which he co-founded, helped establish the professional network as a global digital infrastructure for recruitment, networking and the circulation of work-related information. His career is also associated with investment in technology companies, giving him experience in funding dynamics, network effects and markets where growth depends as much on the product as on its adoption by professional communities.

This experience is particularly relevant in the field of agents. A tool designed to automate computer work must not merely respond correctly to isolated requests. It must integrate with team habits, access rights, sometimes outdated software and validation chains. In a company, a theoretical productivity gain is not enough: the system must be considered sufficiently predictable, managers must be able to identify what it does, and teams must be able to take back control when the context requires it.

For his part, Mark Pincus built Zynga in a different universe, but one that is just as revealing of digital product challenges. Social games had to operate at scale, retain users and adapt to changing behaviors. While office work does not resemble an online game, the two fields share one imperative: designing simple interactions with complex systems. An agent technology can be highly sophisticated in the lab and fail commercially if the user does not understand how to give it an instruction, does not know when to check its action or does not clearly perceive the value obtained.

The bringing together of these profiles underscores a characteristic of the current AI wave: competition is no longer limited to the major labs training foundation models. It also involves entrepreneurs capable of identifying a highly concrete point of friction in digital work. Language models, reasoning capabilities, computer vision and tool-use mechanisms make new products conceivable. But the final product still depends on highly operational choices: which actions to authorize, which data to isolate, which confirmations to request and how to measure the quality of a completed task.

The $100 million mentioned by TechCrunch, if it were actually raised, would place Prentis among young companies with significant resources even before any potential broad release of their product. Such funding can be used to recruit researchers and engineers, access computing capacity, design a security infrastructure or fund the time needed for experimentation. Agentic AI is costly not only because models require resources, but also because evaluation is difficult. Testing an agent in real-world scenarios involves measuring its actions across a variety of software and situations.

Funding does not, however, guarantee a technological lead. In AI, capital can accelerate hiring and access to infrastructure, but it does not automatically solve reliability issues. An agent that succeeds in a browser demonstration can fail when faced with a slightly modified interface, an unexpected error message, a missing authorization or a poorly phrased request. This is precisely why announcements around agents are closely followed, but also viewed cautiously: the move from demonstration to regular deployment remains the decisive challenge.

The word lab, used in the presentation of Prentis, also deserves attention. It may refer to a desire to work on the underlying technology as much as on a single application. But no reported information makes it possible to establish precisely whether Prentis will develop its own models, rely on existing models, or combine several technologies available on the market. This distinction is economically important. Building foundation models requires considerable resources; building a specialized product on top of existing models raises more questions around integration, data, user experience and differentiation.

According to TechCrunch, Prentis sits at the intersection of a potential $100 million fundraising round and a bet on automating repetitive computing tasks.

This framework is nevertheless enough to explain the curiosity generated by the company. The project is not emerging in a competitive vacuum: it is appearing at a time when leading AI providers have begun to display their ambition to make models act in digital environments. Prentis will therefore have to demonstrate not only that an agent can carry out tasks, but also why its approach would be preferable to those of platforms already established in workplace tools.

From code copilot to operational agent: shifting the boundary

The wave of copilots has helped familiarize the public with a simple idea: AI can assist a professional within their very own tool. In software development, this approach has become emblematic. An assistant integrated into the programming environment can suggest code, recommend a fix or speed up a repetitive task. The user generally retains a close steering role, reviews the output and decides whether to integrate it. Responsibility remains human, even if the tool reduces the time spent on certain operations.

The operational agent promises an additional degree of delegation. Rather than merely providing content or a recommendation, it is intended to perform a sequence of actions. The user can express an expected result, and the system must then choose steps: open software, look up information, enter data, navigate an interface or produce a report. This ambition does not necessarily mean total autonomy. In many uses, the most realistic model is supervised delegation, with authorization thresholds and human validation for sensitive actions.

The difference may seem subtle, but it profoundly changes the level of risk. Text generated by an assistant can be reviewed before it is sent. An action carried out in a management tool may have changed data, created a file, triggered a procedure or transmitted information. The agent must therefore know when to act, when to request confirmation and when to stop. It must also produce understandable records so that the user can verify its operational reasoning, at least in the form of a history of the steps performed.

Leading competitors have already presented capabilities in this direction. Anthropic announced in 2024 a feature called Computer Use, designed to allow Claude to interact with a computer interface. OpenAI then presented Operator, an agent intended to carry out certain tasks on the web from a browser. Google also showed Project Mariner, a prototype agent capable of interacting with web pages. These initiatives are not identical in their maturity, access or scope, but they all testify to the same battle: giving models the ability to manipulate interfaces rather than limiting them to generating text.

