“Can I use AI with my clients’ files?” Sometimes, but not with every tool or under every condition. Accounting firms, law firms and clinics are not automatically excluded from cloud AI; the information, obligations and specific offering need to be assessed.
Do not decide whether to upload a document to AI based on the brand name alone. First establish what the information contains, whether you are authorized to process it this way and which terms apply to the actual product. For a small business or nonprofit, public or fictional information is often a more manageable starting point than client files.
A free account, business subscription and enterprise contract may have different conditions. A commitment not to use data for model training does not mean that data is never retained, accessible to administrators or sent to a connected service.
Separate the types of information
These categories can overlap. A client file may contain personal information, confidential business information and a trade secret at the same time.
- Public: published material, still subject to applicable rights. Public availability does not remove every restriction on use.
- Internal: an unpublished procedure or memo. External processing still needs authorization.
- Personal: information that identifies a person, either alone or alongside other details. An address, role and context may be enough.
- Confidential: information protected by a contract, professional obligation or internal rule, such as an unannounced proposal.
- Trade secret: valuable business information whose value depends in part on remaining secret, such as a method or formula.
- Client information: material entrusted to you by a client. Access does not automatically give you permission to disclose it to an AI provider.
When summarizing a file, ask whether the entire document is necessary. An approved extract or fictional example may answer the question without exposing the real details.
Check the product and the data journey
For ChatGPT, Microsoft Copilot, Gemini or Claude, identify the exact version, account type and governing agreement. “Copilot” can describe different experiences; seeing the logo in a browser does not confirm the protections of a managed Microsoft 365 environment.
Review how inputs and outputs are used, retention, deletion, access, processing locations and the contractual commitments you require. Model training, conversation history and security logs are different purposes and should be assessed separately.
Start with the documentation for your offering: OpenAI business privacy, Microsoft Copilot data and privacy, Google Workspace AI privacy and Anthropic commercial terms. These do not necessarily describe every consumer account. Record when and what you checked rather than assuming the terms will remain unchanged.
Review permissions and connected services
An assistant that searches company files can make information with overly broad permissions easier to discover. Review access before enabling a connection. An assistant used by leadership does not necessarily need access to human resources files.
An extension, action or connector may send information to a different provider. Assess that recipient and its scope separately. Also review shared conversation links and who may see the output.
When does local or hybrid AI make sense?
If certain records must stay within your environment, local AI for client information may help. Some models run on a computer or server controlled by the organization. This does not mean professional cloud services are less secure: they can offer substantial protections, subject to the product, contract and configuration.
A hybrid approach may be enough: search client records internally and reserve cloud processing for public or explicitly approved material. Check the extracts being sent, not just the original file. A summary can still contain confidential information. The guide to internal documents and offline AI explains how to examine the full processing chain.
“Local” does not automatically mean “secure” or “compliant.” Devices, networks, administrator access, document permissions and backups still need protection. Search indexes and logs may also contain sensitive information. Plan updates, employee offboarding and recovery tests. Assign responsibility for these tasks before launching an assistant.
Make the decision before uploading
For each use case, keep short answers to these questions:
- Is the information necessary for the task?
- Does the organization have the right to disclose it for this purpose?
- Are the tool, account and settings approved for this information category?
- Are the people receiving the answer allowed to see the source material?
- Who will review the result and handle a possible error?
If an answer is missing, pause the real-data upload and use a fictional example. Put these decisions in an internal AI use policy.
Is removing names enough?
Not always. Dates, amounts, job titles and events can identify someone when combined. Pseudonymized information, which can be linked back to a person using additional information, is different from genuinely anonymized information. For a trial, inventing a realistic case is often safer than partially obscuring a real file.
Practical AI integration includes these boundaries: a usable workflow needs to protect information as well as produce a useful answer.
What should the first trial look like?
Choose one task and a small set of approved or fictional records. Among the practical uses of local AI, summarizing a non-sensitive procedure is easier to govern than opening all client files to an assistant. Test answer quality and whether access to restricted records is actually blocked.
In Quebec and elsewhere in Canada, have your privacy lead and, where needed, legal counsel assess the obligations that apply to your organization. These considerations support a technology decision; they are not legal advice.