| AI data security is the set of rules, settings and habits that protect business information when staff use AI tools. It covers which tools and accounts are approved, what information can go into them, how providers use and keep that information, and what connected AI assistants can reach. |
Staff use AI to draft emails, summarise documents and tidy up spreadsheets. The useful AI data security question for a business is not whether AI is safe in general, but what information each person can put into which tool, using which account.
This AI data security guide answers that question. It covers a simple set of sharing rules, how business and personal AI plans handle data, what to check before approving a tool, why permissions matter when AI connects to your files, what to do if confidential information is uploaded, and how the Privacy Act applies. The aim is to let your team use AI with clear boundaries rather than to block it.
What is AI data security?
AI data security applies familiar information security ideas to a new kind of tool. Cloud services already process business data outside your office, so the change is not that information leaves your network. The change is that AI tools can retrieve, combine, summarise and act on information in new ways, sometimes across every file a person can open.
That creates three practical questions. What information is entering the tool? What does the provider do with it? And what else can the tool reach once it is connected to your email, files or other apps? Good AI data security answers each one for the tools your staff use.
The National Cyber Security Centre’s AI guide for small businesses lists data leaks and privacy breaches among the main risks, alongside unreliable or manipulated outputs and supply chain weaknesses. This guide focuses on the first of those: keeping business information under control.
Which AI data security risks should a business plan for?
The risks are practical rather than exotic. Each one has a clear control, which is why a short set of rules covers a lot of ground.
- Oversharing in prompts. Staff paste more than the task needs, such as a whole client list to draft one email. Sharing rules and minimal inputs address this.
- Personal accounts used for work. Work information ends up in an account the business cannot manage or recover. Approved work accounts address this.
- Overshared files. An assistant connected to your files can surface anything a person can open. Permission reviews address this.
- Connected tools acting on untrusted content. An assistant that reads email or websites can be misled by instructions hidden in that content. Limited permissions and human approval address this.
- Unchecked outputs. AI can produce confident but wrong answers, or include information that should not be shared. A human review before use addresses this.
Treat AI data security as part of your existing security programme rather than a separate project. The same identity controls, permissions and training that protect email and cloud files extend to AI tools, so much of the groundwork may already be in place.
What information can staff put into AI tools?
The starting point for AI data security is the approved work account: check what it is approved to handle. Permission to use an AI tool does not mean permission to upload every business document. Include only the information needed for the task.
| Information | Suggested rule |
|---|---|
| Public material | Use approved tools for published website copy or public product information. Check the output before publishing. |
| Internal working information | Use only an approved business workspace and permitted use case. Remove unnecessary names, identifiers and commercially sensitive detail. |
| Client files, payroll, health records or confidential contracts | Require explicit approval for the data and task, with privacy, retention, access and provider terms checked first. |
| Passwords, recovery codes, API keys or private keys | Keep these out of general-purpose AI chats. Use the approved secure process for managing secrets. |
These are suggested AI data security rules for business policy, not legal categories. Business.govt.nz recommends giving staff a clear reminder of what not to put into AI tools, and a short table like this one does that job.
Does removing names make information safe to share?
Not in every case. Removing a name does not necessarily anonymise a record. A job title, location, date or unusual event may still identify someone, especially in a small community. Use fictional examples where the real information is not necessary for the task.
Our guide to an AI acceptable use policy explains how to turn these AI data security rules into a written policy staff can follow.
How do AI providers handle business data?
Business subscriptions can provide useful contractual and administrative protections. Those protections depend on the product, plan, account, settings and connected services. Payment alone is not a security boundary, and AI data security depends on checking the exact product your staff use.
Business AI plans
Business AI plans can offer commitments about training use, account administration and data protection. Microsoft’s data, privacy and security documentation for Microsoft 365 Copilot says prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, and that Copilot only surfaces organisational data a user already has permission to view. The same documentation says prompts may be processed in other regions, so it does not mean every interaction stays in New Zealand.
Connected features can follow different rules. When web search is turned on, Copilot sends generated search queries to Bing, outside the Microsoft 365 service boundary, with user and tenant identifiers removed. Administrators can turn web search off. Review processing locations, retention, web search and agents before approving sensitive work, as part of your AI data security checks.
OpenAI states in its business privacy commitments that it does not train on business data from its business plans and API by default, with retention controls on some plans. Other providers offer their own terms, and the details change, so check the current terms for the plan you use.
