How to use AI on confidential documents safely

Every organisation handles sensitive documents.

When you paste a contract into a general-purpose chatbot or upload a spreadsheet of client data to an AI summariser, you are trusting that service with information you may be legally obligated to protect. Understanding where the risks lie, and what a properly secure alternative looks like, is the difference between using AI productively and creating a data incident.

The real risks of using AI on sensitive files

The most common way people use AI on documents today is by copying text into a chat window. ChatGPT, Claude, Gemini: each of these tools accepts pasted text and uploaded files. Each also has its own policies on what happens to that data once it arrives.

Some providers retain inputs for a period of days or weeks. Some use conversation data to improve their models unless you explicitly opt out. Some route your data through third-party infrastructure. The policies change, and the defaults are rarely the most private option available.

For a personal query, this may be acceptable. For a confidential contract, a client's financial records, or an employee's performance review, it is not. The moment sensitive text leaves your control and enters a system where it could be stored, logged, or used for training, you have a potential compliance problem.

Beyond training data concerns, there are practical risks. Browser-based AI tools may cache inputs locally. Shared accounts mean shared history. And if a provider suffers a breach, your pasted text may be among the exposed data. Understanding what data AI tools collect is essential before you use any of them on confidential material.

What to look for in a secure AI tool

Not every AI-powered tool handles data the same way. When evaluating whether a platform is safe for confidential documents, there are several things worth checking before you upload anything.

First, look at encryption. The minimum standard is AES-256 encryption both in transit and at rest. This means your files are protected while being sent to the server and while stored there. Without encryption at rest, anyone with access to the underlying storage infrastructure could theoretically read your documents. Learning how end-to-end encryption works will help you distinguish marketing language from genuine protection.

Second, check training policies. A secure AI tool should make an explicit commitment: your data is not used to train AI models. This should not be buried in a footnote or offered only as an opt-out toggle. It should be the default.

Third, look at the scope of AI processing. Does the AI operate exclusively within your workspace, or does it send your data to external services? Workspace-scoped AI means that the model processes your files in the context of your own account and does not mix your data with other users' content.

Fourth, consider compliance. Certifications like CASA Tier 2 indicate that a platform has been independently assessed against security standards. This matters particularly if you work in a regulated industry or handle client data subject to contractual confidentiality obligations.

Finally, examine access controls. Even within a secure platform, you need the ability to restrict who can see what. Granular permissions and password protection for specific files or folders let you maintain confidentiality even when collaborating with others.

How Fabric handles confidential documents

Fabric was designed for people who need AI capabilities without compromising on security. Every file stored in Fabric is protected with AES-256 encryption in transit and at rest. The platform is CASA Tier 2 compliant, meaning it has undergone independent security assessment.

The AI assistant in Fabric works exclusively within your own workspace. When you ask the AI to summarise a contract or search across your files, it processes that request in the context of your account. Your data is not sent to a shared pool, not used to train models, and not accessible to other users.

This workspace-scoped approach is fundamentally different from pasting text into a general-purpose chatbot. With Fabric, your documents remain in your cloud, under your control, subject to your permissions. The AI works on your data where it lives, rather than requiring you to send it somewhere else.

When you need AI to go further than answering questions, Fabric's agents operate under the same security model. You can grant an agent (Instinct, Muse, or Claude) scoped, time-limited access to a specific connected tool: reading a particular Slack channel, checking a Gmail inbox, or searching a Drive folder. The agent works within those boundaries and the permission expires on schedule. This is a meaningful distinction from tools that ask for broad, permanent access to your accounts. With Fabric, every grant is limited in scope and duration, and you approve each one before the agent acts.

The scoped access model is itself a security feature. Rather than granting an AI tool broad, permanent access to your email or messaging accounts, you approve a specific agent for a specific tool for a defined period. When the time expires, the access ends. This means the blast radius of any single permission is small, and you maintain a clear record of what was granted and when.

Connected apps in Fabric use OAuth for authentication, which means no passwords are stored on the platform. When you connect a service like Google Drive or Dropbox through Fabric's connections, the authentication is handled through the provider's own secure token system. This reduces the surface area for credential-related incidents.

Privacy in Fabric is the default, not an opt-in setting. Your workspace is private and secure from the moment you create it. Activity history is fully tracked, so you always know who accessed what and when. And if you close your account, your data is permanently deleted.

This security model extends to how you access Fabric from other tools. You can use your Fabric vault from ChatGPT, Claude, or any MCP-compatible application, and your data stays in your cloud throughout. The cloud drive streams files on demand with zero local storage, so confidential documents never need to sit on a device to be available to AI. The retrieval layer sends only relevant excerpts to the model, regardless of which interface you use to ask the question.

Protecting specific files within your workspace

Even within a secure platform, some documents need an extra layer of protection. A board resolution, a merger agreement, or an employee disciplinary record may need to be accessible only to specific people, even if they share a workspace.

Fabric offers password protection at the file and folder level. You can set a password on any document or folder, meaning that even someone with general workspace access would need the password to view its contents. Combined with granular permissions, this gives you layered security that matches the sensitivity of each document.

