Teaching a model to answer questions is easy. Teaching it to place an order without blowing the budget is not, and Plato just raised $16M to build the rehearsal space for that second problem.
Nava Ventures led the Series A, with participation from Gradient Ventures, Transpose Platform and Microsoft AI chief executive Mustafa Suleyman. Nava founder Freddie Martignetti joins the board, and the round takes Plato’s total funding to $19M after a Gradient-led seed.
Why labs need sandboxes
Training models directly in live software carries real consequences. A model learning to place an order could spend actual money, and one practising an administrative task could alter real records. Plato recreates those workflows in simulations that reset and reuse, so labs train without touching production.
The company works with domain experts to rebuild the tools, situations and expectations behind jobs in customer support, people operations, finance, accounting and healthcare administration.
Grading the how, not just the what
Co-founder and chief technology officer Pranav Putta argues that feedback shapes what a model learns, so Plato evaluates both the outcome and the path taken. Completing a purchase should not count as success if the model exceeded its budget, and the evals check whether a model exploited a weakness in the simulation to score well.
Chief executive Robert Farlow’s version is that mastering complex work comes from experience, and simulations compress years of it into safe, contained environments.
A new front in the training-data race
Every lab chasing reliable agentic behaviour is fighting the same bottleneck: clean, verifiable task environments that reward genuine competence. Synthetic benchmarks stopped convincing anyone once models learned to game them.
Plato’s next step is longer work sequences with simulated colleagues. It has $19M to do it, against competitors inside the labs building the same sandboxes in-house.
