Blog — The Vested Group

What to Put in a NetSuite AI Governance Policy

Written by Joseph Lang | Nov 19, 2026
Written by • Director of inVESTED PRO • 13+ years of NetSuite experience

A NetSuite AI governance policy should say who can authorize an AI connection, what the AI is allowed to do and not just what it can see, and who is accountable when it is wrong. Policies often focus on what employees should not paste into a chatbot. That misses the connection itself and the actions it can take, which are the parts NetSuite lets you control.

Key Takeaways

  • A policy about employee prompts is not enough. It needs to govern the connection and the actions.
  • Name who can approve an AI client, MCP server, or tool, and keep a list of what is approved.
  • Separate query, create, modify, approve, and pay. They carry very different risk.
  • Accountability means naming who reviews which actions, what is logged, and what happens when something unexpected occurs.
  • Write down how to turn AI access off quickly. You will want that page the first time something looks wrong.

Part of the NetSuite Next, Without the Nerves series.

Last reviewed October 5, 2026 against Oracle's published NetSuite Next documentation. NetSuite Next is changing often, so check Oracle's pages for the current state.

A typical AI policy is about people. Do not paste confidential data into a public chatbot. Do not use unapproved tools. That is fine as far as it goes. It does not govern what happens when an AI client is connected to your ERP and acts on its own.

What does a NetSuite AI policy need to cover?

Oracle does not publish a policy template, but its risks, controls, and mitigation document reads like an outline for one. The table below turns its guidance into policy sections. Treat it as a starting structure and have your own counsel and security team review the result. This post is not legal advice.

Policy section What it should say Where Oracle's guidance backs it
Approval Who can authorize an AI client, MCP server, or tool to connect, and the list of approved ones Oracle advises an internal approval process and an approved list, with no unapproved servers or tools
Permitted actions Which actions are allowed: query, create, modify, approve, pay. Each level named separately Oracle lists high-impact actions and advises read-only or low-impact tools first
Access Which roles may hold MCP permission, built per use case, reviewed on a schedule Oracle advises separate roles, no high-privilege users, and periodic review
Human review Which actions need confirmation, who confirms, and what the confirmation shows Oracle prefers clients that require explicit confirmation for high-impact actions
Logging and review What is logged, who reads it, how often, and what counts as unusual Oracle says all MCP tool usage is logged and advises regular review
Incident response How to disable access fast: remove permissions, revoke consent, disable tools Oracle advises documented procedures for quickly disabling MCP access
Training What users are told about AI risks, hidden instructions, and unapproved tools Oracle advises training users on risks and safe usage
Compliance Which data and use cases are excluded, such as ePHI Oracle states the connector is not assessed for HIPAA and points to local regulations

Who can authorize an AI connection?

This is the first gap. Without an answer, any employee with the right permission can connect whatever AI client, MCP server, or tool they prefer. Oracle's guidance is direct: establish an approval process for AI agents, MCP clients, MCP servers, and MCP tools before they connect, keep an internal list of approved servers and tools, and do not let users connect arbitrary ones without administrative approval.

In policy terms, name the role or committee that approves, state what they check, and keep the list somewhere people can find it. Oracle's own checklist for evaluation is useful: whether the vendor addresses prompt injection and hallucination, whether the client shows the full tool name, description, and parameters before it runs, and whether it uses secure authentication with short-lived tokens and narrow scopes.

What is the AI allowed to do?

This is the second gap, and the more common one. Policies describe what data AI may see and say little about what it may do. Reading a vendor balance and changing a vendor record are very different risks.

Write the levels out: query, create, modify, approve, pay. Say which are allowed for which roles and tools. Oracle's list of high-impact actions includes creating, modifying, deleting, approving, paying, exporting large data sets, changing permissions, and accessing new sensitive resources, and it advises preferring clients that require explicit user confirmation for them. Start with read-only or low-impact tools, and expand only after reviewing behavior, logs, and business need.

Who is accountable when the AI is wrong?

This is the third gap. Saying a human is in the loop does not tell anyone what to do. Oracle's FAQ warns that AI clients can still produce incorrect or hallucinated results, so results should be verified against NetSuite data. Its risk document lists hallucination, prompt injection, unintended actions, data corruption, and disclosure of sensitive information, and says prompt injection and hallucination are outside NetSuite's control.

A policy that handles this answers four questions. Which actions require review before they run? Who reviews them? What gets logged? What happens when something unexpected occurs, and who owns the response? Oracle's FAQ says the execution log keeps 21 days in production, so state how often the log is read and by whom, inside that window.

How do you turn it off fast?