Prentis would therefore enter a market where major platforms already have models, computing resources and large user bases. This makes its bet more demanding, but not necessarily impossible. New entrants can target specialized workflows, an experience better suited to certain professions, stronger control guarantees or specific integrations. In enterprise software, the advantage does not always come from the most general-purpose model. It can come from understanding a precise business process and the ability to be deployed in a constrained environment.

Automating ordinary PC tasks also differs from traditional industrial automation. Process automation tools, often referred to by the acronym RPA, were designed to reproduce highly structured sequences: clicking predictable locations, copying fields, moving files or applying defined rules. Their limitation becomes apparent when the context varies, when a document does not follow the expected template or when an interface is modified. AI models could provide more flexibility by interpreting language, images and less rigid situations.

But this flexibility also introduces uncertainty. A deterministic tool will apply the same rule to the same input; a system based on a generative model may interpret an instruction differently or produce an unintended action. For a company, the challenge is therefore to combine AI adaptability with the safeguards of traditional automation. This is one of the sector's central challenges: it is not enough for an agent to succeed often. It is necessary to determine which error rate is acceptable depending on the type of task, and which procedures can limit the impact of a failure.

The most accessible tasks are probably those that are repetitive, reversible and easy to verify. By contrast, operations involving payments, legal decisions, health data or personal information require much stricter protections. The project attributed to Prentis thus fits within a continuum: between the assistant that helps an employee perform their work and the system that carries out, under supervision, part of a process. The line separating these two models will largely determine the speed of adoption.

Reliability, security and governance: the conditions for enterprise deployment

An agent's ability to use a computer opens possibilities, but it also expands the risk surface. A chatbot with access only to text provided by its user can make a mistake in an answer. An agent connected to business applications can potentially encounter confidential data, credentials, internal documents or tools that have effects on the company's operations. Security therefore cannot be treated as an ancillary feature. It becomes a condition of use.

This issue is particularly important when the system reads content originating from outside. A web page, email or document may contain instructions intended to disrupt the model's behavior. Specialists refer in particular to prompt injection: the system may be exposed to text that seeks to make it ignore its objectives or carry out an undesirable action. For an agent tasked with browsing the web or handling communications, the ability to distinguish useful information from an unauthorized instruction is fundamental.

Permission management is another issue. In an organization, employees do not all have the same rights in software. An agent should not become a means of bypassing these rules. Its access must be limited to what is necessary to perform the requested task, and high-impact actions must be capable of being restricted or subject to validation. The promise of automation may lose all its value if it requires, in return, giving a system excessive privileges.

Traceability is just as essential. When an employee performs an operation, enterprise systems often record their account, the time and the change made. With an agent, it becomes necessary to know what objective was given, which steps were attempted, which data were used and which final action was initiated. These elements are used to correct errors, but also to meet internal control obligations. In regulated sectors, they may determine whether using the technology is possible at all.

For Prentis as for its competitors, performance measurement cannot be limited to spectacular demonstrations. Companies will ask for evidence in ordinary situations: what happens if a page loads slowly, if a piece of data is missing, if a form has a new field, if a user gives an incomplete instruction? Robustness is often determined in these less visible situations. A useful agent must know how to recognize its limits, request clarification and avoid concealing uncertainty behind an irreversible action.

Cost is also an operational parameter. An agent that visits many pages, interprets screenshots or makes multiple calls to a model may use more resources than a simple search tool or rules-based automation. To be adopted at scale, it will have to demonstrate a net economic value: time saved, fewer errors, improved processing time or the ability to absorb an increase in volume. The debate will therefore not only concern what AI can do, but what it can do profitably and reliably.

The social question cannot be dismissed either. Automating repetitive tasks does not automatically mean eliminating entire jobs. In many organizations, these tasks are intertwined with broader responsibilities: responding to exceptions, understanding a client, coordinating teams, applying judgment or assuming a decision. However, automation can alter the composition of work. Employees could spend less time on data entry, research or moving from one tool to another, and more on oversight, relationships, analysis and handling non-standard situations.

This transformation will also depend on training. An agent does not necessarily replace skills; it shifts skills toward the ability to formulate a request, verify an output and detect an error. Companies deploying such tools will have to define rules of use and avoid presenting automation as a guarantee of accuracy. Commercial messaging around autonomy will have to be confronted with the reality of professional responsibilities.

What Prentis's emergence may mean for France and Europe

For French and European organizations, the arrival of new labs focused on agents comes in a particular context. Generative AI is already widely discussed in large companies, public administrations and small and medium-sized enterprises, but the move from experimentation to structured deployments raises questions of sovereignty, data protection and integration with existing systems. An agent capable of acting on a computer does not merely handle abstract requests: it may come into direct contact with the tools through which strategic or personal information circulates.

