ServiceNow and NVIDIA announced Project Arc at Knowledge 2026 this week — an enterprise autonomous desktop agent that lives on employee computers and completes complex, multi-step work without a human at the keyboard. The announcement landed alongside a broader expansion of ServiceNow’s AI Control Tower and a deepened integration with Microsoft Agent 365, making this one of the more consequential enterprise AI releases of the year.

Here is what it means for businesses trying to figure out how to deploy AI responsibly at scale.

What Project Arc Actually Does

Project Arc is not a chatbot with extra steps. It is an autonomous agent that runs on an employee’s desktop, accesses their actual tools and systems, and completes full workflows without waiting for step-by-step instruction. It can write code, execute it, adapt when something breaks, read files, call APIs, and navigate enterprise software — all in sequence, without requiring pre-built workflow templates.

The practical difference from existing AI assistants is meaningful. Most AI tools today generate text that a human then acts on. Project Arc acts directly. An employee could tell it to pull the latest contract data from Workday, cross-reference it against a SharePoint folder, and draft a summary report — and then walk away while it finishes.

NVIDIA’s role is the security and containment layer. Every action Project Arc takes runs inside NVIDIA OpenShell, a sandboxed runtime that enforces policy-based management on autonomous activity. Files read, commands executed, and APIs called are all logged. If the agent tries to do something outside its permissions, the runtime blocks it.

ServiceNow’s AI Control Tower sits above that, providing the governance layer: setting policies, monitoring behavior, and giving IT and compliance teams full visibility into what autonomous desktop agents are doing across the organization.

Governance Expands from Desktops to Data Centers

The desktop agent is only half of the story. ServiceNow and NVIDIA also completed an integration that was previewed at NVIDIA GTC in March: AI Control Tower now works inside the NVIDIA Enterprise AI Factory validated design, extending governance to the data center infrastructure layer.

What this means in practice: organizations running large AI model workloads at scale can now apply the same governance framework to their infrastructure-level AI as they apply to desktop agents and employee-facing tools. One control plane, across the full stack, from individual desktops up to data center model deployments.

This matters because most enterprise AI deployments today are fragmented. A business might have one vendor’s agent running in HR, another in IT, a third in customer service, and a separate model infrastructure team handling compute — all with different governance approaches, or none at all. ServiceNow is positioning AI Control Tower as the unifying layer that spans all of it.

Open Benchmarking for Enterprise AI Agents

ServiceNow and NVIDIA also released NOWAI-Bench, an open benchmarking suite for enterprise AI agents, with two frameworks:

EnterpriseOps-Gym evaluates agents across multi-step workflows in IT service management, customer service, and HR. These are the actual business scenarios enterprises care about — not abstract reasoning tasks, but whether an agent can handle a real IT support ticket from start to finish without human intervention.

EVA-Bench is a voice agent evaluation framework designed for enterprise settings, testing how voice-enabled AI agents perform in business communication contexts.

Both frameworks are open-source and NVIDIA is integrating them into its NeMo Gym platform to make agent evaluation reusable across the industry. This is a meaningful move toward standardized ways to compare and audit enterprise AI agents — something that has been largely absent until now.

What This Means for Business

The enterprise AI conversation has been stuck on a specific set of questions for the past two years: which model is best, how do we prompt it, what can it answer. Project Arc signals that the conversation is shifting to a different and harder set of questions: how do we give AI agents access to real systems, how do we contain what they can do, and how do we audit what they have done.

These are questions every business will need to answer as AI moves from assistant to autonomous worker.

A few practical implications:

Governance is now a product requirement, not an afterthought. The NVIDIA OpenShell plus ServiceNow AI Control Tower combination shows that the enterprise market is moving toward treating containment and auditability as non-negotiable requirements, not optional add-ons. Any AI deployment that cannot tell you exactly what it did, and prevent it from acting outside defined boundaries, is going to face increasing resistance from IT and legal teams.

Desktop AI agents are the next infrastructure layer. Most organizations think of AI infrastructure as cloud compute and API calls. Project Arc suggests the endpoint — the employee’s actual computer — is becoming part of the AI deployment architecture. That has implications for device management, endpoint security, and how IT teams think about what constitutes a managed device.

Open benchmarking will accelerate enterprise trust. NOWAI-Bench giving enterprises standardized ways to test agents against real business workflows before deployment is a practical step toward making AI adoption less of a faith exercise. If you can run an agent through EnterpriseOps-Gym and see how it handles your actual IT service management scenarios, you have something more concrete than a vendor’s claims.

For businesses building their AI strategy right now, the ServiceNow and NVIDIA partnership represents a useful model: combine a strong execution layer (the agent doing the work) with a strong governance layer (the system watching and containing it), and extend that framework consistently across every environment where AI operates. That combination of capability and control is what enterprise AI adoption actually requires.

Project Arc is available as an early preview. AI Control Tower enhancements are entering ServiceNow’s Innovation Lab in May, with general availability expected in August 2026.