Plato raises $16 million to build simulated workplaces for training AI
Plato, a San Francisco company that builds simulated workplaces for AI training, has raised a $16 million Series A led by Nava Ventures. The round brings its total funding to $19 million, according to the company.
- Plato raised a $16 million Series A led by Nava Ventures, bringing total funding to $19 million.
- Gradient Ventures, Transpose Platform and Microsoft AI CEO Mustafa Suleyman also participated.
- Plato builds resettable simulations of jobs in customer support, people operations, finance, accounting and healthcare administration.
- Its evaluations check how a model completes a task, including budget limits and exploitation of simulation weaknesses.
- The funding will expand its library by paying professionals to recreate their jobs and by deploying AI agents on real tasks.
Funding and investors
Plato announced the Series A on Oct. 28, 2026. Gradient Ventures, Transpose Platform and Microsoft AI CEO Mustafa Suleyman also took part. Freddie Martignetti, founder of Nava Ventures, joins Plato's board. The company's earlier seed round was led by Gradient Ventures.
Plato was founded in 2025 by Robert Farlow and Pranav Putta and is based in San Francisco. Farlow is co-founder and CEO, and Putta is co-founder and CTO.
Why simulations matter for AI that takes action
Plato's pitch rests on a shift in how AI is used. As models move from answering questions to taking action, the company says they need practice using software, responding to changing circumstances and completing multistep tasks. Plato offers AI labs realistic, repeatable environments where models can learn from feedback before they operate in live systems.
The company points to the risks of training in live software. A model learning to place an order could spend actual money, and one practicing an administrative task could alter real records. Plato recreates such workflows in simulations that can be reset and reused, so labs can train and test models without affecting live operations.
"Mastering real complex work comes from experience doing those tasks," said Farlow, according to the announcement.
What the environments cover and how they are scored
Plato works with domain experts to recreate the tools, situations and expectations of jobs in customer support, people operations, finance, accounting and healthcare administration. AI labs use the environments for reinforcement learning, a process in which models attempt tasks and receive feedback that helps them improve. The company says it is developing richer simulations with longer sequences of work and interactions with simulated people.
Plato says it evaluates both the outcome and the way a model reaches it. Completing a purchase, for example, should not count as success if the model exceeds the budget. The evaluations also check whether models exploit weaknesses in a simulation to earn a high score.
"The feedback a model receives shapes what it learns," said Putta. He said the evaluations are meant to help developers "distinguish a high score from work done correctly," by checking whether a model followed instructions, respected constraints and completed the task.
Founders' background and use of the money
According to the company, Farlow built an annotation platform at Yutori to support training computer-use models. Putta has done research in computer vision, robotics and reinforcement learning, and worked on training computer-use models at MultiOn.
Martignetti said the founders have "developed a unique way to create high fidelity environments at scale," and described Plato as "really a data lab" working at the frontier of the field.
Plato plans to use the funding to expand its library of real-world work experience. It will pay professionals to recreate their jobs in simulated environments and deploy AI agents on real tasks to collect additional experience.
What it means for AI developers
For AI labs, the offer is a contained place to train and evaluate agents on job-like tasks in areas such as finance and healthcare administration, without risk to real money or records. The emphasis on scoring how a task is completed, and not only whether it is completed, targets a known weakness of reward-based training, where models can chase high scores in ways that do not reflect correct work. The source does not name customers or give performance figures, so the practical impact remains to be shown.
