ServiceNow and Nvidia launch Arc AI agent for workplace tasks

Autonomous workplace planning

ServiceNow and Nvidia have launched Arc, a personal enterprise AI agent designed to handle extended workplace assignments. Unlike simple chatbots, Arc is built to take over tasks that might span several days or involve multiple different applications. An employee can provide a high-level goal, and the agent develops a plan to execute it independently.

The agent identifies unfinished work by scanning an employee's inbox, task backlog, and meeting conversations. It can then coordinate automated workflows across various productivity tools, including Microsoft Teams and email. If a specific connector is missing for a required task, the software is capable of writing a new integration to communicate with the necessary application.

To ensure accuracy, Arc is designed to check its own progress and adjust its actions as it moves through a plan. This allows it to handle recurring assignments by developing memory and skills over time. It aims to reduce the need for humans to supervise every individual step of a multi-step business process.

Hardware and model scaling

The system uses a tiered approach to processing power and model size to balance cost and performance. For everyday workloads, Arc can run on smaller local models using Nvidia RTX Spark hardware. This allows for faster and potentially cheaper processing of routine tasks that do not require massive computing power.

When a request is more demanding or complex, the agent can switch to frontier models to handle the heavy lifting. This flexibility is intended to optimize how the software handles different types of requests based on their difficulty. The tool can operate directly on an employee's desktop or on a separate computer located within a secure environment.

ServiceNow is including Arc in several of its existing subscription tiers, such as EmployeeWorks Foundation, Advanced, and Prime. While the tool is currently available on Mac, the company expects to bring Windows support to the platform within a few weeks. This rollout makes the agent accessible to a wide range of enterprise users quickly.

Security and permission layers

Security is a primary concern for enterprises allowing AI to access internal data and perform actions. Administrators can define specific permissions for Arc through a tool called the Super Intelligence Control Tower. This allows IT leaders to specify exactly which applications and data sets the agent is allowed to interact with.

Nvidia provides a sandboxed environment called OpenShell to enforce these restrictions during the agent's execution. This ensures that even if an agent is performing a task, it cannot step outside of the boundaries set by the company. Any integrations created by the assistant must also operate within these same security limits.

Looking ahead, ServiceNow plans to introduce AI Passport, which acts as an authorization record for every agent action. This will eventually allow the software and hardware layers to recognize when permissions are revoked. While not part of the current release, this will provide a more dynamic way to stop unauthorized activity in real time.

Benchmarking and performance

Research from ServiceNow's March EnterpriseOps-Gym study highlights how planning affects AI performance. The study examined 1,150 simulated workplace tasks and found that providing models with expert-written plans significantly improved success rates. In one experiment, a smaller four-billion-parameter model outperformed a much larger 30-billion-parameter model when given these plans.

The data showed that human-written plans improved task completion by approximately 14 to 35 percentage points compared to models working without them. However, the researchers noted that plans generated by other AI models provided smaller improvements than those written by humans. This suggests that the quality of the initial instructions remains a critical factor for success.

For businesses, these findings mean that using smaller models locally can reduce costs, but only if the workflow is well-defined. If a workflow requires frequent human corrections, it may still consume significant staff time regardless of the model size. Experts suggest that companies should test Arc on familiar assignments to compare the agent's planned actions against established human procedures.