Are employees using their own ChatGPT or Claude accounts to draft emails, summarise documents or work through business problems? They may be saving time. They may also be putting company information into an environment your organisation cannot administer.
This is often called shadow AI: people using AI tools for work outside the organisation’s approved arrangements. The practical response is to understand the useful work being done, provide an approved business workspace and apply controls that keep business information within agreed boundaries.
Here is how to keep the benefits while bringing account ownership, access and data handling under control.
The good: why your team is already using AI
ChatGPT and Claude can help people structure a first draft, explain unfamiliar material, brainstorm options and turn rough notes into clearer writing. Used with appropriate information and human review, they can reduce repetitive work and help staff get started faster.
Personal accounts are convenient. Someone may already pay for a subscription and assume using it at work is helpful. Start by asking which tasks people use AI for, what information they provide and what they find useful. A discussion about real workflows is more productive than assuming every user has acted carelessly.
Those answers help you design an approved alternative that staff will actually use. A restriction without a workable replacement gives people a reason to find another route.
The bad: what personal accounts leave outside your control
Business information can leave approved systems
A prompt can contain customer names, employee details, pricing, contracts, source code or internal plans. Uploading a whole document can disclose more than the employee intended. Removing a name may not make the remaining information anonymous.
The issue is broader than model training. Your organisation needs to understand the provider’s terms, storage, access, retention, sharing and connected services. Turning off training in a personal account does not give your IT team ownership of that account.
Consumer and commercial settings differ
A paid personal subscription is not the same as a company-managed workspace. Nor should you assume every personal conversation is automatically used for training. For example, Anthropic’s consumer privacy guidance explains that users can choose whether their conversations are used to improve its models. Check the actual product, terms and settings rather than relying on the word “paid”.
Offboarding becomes harder
You can remove an employee’s access to company email and still have business material sitting in their independently controlled AI account. Saved conversations, uploaded files, shared links and connected services need consideration. Paying an expense claim for a personal subscription does not establish central administration or a reliable exit process.
Useful answers can still be wrong
AI can produce convincing errors, unsupported references and insecure code. Moving onto a business plan does not remove the need to verify factual claims, review code and approve external communications. If an AI tool can act in connected systems, restrict its permissions and require appropriate review before consequential actions.
Move onto business accounts with the controls you need
Choose an organisation-managed offering, such as ChatGPT Business or Enterprise, or Claude Team or Enterprise. Give each employee their own company-managed seat and assign clear administrative ownership. Avoid shared logins.
OpenAI’s official plan guidance lists a dedicated workspace, SAML single sign-on and no training on business data by default for ChatGPT Business. Enterprise adds controls such as SCIM provisioning, audit capabilities and retention options. These additional controls are not all included in Business.
Anthropic documents single sign-on for Claude Team and Enterprise. Its Enterprise offering includes further identity, retention and compliance capabilities and states that inputs and outputs are not used for model training by default. Claude audit logs are an Enterprise capability; activity logs should not be confused with a complete transcript of every conversation.
Before buying, agree which controls are essential:
- Identity: company sign-in, multi-factor authentication and an effective process for removing leavers.
- Data: approved information types, retention requirements and contractual terms suitable for your work.
- Features: which apps, connectors, sharing options and agent actions employees may use.
- Visibility: the usage reports and investigation records your administrators actually need.
- Device access: whether work is permitted from unmanaged computers or personal phones.
Check each requirement against the exact plan and configuration. A business workspace does not automatically mean zero retention, unrestricted administrator access to every chat or permission to upload any confidential document.
How to restrict personal AI use in practice
Single sign-on secures access to your company workspace. It does not, on its own, prevent someone opening a personal account elsewhere. Personal and business accounts can use the same website address, so blocking a domain cannot normally distinguish them.
Combine clear rules with controls at the device, browser and network layers. Choose measures your IT team can maintain and test.
1. Publish a short acceptable-use policy
Name the approved tools and workspaces, explain what information employees may use and identify what needs prior approval. Give examples: a public product description may be suitable; passwords, customer records and personnel files need much tighter handling. Provide a simple route to request a tool or report a mistake.