Personal AI accounts
Some personal tools let users switch off model training. In ChatGPT, for example, users can turn off training in data controls, and temporary chats are deleted from OpenAI’s systems after 30 days. Those settings help, but they do not replace business approval, contractual protections or central administration. A personal account belongs to the person, not the business, and the business cannot manage its settings, sharing or access when that person leaves. For AI data security, approve the account as well as the tool.
Training, retention and access are separate AI data security questions
A no-training commitment does not mean no storage or no disclosure risk. Training use, retention, processing location, access, connected services and oversight are separate questions, and each one needs an answer before a tool is approved for sensitive work. Audit logging also varies: check what is recorded under your licence and configuration rather than assuming every prompt can be reviewed.
Is shadow AI a type of product?
No. Shadow AI describes approval status, not a product category. A paid business subscription that nobody approved is still shadow AI, while a free tool approved for public material may be acceptable. Our guide to shadow IT covers how to find and manage unapproved tools. Shadow AI is one of the AI data security risks an approval process is designed to catch.
What should you check before approving an AI tool?
Approval should cover a specific use, not just a brand name. Record the following for each tool, so your AI data security decisions are consistent and easy to revisit.
- Product and plan. The exact product, plan and work account staff will use.
- Permitted tasks and data. What the tool is approved for, and which types of information it may handle.
- Provider terms. Training use, retention, deletion and processing locations for that plan.
- Access and connections. Which permissions it inherits, which apps it connects to and whether web search is on.
- Oversight. Which logs are available under your licence and who reviews them.
- Ownership. A named owner, a review date and a route for staff to request another tool.
A quick, clear approval route matters. If requesting a tool takes weeks, staff find their own, and the approved list stops reflecting what is in use.
Why do permissions matter when AI connects to your files?
Permissions are where AI data security meets your existing Microsoft 365 settings. An AI assistant that searches your files works within each person’s existing access. If a folder is shared with the whole company, the assistant can find it for anyone who asks. Copilot does not create new access, but it can make information that is already overshared easier to find.
Review broad SharePoint access, sharing links and connected sources before rollout. A Microsoft 365 permissions audit finds overshared sites and stale access, and data loss prevention with sensitivity labels can limit what sensitive content an assistant can use. Detailed configuration belongs on our Microsoft Copilot page.
What changes when AI tools connect to email and apps?
An AI assistant connected to email or files can access more than the text you paste. This is a newer AI data security risk. Limit its permissions and check what external apps receive. Treat instructions found inside documents, messages and websites as untrusted, because content can be written to trick an assistant into doing something you did not ask. Require human approval before sensitive sharing or consequential actions.
Does MFA protect AI accounts?
Multi-factor authentication reduces the risk that a stolen password gives someone access to an AI account and everything it can reach. It does not fix permissions that are already too broad, so pair it with access reviews and identity and access management.
What if someone uploads confidential information?
Mistakes happen, and good AI data security depends on quick reporting rather than hidden mistakes. Give staff a clear, no-blame way to say what happened.
Stop and report
Stop further sharing and tell the person responsible for IT or privacy promptly.
Record the details
Note the tool, account, time and information involved.
Limit further exposure
Restrict shared links or connected access as instructed, and contact the provider where needed.
Assess the exposure
Work out what was exposed and whether any privacy or contractual notification duties apply.
Rotate any credentials
If passwords, keys or recovery codes were included, revoke or rotate them through the approved process.
Deleting a chat does not prove that all retained copies have disappeared, because retention depends on the provider, plan and settings. Our incident response plan guide explains how to prepare a process that covers AI data security incidents, such as uploads, alongside other incidents.
How does the Privacy Act apply to AI tools?
Using AI does not automatically breach the Privacy Act, and overseas processing is not automatically unlawful. Check the purpose for using personal information, the security safeguards in place and whether the provider acts only on your behalf or uses the information for its own purposes. Privacy and AI data security overlap, so assess both together.
The Privacy Commissioner’s guidance on AI and the information privacy principles recommends a privacy impact assessment before introducing AI uses involving personal information. Where a provider uses information for its own purposes, an overseas disclosure may be covered by IPP 12, and the Commissioner explains the requirements for sending information overseas.
