Plato Raises $16M Series A for AI Workplace Simulations

In Plato's accounting simulation, a quarterly report reaches the CFO with an inflated revenue figure inside. Several steps went correctly: the file was prepared, attached and delivered. A separate check catches the currency-conversion error, putting a very ordinary office problem inside an AI training exercise.

The San Francisco company announced a $16M Series A on October 28, 2026, led by Nava Ventures, with participation from Gradient Ventures, Transpose Platform and Microsoft AI CEO Mustafa Suleyman. Co-founders Robert Farlow, CEO, and Pranav Putta, CTO, are building simulated workplaces for model developers to train and evaluate agents. The funding brings Plato's total to $19M and supports a business that turns professional experience into repeatable practice.

For AI labs, that means buying access to the requirements surrounding a task as well as the software used to perform it. For the professionals helping create those environments, a familiar working day becomes something they must explain precisely enough for another system to attempt and someone else to grade.

What Plato's Series A finances

Plato says the financing will expand its library of work experience, pay professionals to reconstruct their jobs in simulations and support collecting additional experience from agents performing real tasks. Its announcement identifies customer support, people operations, finance, accounting and healthcare administration among the work being recreated. The company is also developing longer sequences of activity and interactions with simulated people, extending the assignment beyond isolated actions.

Nava Ventures led the round, and its founder Freddie Martignetti is joining Plato's board. Gradient Ventures previously led the seed financing and is participating again. The announcement names Suleyman individually; it does not identify Microsoft as a corporate investor. Plato was founded in 2025, putting this financing early in the company's effort to make work environments a recurring input to model development.

The founders bring relevant experience from building training systems. Farlow developed an annotation platform at Yutori for computer-use model training, while Putta worked on computer-use models at MultiOn. Their commercial assignment now reaches into the supply of useful experience itself: finding people who understand a job, recreating the setting and preserving the requirements that make its completion meaningful.

How simulated work becomes a training product

Plato's Computer Work Val v1.0 research, published October 27, describes 300 tasks and 210 simulated applications. Those are company-reported figures. The environments combine replicas of workplace tools with mock records and checks on the state an agent leaves behind, allowing developers to repeat an assignment without repeatedly changing a live business system.

The accounting illustration makes the grading problem concrete. Sending a spreadsheet and calculating revenue correctly receive separate checks. That separation lets a developer see a mistake that a simple delivery notification would conceal, and it gives the task author a way to express which requirements must survive the handoff. The familiar application names in the demonstration describe simulated tools, rather than announced commercial partnerships.

Reproducibility matters to the buyer because it makes attempts comparable. When the environment can return to its starting condition, a developer can examine whether a change in the model changes its behavior on the same assignment. The usefulness of that comparison still depends on the exercise reflecting the work faithfully. An unrealistic shortcut can teach a model something that looks efficient inside the exercise and becomes expensive in the office it eventually enters.

The professional supplies more than instructions

Paying someone to reconstruct a job creates a demanding knowledge-transfer relationship. A professional has to make explicit the details a competent colleague might already understand: which record is authoritative, which instruction constrains an action and which mistake will remain hidden after a polished delivery. Turning that understanding into a repeatable exercise takes a different kind of effort from simply demonstrating the final answer.

For Plato, the quality of those contributions belongs inside the product. Its research describes expert review of tasks and internal checks for defects and grading mismatches. The resulting business depends on the people who can recognize when an apparently plausible workflow has drifted away from what their occupation actually requires. For a laboratory choosing a training supplier, the expertise behind those requirements becomes part of the purchasing decision.

AI labs evaluating this kind of infrastructure have reason to ask how examples are selected, how scoring is maintained and how closely the tasks resemble the work their models will encounter. Purchasing more exercises expands practice opportunities; purchasing well-defined exercises also changes what a developer can learn from a failed attempt. Plato's approach puts the definition of competent work into the same commercial conversation as the environment in which an agent practices it.

Building the organization around the environments

DevCuration's coverage of Labelbox's financing for training environments and enterprise agents and micro1's enterprise-data purchasing commitment offers adjacent context for this buyer relationship. Plato's announcement focuses specifically on constructing work simulations with domain experts, placing professional knowledge inside the environment a model developer acquires.

Plato's current careers page is linked from its official site, which advertises hiring across infrastructure, research and engineering in San Francisco. That combination fits the work described in the announcement: maintain usable software replicas, develop evaluation methods and expand the library. The public materials reviewed do not disclose revenue, named paying customers, valuation or an independently validated measure of improvement after deployment.

Those gaps leave the commercial scale open while the product mechanism is visible. A model developer can inspect the company's published methodology and then ask what access, customization and evaluation support a working relationship would provide. For an enterprise operator further downstream, the relevant question is how well that practice prepares an agent for the approvals, records and constraints inside the operator's own organization.

The $16M round gives Farlow and Putta resources to keep assembling that experience. Each additional profession brings its own software, expectations and costly mistakes into the library, with a person responsible for explaining why the tidy-looking answer would still come back for correction.