2. Use managed browser policies
On managed devices, enforce browser configuration and an approved extension list. Microsoft Edge’s URLBlocklist policy, with URLAllowlist exceptions, can restrict website access. Comparable controls should be assessed for your other supported browsers.
If blocking an unapproved service entirely, addresses such as chatgpt.com and claude.ai are starting points to investigate, not a complete blocklist. Apps, APIs, authentication services and new domains may use other endpoints. Maintain rules against current vendor documentation and observed traffic.
Allowing the approved AI website through a URL exception also allows whatever account types that website permits. A managed browser profile alone does not establish which AI workspace is in use.
3. Apply web gateway and account restrictions where supported
A secure web gateway or suitable cloud access security broker can apply application policies to managed traffic. Where your chosen product supports the AI provider’s workspace or tenant restrictions, allow the approved organisation and deny personal or unapproved workspaces. Verify support separately for ChatGPT and Claude; do not assume one rule works for both.
If reliable account-level enforcement is unavailable, consider blocking the service outside a controlled access environment while providing a tested approved route. Office-only filtering is insufficient for laptops that leave the building: the relevant traffic must continue through your controls when staff work remotely.
4. Control uploads and sensitive content
Data loss prevention (DLP) can help identify or restrict sensitive information leaving managed environments. Some gateways also control individual AI actions. For example, Cloudflare documents application controls that can block ChatGPT uploads, subject to its setup requirements, including a device certificate, TLS decryption and gateway proxying.
Blocking file uploads does not stop someone pasting text, submitting screenshots or using another channel. Test which actions and clients your controls cover, and review false positives before wider enforcement.
5. Manage apps, extensions and alternative routes
Cover desktop applications, browser extensions, mobile apps and developer tools as well as websites. Application allowlisting can limit unapproved programs, while controlled administrative rights make it harder to install unmanaged tools.
Use managed DNS and network policies where appropriate, and assess unapproved VPNs, proxies and encrypted DNS that could bypass the intended route. An office firewall cannot control an unmanaged personal phone on mobile data. For sensitive work, restrict access to company data to approved devices and environments.
6. Test the policy before claiming success
Use authorised test accounts to check the approved workspace, a personal account, signed-out access, a private browsing window, desktop apps and remote working. Include API and extension routes where relevant. Confirm that required business workflows still work, and record both the expected result and any gaps.
Pilot changes with a small group, retain the previous settings and prepare a rollback. Review rules as products and workflows change.
Move the work carefully
Invite employees into the approved workspace and show them how to confirm they are using it. Do not automatically merge entire personal histories or upload old exports: these may mix private material with business information. Review what needs preserving and use the provider’s supported migration route where appropriate.
Agree how existing work material in personal accounts should be handled. Where sensitive information may already have been disclosed, involve your security and data-protection leads, preserve relevant evidence and assess the incident before removing records. Rotate exposed credentials promptly through your established response process.
Be clear about monitoring. Explain what is collected, why, who can access it and how long it is kept. The ICO’s worker-monitoring guidance emphasises necessity, proportionality and transparency. Choose logging that serves an agreed security purpose.
How Astley Digital can help
You do not need to work through this alone. Astley Digital can help assess how AI fits into your existing technology and security arrangements, identify the controls your organisation needs and plan a practical next step.
Our cyber security consulting and virtual CISO services support risk assessment, policies and governance. Fully Managed IT Solutions bring device management, user access and ongoing support together. Where application controls are appropriate, our Fully Managed ThreatLocker service offers a route to managing approved software and privileges, with the modules and scope agreed around your environment.
The aim is to give your team useful AI tools with clear ownership, sensible boundaries and controls you can verify.
Speak to Astley Digital about managing AI use in your organisation. Tell us which tools your team uses, what information they handle and where you need more control.
Product capabilities and official guidance checked on 11 November 2026. Availability depends on the selected plan, configuration and service terms.
Ready for a clearer technology plan?
Tell us what is slowing your organisation down. We will help you identify the practical next step.