This matters for data minimisation in practice. Rather than granting broad access and hoping people exercise restraint, you can enforce access limits structurally. The AI assistant respects these boundaries: it only processes files that the requesting user has permission to access.

The difference between secure AI and no AI

Some organisations respond to AI security concerns by banning AI tools entirely. This is understandable but counterproductive. People will find ways to use AI regardless, and unofficial use is far more dangerous than sanctioned use of a secure tool.

The better approach is to provide a tool that meets your security requirements and makes AI accessible within those guardrails. When people can use AI without losing their data, they do not need to paste confidential text into a consumer chatbot. When search is built into a secure workspace, there is no reason to upload files to an external summarisation service.

Fabric's search works across everything in your workspace, including connected apps, making it possible to find information without moving files to less secure locations. The AI understands the content of your documents, not just filenames, so you get meaningful results without exposing your data to third parties.

For organisations evaluating the tradeoffs, the AI without the risk overview covers how Fabric balances capability with protection.

Building a practical approach to confidential document AI

The goal is not to eliminate risk entirely. That would mean eliminating AI, which would mean eliminating its considerable benefits. The goal is to reduce risk to an acceptable level through a combination of platform choice, access controls, and clear policies.

Start by understanding who owns your data in cloud storage. If you are using a platform where the provider claims broad rights to your content, that is a foundational problem no amount of careful usage can solve.

Next, establish what counts as confidential in your context. Not every document requires the same level of protection. A marketing brief and a client's financial statement have different risk profiles, and your tooling should reflect that.

Then choose a platform that treats security as a default rather than a premium add-on. A personal data vault approach, where your data stays under your control and AI processes it in place, is significantly more secure than a model that requires you to export and upload files. Fabric's self-writing documentation can also help with compliance visibility: as your workspace grows, Fabric generates and maintains wikis from your content, giving you a browsable record of what your library contains without requiring manual cataloguing.

For further reading on the broader landscape, the AI and privacy guide covers what has changed in 2026 and what to expect going forward.

Frequently asked questions

Is it safe to paste confidential text into ChatGPT or Claude?

It depends on your settings and the provider's current policies. By default, some providers may use your inputs for model training, and most retain data for some period. For highly confidential material, such as contracts, client data, or HR records, it is safer to use a tool that explicitly commits to not training on your data and processes files within your own workspace. Understanding what data AI tools collect will help you make informed choices about which tools to trust with sensitive information.

Does Fabric use my documents to train AI models?

No. Fabric does not use your data to train AI models. Your files are processed by the AI assistant exclusively within your own workspace, and the platform's privacy-by-default design means your content is never shared with other users or used to improve models.

What encryption does Fabric use for stored documents?

Fabric uses AES-256 encryption for data both in transit and at rest. AES-256 is widely regarded as the standard for protecting sensitive data. You can read more about what zero-knowledge encryption means for cloud storage and how different encryption approaches compare.

Can I restrict AI access to specific files in my workspace?

Yes. Fabric's AI assistant respects the permission and access controls you set. If a file is password-protected or restricted to specific users, the AI will only process it for users who have the appropriate access. This means you can store highly sensitive documents alongside everyday files without worrying about unauthorised AI processing.

What is CASA Tier 2 compliance?

CASA (Cloud Application Security Assessment) is a framework for evaluating the security of cloud applications. Tier 2 compliance means that Fabric has undergone independent assessment against a defined set of security requirements. It is worth noting that platform-level compliance supports but does not automatically guarantee compliance for every specific regulatory framework your organisation may be subject to. Your deployment context and usage patterns also matter.

How is Fabric different from using a regular AI chatbot with my files?

When you use a regular chatbot, you are sending your data to an external service with its own retention and training policies. With Fabric, the AI works within your workspace. Your files stay in your cloud, encrypted and access-controlled. There is no need to upload files every time you want AI assistance, and no need to worry about where your data ends up after the conversation.

What happens to my data if I close my Fabric account?

Your data is permanently deleted when you close your account. Fabric does not retain your files, metadata, or AI interaction history after account closure. This is consistent with a data minimisation approach where platforms hold only what is necessary for as long as it is needed.

Can I use Fabric with files from other cloud services?

Yes. Fabric supports connections to other cloud services through OAuth-based authentication. This means you can bring files from services like Google Drive or Dropbox into your Fabric workspace without storing additional passwords. The connected files benefit from Fabric's AI capabilities and security protections.

How do I know who has accessed my confidential documents?

Fabric maintains a full activity history for your workspace. You can see who accessed a document, when they accessed it, and what actions they took. This audit trail is important for compliance purposes and for understanding data breaches if they occur elsewhere in your toolchain.

Should my organisation ban AI tools to protect confidential data?

Banning AI tools outright often backfires. People find workarounds, and unsanctioned use of consumer AI tools is typically far less secure than providing an approved platform. A more effective approach is to select a tool that meets your security requirements, such as one that offers AES-256 encryption, does not train on your data, and scopes AI processing to individual workspaces, and then establish clear guidelines for its use. The AI without the risk overview covers how to think about this balance.