AI-assisted tax preparation in Canada: what still needs human review?

AI can assist preparation; people still need to verify the result

AI-assisted tax preparation is most useful when the task is defined: organize records, extract proposed facts, draft questions, or assemble material for a reviewer. These are possible workflow designs to evaluate, not a guarantee about any particular tool’s accuracy or available features.

This article is about designing a Canadian tax preparation workflow. It does not determine an individual’s tax position. The appropriate professional needs to assess the applicable rules, tax year, jurisdiction, and client circumstances before a conclusion is used.

Imported information is not a completeness check

The CRA says Auto-fill My Return can supply only information it has when the request is made. Before filing, the information must be checked for accuracy and completeness, including items the service did not provide. Taxpayers remain responsible for reporting income accurately.CRA: Auto-fill My Return and your responsibilities

Auto-fill My Return is a CRA data service, not a generative AI agent. Its guidance nevertheless illustrates an important workflow distinction: transferring available information and verifying a complete return are different tasks.

Our recommendation is to reconcile expected documents with received documents before treating a file as ready. If a client mentions an income source that has no corresponding record, create an open question rather than assuming the data import is complete.

Check the evidence behind an AI explanation

NIST warns that generative AI can produce incorrect content and fabricated logic or citations. A source-looking reference is not enough; the reviewer needs to open the source and check whether it supports the claim.NIST: Generative AI Risk Management Profile, section 2.2

For a proposed tax explanation, inspect the actual publication and its scope. Confirm the relevant year and jurisdiction, separate the client’s documented facts from assumptions, and identify anything that depends on missing information. If the source and the facts do not support the conclusion, return it for correction.

A document summary also needs comparison with the original. Check the taxpayer or business identity, period, amounts, and whether the record has been amended. The precise checks should follow the engagement and the reviewer’s professional requirements.

Make the review package easy to inspect

We recommend putting the proposed result, supporting records, and unresolved questions together. The reviewer should not need to repeat the whole collection process just to understand what the agent did.

Use separate states for prepared, returned for changes, and approved. Record who made the decision and which version they reviewed. Those are suggested design choices for accountable handoffs, not a claim that a status label by itself satisfies professional obligations.

• Proposed facts linked to the records they came from.
• Official guidance used for the explanation.
• Assumptions and information still missing.
• Disagreements or inconsistencies the agent identified.
• A named reviewer and a clear approval or return decision.

Keep responsibility and client communication clear

The Canadian privacy authorities’ AI principles state that accountability for decisions remains with the organization using the system. Their guidance also calls for explaining AI’s role in decision-making and the safeguards involved.Canadian privacy authorities: Principles for generative AI

For your team, define who can release a client explanation or final deliverable. A completed preparation task should lead to a review decision, not silently become an approved conclusion. JAN AI is being designed around that separation: agents prepare work, and people decide what goes out.