IPP 3A, which took effect on 1 May 2026, adds notification duties when personal information is collected indirectly. It is not a blanket rule about AI, but it matters if an AI tool gathers information about people from other sources. Our NZ Privacy Act compliance guide covers the wider obligations, and our guide to Privacy Act security safeguards covers the security duty.
How do you keep AI use under control over time?
AI tools, plans and features change quickly, and a decision made at approval can drift. Build AI data security into normal routines rather than treating it as a one-off project.
- Keep the approved list current. Add tools that meet a real need, record their permitted use and remove ones nobody uses.
- Review on a schedule. Revisit each approval on its review date, and sooner when a provider changes its terms or adds connected features.
- Watch for new built-in AI. Features added to existing software may need the same checks as a new tool.
- Train with real examples. Short, practical AI training shows staff the approved tools and the sharing rules in context.
- Keep the request route open. A fast way to request a tool reduces the pull towards unapproved alternatives.
General security awareness training supports the same habits, such as checking before sharing and reporting mistakes early.
AI data security support from Exodesk
Exodesk helps businesses put AI data security in place: deciding which AI tools and data their team can use, setting up approved accounts and permissions, and training staff. Our AI Solutions page covers implementation help, and the Microsoft Copilot page covers the Microsoft-specific route. Exodesk has supported New Zealand businesses since 1989. Today, our Christchurch and Dunedin teams support businesses nationwide.
NEXT STEP
Need help deciding which AI tools and data your team can use?
Explore Exodesk’s AI Solutions, or talk to us about approved tools, permissions and staff training.
Or see our AI training.
Frequently Asked Questions
What is AI data security?
AI data security is the set of rules, settings and habits that protect business information when staff use AI tools. It covers approved tools and accounts, what information can go into them, how providers use and keep it, and what connected assistants can reach.
Is it safe to use AI with business data?
For AI data security, only use tools and work accounts approved for that information and task. Check provider terms, access, retention and connected services. A business subscription does not make every upload appropriate.
What information should staff keep out of AI tools?
Passwords, recovery codes and keys should stay out of general-purpose AI chats. Client files, payroll, health records and confidential contracts need explicit approval for the data and task first. Public material can go into approved tools, with the output checked before use.
Does turning off training make a personal account suitable?
Not by itself. Training controls do not settle retention, central administration, sharing or contractual requirements. Your business still needs to approve the account and use case.
What is the difference between consumer and business AI?
Business plans may add contractual protections and administration. The exact features vary. Compare the actual plan and settings rather than assuming that free means unsafe or paid means safe.
Does Microsoft 365 Copilot keep data in New Zealand?
Not necessarily. Microsoft says prompts, responses and Microsoft Graph data are not used to train foundation models and that Copilot follows existing permissions, but prompts may be processed in other regions. Web search queries go to Bing outside the Microsoft 365 boundary unless web search is turned off.
What is shadow AI?
Shadow AI is AI used for work without the approval or oversight of the business. It describes approval status, not a type of product, so an unapproved paid subscription is still shadow AI. Give staff approved options and a quick way to request new tools.
Can Copilot show staff files they should not see?
Copilot only surfaces content a person already has permission to view. If files are overshared, it can make them easier to find. Review broad sharing and permissions before rollout.
Does using AI breach the Privacy Act?
Not automatically. The answer depends on the information, purpose, safeguards and how the provider handles it. Assess the proposed use before uploading personal information.
Should we do a privacy impact assessment before using AI?
The Privacy Commissioner recommends one before introducing AI uses involving personal information. It helps you check purpose, safeguards, provider terms and any overseas disclosure before the tool is in use.
What should we do if someone uploads confidential information?
Stop further sharing and tell the person responsible for IT or privacy promptly. Record the tool, account, time and information involved, then assess what was exposed and any notification duties. If credentials were included, revoke or rotate them.
Does deleting a chat remove the data?
Not necessarily. Retention depends on the provider, plan and settings, and copies may be kept for a period after deletion. Check the provider terms and record what happened.
Do AI accounts need multi-factor authentication?
Yes. Multi-factor authentication reduces the risk that a stolen password gives someone access to an AI account and the information it can reach. It does not fix permissions that are already too broad.
Do small businesses need AI data security?
Yes. Small teams adopt new tools quickly and handle client information like larger organisations. A short approved list, clear sharing rules and a reporting route cover the main AI data security risks.
Contact us or call 0800 396 337 to talk about AI data security.
