Personal tax season is a logistics problem wearing a tax problem's clothes. AI is worth something on the logistics; it is worth nothing on the position you take, and the Act is clear about who wears that.
Key takeaways
Canadian firms use AI in T1 season on the parts of the job that are clerical volume rather than judgment: triaging the slips and receipts a client dumps in the portal, reading the figures off them into a staging sheet, comparing this year's document set against last year's return to produce a missing-items list, drafting the client query email, and drafting the covering letter that goes out with the finished return. Every one of those is a drafting or extraction task with a human check at the end. None of them is the return.
What AI is not doing, in any firm that intends to still be licensed in April, is deciding whether an expense is deductible, whether a taxpayer is a resident, or how to characterise a disposition. Those are positions, and a position needs a person who has read the source documents and is prepared to be assessed on it.
The reason a tax practice compresses at all is written into the Act. Under paragraph 150(1)(d) of the Income Tax Act, an individual's return for a taxation year is due on the following 30 April; subparagraph (ii) moves that to the following 15 June where the person carried on a business in the year, subject to the tax-shelter exclusion in that subparagraph. Nothing about AI changes that date. What it can change is how much of the work sits in the last fortnight.
The second structural fact is that the filing itself is electronic by law for anyone at scale. Subsection 150.1(2.2) defines a tax preparer as a person or partnership who in a calendar year accepts consideration to prepare more than five returns of corporations, more than five returns of individuals other than trusts, or more than five returns of estates or trusts. Subsection 150.1(2.3) then requires that preparer to file by electronic filing, allowing five of each category to go by another route. Miss it and subsection 162(7.3) sets the penalty at $25 for each individual return and $100 for each corporate return. That is a small number per file and a large number across a season, and it is the sort of tracking a workflow tool should be doing rather than a person.
Note what is common to all four: the AI produces something a person then checks against a document that already exists. That is a materially different risk profile from asking a model a tax question, where there is nothing to check the answer against except another answer.
The hard line is the position. Section 163.2 imposes a penalty on a person who makes, participates in, assents to or acquiesces in the making of a false statement that could be used by another person for a purpose of the Act, where the person knows it is false or would reasonably be expected to know but for circumstances amounting to culpable conduct. That phrase is defined in subsection 163.2(1) as conduct tantamount to intentional conduct, showing indifference as to whether the Act is complied with, or showing a wilful, reckless or wanton disregard of the law. Subsection 163.2(5) sets the penalty at the greater of $1,000 and a further amount computed by reference to what the taxpayer would have faced.
Read that definition again with an AI-drafted schedule in mind. ‘Indifference as to whether the Act is complied with’ is exactly what pasting an unverified model output into a return looks like from the outside. The defensible posture is the ordinary one: the preparer read the source document, formed the view, and can say why.
Two provisions decide what your file has to contain, and both survive any change in tooling. Subsection 150.1(4) provides that where a return is filed electronically by a filer other than the person required to file it, that person must make an information return in prescribed form, retain a copy and provide it to the filer — and the Act deems that return and the copy to be a record referred to in section 230 for both of them. So the signed authorisation is not administrative courtesy; it is a statutory record.
Section 230 then sets the retention rule. Subsection 230(1) requires every person carrying on business to keep records and books of account in a form and containing information that will enable the taxes payable to be determined; paragraph 230(4)(b) sets a general floor of six years from the end of the last taxation year to which the records relate; and subsection 230(4.1) adds that records kept electronically must be retained in an electronically readable format for that period. A model's chat history is not a record in that sense. The staging sheet, the source documents and the reviewer's notes are. Treadstone Law's plain-language guide to corporate record retention covers the corporate-law side of the same question, and its note on whether the CRA can reopen closed years explains why the retention floor and the reassessment window are not the same number.
A T1 file is a dense concentration of personal information: SIN, income, medical expenses, dependants, sometimes marital breakdown. The Office of the Privacy Commissioner's principles for responsible, trustworthy and privacy-protective generative AI make the framing explicit — generative AI does not sit outside existing privacy law, and organisations developing, providing or using it must ensure their activities comply with the applicable Canadian privacy legislation. In practice that means the deployment question comes before the capability question: which tenant the data sits in, whether it is used for training, and who at the vendor can see it. The federal Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems is worth reading alongside it, because it sets out the outcomes — accountability, safety, fairness and equity, transparency, human oversight and monitoring, validity and robustness — that a vendor should be able to speak to when you ask.
A three-partner practice in Ontario runs roughly 900 personal returns between February and April with six seasonal staff. Its bottleneck is not preparation time; it is the number of times a file is opened, found to be short a document, closed and re-queued. The partners instrument that: for one season they record, per file, how many times it was picked up before it was signed.
The following year they change one thing. On intake, every document set is classified and reconciled against the prior-year return, and a single consolidated query list goes to the client before a preparer ever opens the file. The reviewer confirms the list, because the list is a client communication and carries the firm's name.
The measurable is the same one they already had: touches per file, and the share of files signed on first preparation. Both are countable from the practice-management system before and after. ‘Hours saved’ is not countable, which is why it makes a poor business case and a worse post-mortem.
Two adjacent jobs get confused with this one. If what you are trying to fix is the front desk — who answers, who books, who reminds — that is a practice-operations question and it lives on the professional practice owners page. If what you are trying to fix is your own firm's bookkeeping and month-end rather than the client work, that is the bookkeeping automation page. This article is about a practice delivering personal tax work to a book of clients.
Can AI prepare the return itself?
It can populate a staging sheet from source documents. It cannot take the position, and it cannot be the person who is expected to know whether a statement is false under section 163.2. The preparer remains the preparer.
Does using AI change my filing obligations?
No. The dates in section 150 and the electronic-filing requirement in section 150.1 are indifferent to how the file was assembled.
What is the first thing to automate?
The query list. It is the step that determines how many times every other step has to be repeated, and it is checkable against two documents you already hold.
A 30-minute call is enough to tell you whether AI pays for itself here.