The European data protection framework is therefore immediately relevant. When a system processes personal data, organizations must ask which data are sent, where they are hosted, how long they are retained and who can access them. These questions are not specific to Prentis, whose offering is not detailed in the reported information, but they will apply to any provider seeking to offer agents to European companies. Data location, subcontracting contracts and security mechanisms will be part of the assessment just as much as model capabilities.

The European AI Act adds another layer of context. The European regulation on artificial intelligence establishes a risk-based approach, with varying obligations depending on use cases. Not all office agents will necessarily fall into the same categories, but the use of AI systems in sensitive domains, particularly when they influence decisions concerning people, calls for increased vigilance. Providers will have to clarify the limitations of their products, and users will have to avoid confusing operational assistance with the delegation of consequential decisions.

Language is also a concrete issue. An agent used in a French company must be able to understand French wording, documents and interfaces, including professional abbreviations and the particularities of each organization. But linguistic mastery is not enough. Work processes vary by country, sector and the tools used. A solution designed for American procedures cannot be assumed to be suited, without effort, to French administrative, accounting or regulatory practices. This is where integrators, local software vendors and business teams can retain an important role.

The European market is not limited to purchasing products designed in the United States. Local players are working on models, cybersecurity, automation and business software. The rise of agents could create new opportunities for these companies, particularly in areas where proximity to customers, regulatory constraints and knowledge of local processes matter more than model size alone. French companies specializing in enterprise resource planning software, customer relations, human resources or public services will have an interest in observing how agentic interfaces change user expectations.

For user companies, the challenge will be to avoid two opposite mistakes. The first would be to ignore these technologies on the grounds that they are imperfect. Agents can already push organizations to map their repetitive tasks and identify poorly documented procedures. The second would be to deploy them too quickly, granting broad access to critical systems without testing or supervision. Between these two extremes, gradual adoption appears more coherent: select limited processes, define success criteria, retain human validation and measure incidents as closely as time savings.

The Prentis case also recalls that competition is now being fought on several levels. Foundation models remain essential, but market access will depend on the application layer: connectors with existing software, identity management, logging, administrator interfaces and adaptation to use cases. For Europe, the question is therefore not only whether to have competitive models. It is also whether to master the tools that will allow agents to enter work environments without creating new uncontrolled dependencies.

A battle still open for control of digital work

The $100 million fundraising round being discussed by Prentis, if it materializes, would illustrate the intensity of capital being mobilized around AI agents. Investors are no longer betting solely on text or image generation. They are seeking companies capable of capturing a share of budgets currently devoted to productivity software, outsourced services, automation and support operations. The prospect is immense because digital work involves a multitude of micro-tasks, often of little value individually but substantial when repeated across an organization.

The difficulty is that this opportunity is coveted by very different players. Major labs have general-purpose models and research capabilities. Enterprise software providers already possess the data, users and integration points. Automation specialists know structured processes. Young startups, for their part, can try to build a more direct experience or focus on use cases overlooked by general-purpose platforms. Prentis will have to find its place in this value chain.

The strategic question will be which layer retains the most value. If models become widely interchangeable, the advantage could shift toward business data, integrations and enterprise trust. If, on the other hand, certain agent capabilities remain strongly tied to proprietary models and specific computing infrastructure, major labs could strengthen their position. In both cases, the end user will probably seek less a model than an outcome: a completed, controlled, documented task performed without multiplying manual steps.

It is still too early to know whether the term AI agent will tomorrow describe a universal autonomous product or a collection of highly specialized tools. The latter hypothesis appears compatible with the real diversity of office work. A finance team, customer service department, legal department or public administration do not handle the same information and do not accept the same margins of error. The most useful systems may therefore be those that combine general intelligence with rules, permissions and knowledge specific to each context.

Prentis thus deserves attention not because its success is assured, but because the project crystallizes a change in target. After copilots that suggest and write, companies are trying to build systems that execute. According to the account reported by TechCrunch, Prentis's founders are betting on repetitive computing tasks, a territory far broader than code alone and far more difficult to automate safely.

In the long term, the real test for this new generation of companies will not be showing that an agent can click on a screen. It will be demonstrating that it can become a reliable component of daily work: capable of acting within a clear scope, complying with rules, flagging its uncertainties and leaving humans in control of important decisions. It is on this ability to turn spectacular automation into professional infrastructure that the next stage of the agent market will be decided, in North America as well as in France and the rest of Europe.