Of all the announcements at KN26, Project Arc is the one that most directly challenges what we consider the boundary of automation. Previous automation tools — RPA, Flow Designer, IntegrationHub — required you to define the process upfront. Project Arc does not. It can figure out the steps itself.
What is Project Arc?
Project Arc is an enterprise autonomous desktop agent, jointly developed by ServiceNow and NVIDIA, that lives on employee desktops and can autonomously complete complex multi-step work across enterprise tools and systems.
The key differentiator from existing automation: Project Arc does not require pre-built workflows. It can navigate desktop applications, write and execute code, read and process documents, interact with web interfaces, and chain together actions across multiple systems — adapting in real time when things do not go as expected.
It is currently available as an early preview.
The Three Components of Project Arc
1. The Agent Itself
The core of Project Arc is an AI agent capable of computer use — the same capability that enables AI models to navigate GUIs, click buttons, fill forms, read screen content, and interact with desktop applications the way a human would. Applied to an enterprise context, this means the agent can work in any application an employee can work in, without requiring that application to have an API or a ServiceNow integration.
The agent thinks, writes code when needed, executes actions, monitors outcomes, and adapts when a step fails or produces an unexpected result. It handles multi-step processes across applications that were never designed to work together.
2. NVIDIA OpenShell — The Security Runtime
Every action Project Arc takes runs inside NVIDIA OpenShell, a sandboxed runtime environment. OpenShell provides:
- Containment — the agent operates inside a defined boundary; it cannot reach beyond the applications and data sources it has been granted access to
- Policy-based management — enterprise security policies are enforced at the runtime level, not just at the application level
- Auditability — every action taken inside OpenShell is logged, including files read, commands executed, and APIs called
- Sandboxing — even if the agent produces unexpected or anomalous behaviour, the sandbox limits the blast radius
The OpenShell runtime is the reason enterprise security teams can approve Project Arc deployments. Without a sandboxed environment, an agent with computer use capabilities represents significant security risk. OpenShell addresses that directly at the infrastructure level.
3. ServiceNow AI Control Tower — The Governance Layer
AI Control Tower governs what Project Arc can do. It sets policies, monitors runtime behaviour, and logs every action the agent takes. When the agent attempts an action that falls outside its governance policy, AI Control Tower blocks it.
Project Arc is also powered by Action Fabric and grounded in the ServiceNow CMDB, giving it enterprise-level intelligence about the environment it is operating in — which systems are available, which users have which access levels, which workflows are in place for which types of requests. That context enables the agent to make better decisions about how to complete its assigned tasks. — though access to that intelligence still flows through standard ACL security rules.
What Can Project Arc Actually Do?
ServiceNow described Project Arc's capability as completing complex multi-step work across enterprise tools without pre-built workflows. In practice, that covers tasks like:
- Researching information across multiple internal systems and compiling it into a report
- Processing a batch of employee onboarding tasks across HR, IT, and facilities systems that are not natively integrated
- Investigating an incident by pulling data from monitoring tools, CMDB, and change records, then drafting a root cause summary
- Executing a software deployment by navigating a deployment pipeline, checking approvals, running tests, and publishing results
- Processing legacy application data that has no API by navigating the UI directly
The CMDB grounding is important here. Because Project Arc knows the enterprise environment through the CMDB, it can make contextually appropriate decisions. When it needs to access a system, it knows which system it is, who owns it, what access controls apply, and what the approved processes are.
The NOWAI-Bench Announcement
Alongside Project Arc, ServiceNow and NVIDIA announced NOWAI-Bench — an open benchmarking standard for AI agents in enterprise environments. It includes EnterpriseOps-Gym and EVA-Bench, released as open source.
The benchmark matters because enterprise AI evaluation is still immature. Most AI benchmarks test model capabilities in isolation, not agent performance in real enterprise workflow environments. NOWAI-Bench is designed to fill that gap — giving organisations a standardised way to measure how well AI agents actually perform on enterprise tasks before deploying them in production.
AI Control Tower Integration with NVIDIA Enterprise AI Factory
Alongside Project Arc, ServiceNow announced that AI Control Tower is now integrated with the NVIDIA Enterprise AI Factory validated design. This extends governance to large-scale model workloads at data centre scale — not just desktop agents, but the model infrastructure behind them.
The significance: AI governance that starts at the desktop agent and extends through the AI infrastructure stack to the model training and inference layer. A single governance framework from the endpoint to the data centre.
What Project Arc Means for ServiceNow Professionals
Project Arc is not a replacement for ServiceNow expertise — it is an expansion of what ServiceNow can automate. The difference is important.
Current automation tools require a ServiceNow professional to design and build the integration, the flow, or the catalog item before the automation can run. Project Arc can automate tasks in applications that have no integration and no pre-built flow. But someone still needs to define what tasks the agent should perform, configure the governance policies that control it, review and approve its actions in contexts that require human oversight, and maintain the CMDB data that grounds its decisions.
The skills that become more valuable as Project Arc adoption grows: CMDB management and data quality, AI governance policy design, integration architecture, and the ability to evaluate AI agent outputs critically. Knowing when an agent's action is correct, when it needs to be reviewed, and when a process is too sensitive to delegate to an agent is human judgement that the technology does not replace.
Availability
Project Arc is available as an early preview as of Knowledge 2026. Check the ServiceNow Australia release notes and the NVIDIA partnership announcement for the current access and preview programme details. The AI Control Tower integration with NVIDIA Enterprise AI Factory validated design is generally available.
Key Takeaways
- Project Arc is an enterprise autonomous desktop agent that completes multi-step work without pre-built workflows
- It runs inside NVIDIA OpenShell — a sandboxed, auditable, policy-controlled runtime environment
- ServiceNow AI Control Tower governs its actions; CMDB provides enterprise context
- It is powered by Action Fabric — giving it access to the ServiceNow system of action
- Currently available as an early preview; full GA timeline pending
- NOWAI-Bench is an open benchmarking standard released alongside Project Arc