Tax professionals regularly work with information obtained from the Canada Revenue Agency (CRA): notices of assessment, tax slips, account balances, carry-forward amounts, correspondence, benefit information, and other taxpayer records.
Generative AI creates a new question for firms handling this information:
How much of it should ever reach an AI tool?
There is no blanket rule saying that every piece of information obtained from the CRA can or cannot be used with generative AI. The answer depends on the information involved, the purpose, the AI service being used, applicable privacy requirements, and the firm's authorization to handle the information.
But one principle provides a useful starting point: access to taxpayer information does not mean that information should automatically be copied into an AI system.
CRA Information Should Be Treated as Sensitive
Information retrieved from CRA systems can reveal extensive details about an individual or business.
Examples can include:
- Social Insurance Numbers
- Names and addresses
- Income information
- Tax balances
- Employer information
- Benefit information
- Registered account details
- Tax slips
- Assessment information
- Correspondence with the CRA
The CRA itself describes taxpayer-information protection as a significant responsibility. Its internal security practices include need-to-know access, employee awareness, risk assessments, controls against unauthorized access, and procedures for investigating suspected breaches.
Those specific internal CRA controls do not automatically become legal requirements for private tax firms. However, they illustrate an important security principle: sensitive tax information should only be accessible and used when necessary.
The same principle is useful when evaluating AI workflows.
Rule 1: Do Not Give AI More CRA Data Than It Needs
Suppose a tax professional wants an AI assistant to explain a CRA notice in simpler language.
The AI may need the wording of a particular paragraph.
It probably does not need the taxpayer's:
- Name
- Address
- Social Insurance Number
- Account number
- Complete tax history
- Other unrelated pages of the notice
Instead of uploading the complete document, the tax professional could extract the relevant section and remove identifiers.
Canadian privacy regulators recommend using anonymized, synthetic, or de-identified information rather than personal information when identifiable information is unnecessary. They also recommend limiting collection, use, and disclosure to what is needed for the identified purpose.
For tax teams, this translates into a simple workflow:
Minimize first. Prompt second.
Rule 2: Redact Before Uploading Documents
Tax documents often contain identifying information in places employees may overlook.
A notice might include a name and address at the top, an account number on another page, or identifying information in headers and footers.
Before information is submitted to an AI tool, tax teams should consider removing unnecessary identifiers and unrelated financial information.
Redaction may include removing or replacing:
- Social Insurance Numbers
- Taxpayer account numbers
- Names
- Addresses
- Dates of birth
- Banking details
- Signatures
- Unrelated income figures
- Information about dependants
Importantly, redaction should happen before the information reaches the AI service.
Asking an AI system to remove sensitive information after uploading the original document does not prevent the original information from having already been submitted.
Rule 3: Do Not Assume a Public AI Chatbot Is Appropriate for Taxpayer Data
The CRA's own public generative AI chatbot provides a useful illustration of this issue.
The CRA explicitly tells users not to provide personal details such as names, addresses, Social Insurance Numbers, dates of birth, or financial information to its GenAI chatbot. The CRA explains that the chatbot is not designed to securely handle personal or sensitive information.
This notice applies specifically to the CRA chatbot; it is not a universal rule governing every commercial AI product.
But it demonstrates why tax teams should examine the intended use and security design of an AI service before supplying sensitive information.
A convenient chatbot and an enterprise system approved to process confidential business information are not automatically equivalent.
Rule 4: Understand What the AI Vendor Does With Prompts
Before allowing taxpayer information into an AI service, firms should know what happens after the employee clicks Send.
Questions should include:
- Are prompts stored?
- How long are they retained?
- Can prompts be used to train or improve models?
- Who can access submitted information?
- Where is information processed?
- Are subcontractors involved?
- Can administrators control employee accounts?
- Can information be deleted?
- What happens when the firm's contract ends?
Canadian privacy regulators recommend that organizations understand secondary purposes for personal information, establish appropriate retention schedules, and ensure sensitive or confidential information is only entered into generative AI prompts when authorized.
A firm's AI policy should therefore distinguish between approved systems and unsanctioned consumer tools.
Rule 5: CRA Authorization Does Not Equal Unlimited Data Permission
A taxpayer may authorize a tax professional to access information through the CRA's Represent a Client service.
That authorization gives the representative access to information necessary to perform authorized tax-related activities. CRA has continued strengthening the authorization process so that only authorized representatives receive access to taxpayers' information.
But obtaining access does not answer every privacy question about what happens to the data afterward.
A tax firm still needs to consider whether transferring that information into another system is necessary, appropriate, secure, and consistent with its authority and privacy obligations.
In other words:
Authorized access to data is not the same thing as authorization for every possible use of that data.
Rule 6: Verify AI-Generated Tax Information
Privacy is not the only risk.
Generative AI can provide incomplete, outdated, or inaccurate tax information.
The CRA itself warns users of its GenAI chatbot that responses may not be fully accurate or appropriate to an individual's circumstances and should not be treated as professional or legal advice.
Similarly, NIST's Generative AI Profile is designed to help organizations identify and manage risks specific to generative AI across its lifecycle.
Tax professionals should therefore verify AI-generated conclusions against authoritative sources before using them for:
- Filing decisions
- Tax calculations
- Eligibility determinations
- Deadlines
- Interpretations of CRA guidance
- Client recommendations
AI can accelerate research and drafting. Professional review still needs to determine whether the result is correct.
Build a Safer CRA-to-AI Workflow
A practical tax-team workflow might look like this:
CRA information → determine necessity → remove unnecessary data → redact identifiers → use approved AI system → review output → verify against authoritative sources.
This approach allows tax teams to benefit from generative AI without treating entire taxpayer records as routine prompt material.
The key principle is not complicated:
Just because information is available to the tax professional does not mean it needs to be available to the AI.
When tax teams minimize data, use approved systems, understand vendor practices, and independently verify outputs, generative AI becomes easier to integrate without unnecessarily increasing the exposure of sensitive taxpayer information.
The workflow above only works if the redaction step happens before the prompt is sent, because once a prompt is submitted it cannot be recalled. A Social Insurance Number in a pasted notice, an account number in a footer, or an unrelated page of a CRA letter is already gone the moment someone clicks Send. This is the principle Sanitized Ai is built on, catching and redacting taxpayer identifiers and other sensitive CRA-derived information before the prompt reaches the tool, so the minimized input is the only input that goes out.
This quarter, take one routine task where your team uses CRA information with AI, such as explaining a notice of assessment, and walk it through the workflow above step by step. Write down which identifiers can be removed without changing the answer, and make that the standard for the task. If you want to see how redaction can happen before submission rather than after, request a demo and walk through your own workflow with us.