AI agents are getting ready for the office, but first they need a safe place to learn how to do the actual work.
San Francisco-based startup Plato has raised $16 million in Series A funding to build simulated workplaces where AI models can practice professional tasks before interacting with real business systems.
Nava Ventures led the round, with participation from Gradient Ventures, Transpose Platform and Microsoft AI CEO Mustafa Suleyman. The investment brings Plato's total funding to $19 million following an earlier seed round led by Gradient Ventures. Freddie Martignetti, founder of Nava Ventures, will join Plato's board.
AI agents need more than chatbot training
Most AI models are trained to generate responses, retrieve information or follow instructions. However, building agents that can independently complete business tasks introduces a different challenge.
An agent handling a customer request, processing an invoice, or managing an employee record must navigate software, make decisions, and respond to unexpected situations without breaking established rules.
Training directly inside business applications can create expensive problems. An AI agent practicing how to place an order could spend actual money, while one learning administrative workflows could accidentally modify company records.
Plato addresses this by recreating workplace software and processes in controlled environments that AI developers can reset and reuse. The company works with professionals to reproduce workflows across customer support, people operations, finance, accounting and healthcare administration.
These environments support reinforcement learning, allowing AI models to attempt tasks, receive feedback and improve through repeated practice.
Completing a task isn't necessarily doing it correctly
Plato is also tackling a problem that could become increasingly important as companies give AI agents access to sensitive business operations: measuring whether an agent actually follows instructions.
For instance, an agent that successfully completes a purchase but exceeds its authorized budget should not receive a passing score simply because the transaction went through. The company evaluates both the outcome and the decisions behind it, including whether models respect operational constraints or exploit weaknesses in simulations to achieve higher scores.
"The feedback a model receives shapes what it learns," said Pranav Putta, Plato's co-founder and chief technology officer.
The approach could help AI developers distinguish between agents that merely find shortcuts and those that can follow the rules governing actual business processes.
A growing market for AI training infrastructure
Plato was founded in 2025 by Robert Farlow and Putta, who bring experience in computer-use model training, robotics, computer vision and reinforcement learning. The company plans to use its new funding to expand its simulation library, pay professionals to recreate their jobs, and deploy AI agents on real tasks to gather additional experience.
It is entering a market attracting other well-funded players. AI training company micro1 has announced plans to spend $1 billion on enterprise data for agent-training environments, while Patronus AI has raised $50 million to test AI agents in simulated settings.
However, building realistic workplace simulations is not simply a matter of copying software interfaces. Real business operations involve incomplete information, changing priorities, human judgment, and exceptions that are difficult to reproduce.
What eWEEK found: AI's next infrastructure battle could be about real-world experience
Plato's $16 million funding round highlights an emerging challenge in enterprise AI: Building increasingly powerful models is only part of the equation. Those models also need practical experience navigating the complicated software, rules, and decisions that define everyday business operations.
As companies move beyond chatbots toward AI agents capable of executing entire workflows, realistic training environments could become an increasingly important part of AI development. Just as autonomous vehicles benefit from simulated driving scenarios, enterprise AI agents may need thousands of opportunities to practice handling invoices, resolving customer complaints, and managing administrative tasks before they can be trusted with real business systems.
For enterprises, the potential benefits extend beyond better AI performance. Agents trained to recognize spending limits, follow approval processes, and handle unexpected situations could reduce costly automation errors. More realistic simulations could also help organizations evaluate AI systems before granting them access to sensitive financial records, customer information, or operational tools.
The opportunity also extends to the companies building AI models. If simulated workplaces can produce measurable improvements in agent reliability, developers may increasingly compete on the quality of their training environments and feedback systems, alongside model size, computing power, and training data.
However, significant questions remain. Simulated environments cannot fully reproduce the unpredictable nature of real workplaces, where employees make judgment calls, software changes without warning, and unusual situations frequently fall outside established procedures. An agent that performs well in training could still struggle when confronted with circumstances it has never encountered.
There is also a gap between demonstrating that an agent can complete a task and proving that it can do so consistently, securely, and in accordance with an organization's policies. Businesses will still need real-world testing and oversight before deploying autonomous systems across critical operations.
The bigger question is whether AI's next competitive advantage will come from giving models more knowledge or giving them better experience. Plato is betting that realistic workplace training will help close the gap between agents that perform well in demonstrations and those businesses can reliably put to work.
Related reading: A recent eWEEK report on AI agent performance examines how a leading AI model successfully completed just 35% of enterprise tasks in benchmark testing, highlighting the reliability challenges Plato hopes to address.