ServiceNow and Nvidia Launch Arc: The Long-Running AI Agent That Could Reshape Enterprise Work

ServiceNow and Nvidia have launched Arc, a personal, long-running enterprise AI agent. Its job is to handle productivity tasks across email and Teams. Arc is built for RTX Spark and Nvidia OpenShell. On the surface, that is a product launch. In practice, it is a clear signal about where AI at work is going next.

The next wave of enterprise AI is not just a smarter chatbot. It is an agent that stays with a worker, keeps context, and takes action inside the tools people already use every day. That changes how work gets done, how teams are managed, and how companies measure the value of AI.

What Arc Is, and Why Every Word Matters

The description of Arc is short, but each word carries weight: personal, long-running, enterprise, and productivity tasks across email and Teams. Together, they describe a different kind of AI than the one most workers have used so far.

Personal means it is built around one worker

A personal agent is not a shared help desk bot. It is designed around an individual employee’s work. That raises the promise of less repetitive effort and faster answers. It also raises hard questions about privacy, data boundaries, and what happens when an agent can see a person’s inbox or chat history.

Long-running means it does not stop after one answer

Most AI tools today work in short bursts. You ask a question, you get a reply, and the session ends. A long-running agent is different. It is meant to continue work across steps and sessions. It can pick up a task, keep track of what happened, and carry context forward. That is the difference between a search box and a coworker.

Enterprise means it has to be governed

Consumer AI can get away with rough edges. Enterprise AI cannot. An agent that touches email and Teams needs clear permissions, audit trails, and a way to stop bad actions. It also needs to fit into company security rules. The word enterprise is not decoration. It is the hardest part of the product.

Email and Teams are the front door to work

The stated scope is productivity tasks across email and Teams. Those are not niche tools. They are where messages, decisions, requests, and follow-ups live. An agent that works there sits close to the center of daily work. The obvious use cases are inbox triage, thread summaries, draft replies, meeting follow-ups, and routing requests to the right person. Arc may not do all of that on day one, but the launch points in that direction.

The Big Shift: From Answers to Delegation

The first era of enterprise AI was about answers. Workers asked a model to write a sentence, explain a concept, or summarize a document. That was useful, but it was still a tool you picked up and put down.

Arc points to the next era: delegation. Instead of asking for a paragraph, a worker can hand off a task. Instead of copying text into a chat window, the agent can work inside the flow of email and Teams. The human moves from doing every step to directing, reviewing, and approving.

That shift sounds small. It is not. Delegation requires trust. A worker must believe the agent understands the goal, has the right access, and will not embarrass them in front of a customer. A company must believe the agent will follow policy and leave a record. Those are not model problems alone. They are product, security, and management problems.

Why the Nvidia Stack Matters

Arc is built for RTX Spark and Nvidia OpenShell. Those names matter because long-running agents need more than a language model. They need compute, a runtime, and a way to connect to enterprise systems. The exact technical details will matter to buyers, but the strategic point is clear: agentic AI is becoming an infrastructure story.

The pairing of ServiceNow and Nvidia is also a signal. One company brings enterprise software context. The other brings AI computing. The bet is that agents will be sold as part of a stack, not as a standalone app. That stack has to handle identity, tools, data, and scale. Companies that can put those pieces together will have an edge.

For technical teams, this is the moment to think about where agents will run. Will they sit in the cloud, on a local device, or in a hybrid setup? Each choice changes cost, latency, privacy, and control. A personal agent that handles email and Teams may need local access for speed and context, but central governance for safety. The right answer will depend on the workflow.