Most policies skip this, and it is the page you will want first. Oracle advises documenting procedures for quickly disabling MCP access, removing MCP permissions, revoking AI agent consent, disabling specific tools, or suspending access to third-party MCP servers. Write the steps, name who is allowed to take them, and test them once before you need them.

What about data and compliance limits?

State which data and use cases are out of bounds. Oracle's FAQ says the AI Connector has not been assessed for HIPAA compliance and should not be used for electronic protected health information unless you have independently decided that use meets your own obligations. It also notes that regulations vary by region, with specific requirements for use cases such as HR and financial processes. Ask your counsel which of those apply to you.

One more line belongs here. Data sent to an outside AI provider is governed by that provider's privacy policy, per Oracle. Your policy should say which providers are approved for which data.

Does AI usage need cost governance too?

It can. Oracle's AI Units FAQ says administrators can monitor AI Units usage on the Billing Information page and that, when units run out, customers are expected to buy more or pause AI feature use. A policy can say who watches usage, at what threshold they act, and who may approve extra purchases. The AI Connector does not consume AI Units, per Oracle, so cost governance there means the outside AI plan.

Joe's take: govern the connection and the actions

Three things stand out to me that most templates leave out. First, who can authorize an AI connection to NetSuite. Employees shouldn't be able to connect whatever AI client, MCP server, or tool they happen to prefer. There needs to be an approval process.

Second, what the AI is allowed to do, not just what data it can see. Reading a vendor balance and changing a vendor record are very different risk levels, so a policy has to distinguish between query, creation, modification, approval, and other actions.

Third, who is accountable when the AI is wrong. Human in the loop isn't enough. Which actions require human review? Who reviews them? What gets logged? What happens when something unexpected occurs? Oracle's own guidance calls out hallucination, prompt injection, unintended actions, data corruption, and disclosure of sensitive information. Those aren't solved by a policy that simply says employees shouldn't put confidential information into AI. The policy needs to govern the connection and the actions, not just employee prompts.

Frequently Asked Questions

What should a NetSuite AI governance policy include?

At minimum: who can approve AI connections, which actions the AI may take, which roles and tools are permitted, which actions need human review, what is logged and who reads it, how to disable access quickly, and which data is off limits. Oracle's risk guidance covers each of these areas and works as an outline.

Who should approve AI connections to NetSuite?

Someone with authority over security and finance systems, named in the policy. Oracle advises an internal approval process for AI agents, MCP clients, servers, and tools, with a list of approved ones and no arbitrary connections. Whoever approves should check how the vendor handles prompt injection, hallucination, authentication, and what the client shows before it runs a tool.

What should an AI policy say about what the AI can do?

It should separate query, create, modify, approve, and pay, and state which are allowed for which roles and tools. Reading a balance and changing a record carry different risk. Oracle recommends preferring read-only or low-impact tools first and clients that require explicit confirmation for high-impact actions.

Is a human in the loop enough for AI in NetSuite?

Not by itself. A policy has to say which actions need review, who reviews them, what gets logged, and what happens when something unexpected occurs. Oracle notes AI clients can still hallucinate, so results should be checked against NetSuite data, and it advises confirmation prompts for creating, modifying, deleting, approving, and paying.

How do I disable AI access in NetSuite quickly?

Write the procedure in advance. Oracle advises documenting steps for removing MCP permissions, revoking AI agent consent, disabling specific tools, or suspending access to third-party MCP servers. Name who may act and test the steps once. NetSuite has no single account-level switch to disable all AI features, so the policy should cover each control.

Does the NetSuite AI Connector meet HIPAA requirements?

Oracle says the AI Connector has not been assessed for HIPAA compliance and should not be used to store, process, or transmit electronic protected health information unless you have independently decided that use is consistent with your own obligations and applicable law. Ask your counsel before connecting any workflow that touches health data.

Want a policy that matches how NetSuite really works?

The strongest policies are short, specific, and tied to controls you can actually switch on. A page that names the approver, the permitted actions, and the shutoff steps will serve you better than ten pages of principles.

TVG can help turn Oracle's guidance into a policy for your account and set up the roles, logging, and review that back it. See our AI services, ask about a NetSuite Health Assessment, or start with the NetSuite Next guide.

Work With The Vested Group

Ready to Get More From NetSuite?

Whether you are optimizing your current environment, planning your next project, or looking for a partner who does more than close tickets, our NetSuite team is ready to help.

About the Author

Joseph Lang is the Director of inVESTED PRO at The Vested Group. With more than 13 years of NetSuite experience, he helps organizations maximize the value of their ERP investment through strategic guidance, operational improvements, and ongoing managed services. Joseph specializes in NetSuite optimization, manufacturing, warehouse management, inventory management, procurement, and SuiteCommerce, helping clients continuously improve their business processes and long-term ERP performance.