The Technical View: What a Long-Running Agent Needs

Any agent that runs for a long time and touches company communication needs a strong foundation. Arc will be judged on these basics:

• Identity and permissions. The agent must act as the right user, with the right access. It should not see more than it needs.
• Memory and context boundaries. It must remember useful context without leaking sensitive information from one task or person to another.
• Tool use and audit logs. It must show what it did, when, and why. A log is not optional when an agent can send messages or change records.
• Human approval. High-risk actions should require a human check. The agent should know when to ask.
• Cost controls. Long-running agents can burn compute. Teams need limits, routing, and visibility into spend.
• Security against prompt injection. Email and chat are full of untrusted text. An agent must not be tricked into ignoring rules or leaking data.

These are not nice-to-haves. They are the price of entry for enterprise agents. A demo can skip them. A product that runs inside real companies cannot.

Business Implications: Small Savings, Large Scale

Email and Teams are high-volume tools. A small time saving on each message or thread can add up across a large workforce. That is the business case for Arc. But the return on investment will not come from licenses alone. It will come from changed workflows.

The first wins will likely be narrow and repeatable. A sales team might use an agent to prepare account summaries from email threads. A support team might use it to draft replies and route cases. A project team might use it to turn chat into clear action items. Each use case needs a clear owner, a success metric, and a review step.

Managers should measure more than time saved. They should track quality, rework, risk events, and adoption. An agent that saves ten minutes but creates a compliance problem is not a win. An agent that saves five minutes and improves consistency may be. The metrics have to match the risk.

There is also a people story. Workers will need to learn how to delegate to an agent. That means writing clear requests, setting boundaries, and checking output. It is a new skill, not a replacement for judgment. Companies that train for it will get more value than companies that simply hand out access.

Risks and Guardrails

The launch of Arc also makes the risks of personal enterprise agents more concrete. An agent with access to email and Teams is a high-value target. It sees messages, names, plans, and maybe sensitive data. That calls for serious guardrails.

• Privacy. Workers need to know what the agent can see and how long it keeps data. Consent and transparency matter.
• Security. Access should follow least privilege. The agent should not have broad rights just because it is convenient.
• Accuracy. A confident wrong answer in a customer email is costly. Human review should match the risk.
• Accountability. If an agent sends the wrong message, who owns the outcome? Companies need a clear answer before scale.
• Overreliance. Workers may stop checking. Training and design should keep humans in the loop.
• Lock-in. Agents tied to one stack can be hard to move. Buyers should ask about openness and data portability.

These risks do not mean companies should wait. They mean companies should pilot with eyes open. Start small. Limit access. Log actions. Review results. Then expand.

Actionable Insights for Leaders

For business leaders, the Arc launch is a prompt to act, not just a product to watch. Here is a practical path:

• Pick one workflow. Choose a high-volume task in email or Teams that has clear inputs and outputs. Avoid starting with the most sensitive work.
• Map data and permissions. Know what the agent needs to see and what it must never see. Set limits before the pilot.
• Run a small pilot. Use a group that can give feedback. Compare results with a control group if possible.
• Define success. Use metrics for time, quality, risk, and adoption. Do not rely on vendor demos.
• Set guardrails. Require approval for high-risk actions. Keep logs. Test for prompt injection and data leaks.
• Train workers. Teach people how to delegate, review, and correct the agent. Make it safe to report problems.
• Review and scale. Expand only when the workflow is stable and the risk is understood.

For technical teams, the message is to design for the long run. Build identity, logging, and tool permissions into the agent from the start. Treat context as a security boundary. Plan for cost and latency. The agents that win will be the ones that are trusted, not just the ones that are clever.

The Future of AI at Work Is Persistent, Personal, and Governed

Arc is a sign of where enterprise AI is headed. The future is not a single model that answers every question. It is a set of personal agents that work alongside people, across tools, over time. These agents will handle routine work, surface what matters, and help teams move faster. They will also force companies to rethink permission, privacy, and accountability.

The winners will not be the companies that buy the most agents. They will be the companies that redesign work around them. That means clear tasks, clean data, strong guardrails, and people who know how to direct and check the machine. ServiceNow and Nvidia have put a marker down. The next question is whether enterprises are ready to manage a workforce that includes persistent AI.