Yes, and nothing prohibits it. What does not move is who signs, and the standard of care that applies if the engagement later goes wrong.
Key takeaways
No Canadian accounting body prohibits the use of AI on client work. What does not change is who is responsible for the file: a practitioner signs, and the standard they are judged against afterwards does not soften because a tool produced the first draft.
So the useful question is not permission. It is which steps in an engagement are clerical, which are judgement, and whether your firm can still tell the difference once the drafting gets fast.
A practice delivering to a book of clients does four broadly separable things: it gathers records, it processes them, it forms a conclusion, and it reports. AI is strong on the first two, useful on the fourth, and has no business in the third.
Gathering. Building the query list from last year’s file, chasing missing documents, reading and indexing what arrives, renaming and filing it consistently.Processing. Extracting fields from statements and slips into your working papers, flagging outliers and inconsistencies for a person to look at, drafting the lead schedules’ commentary from figures you produced.Concluding. Nothing. Materiality, going concern, the sufficiency of evidence, whether a treatment is appropriate — these are professional judgements attached to a named person.Reporting. First drafts of the covering letter, the management letter and the client-facing explanation, from your conclusions.How much can be automated depends on what you have been engaged to do. the Auditing and Assurance Standards Board covers audit, review, compilation and other assurance engagements, and each of those carries different responsibilities, different evidence requirements and a different report. A shortcut that is unremarkable in a compilation can be a file deficiency in an assurance engagement.
Note also the collision to avoid: this is about a practice delivering to clients. If what you are automating is your own firm’s bookkeeping or month-end, that belongs with AI in accounting operations and AI in bookkeeping, and the risk profile is completely different because nobody outside is relying on it.
Client records are full of personal information — SINs, salaries, health spending, family details. The fair information principles in Schedule 1 of PIPEDA require accountability, identified purposes, limited collection, limited use and disclosure, safeguards and openness.
For generative tools specifically, the Privacy Commissioner’s principles for responsible, trustworthy and privacy-protective generative AI expect each party to know and document its legal authority for the collection, use, disclosure and deletion of personal information across the system’s life, and to be open about it. Alberta, British Columbia and Quebec have their own private-sector privacy laws; where information crosses a border in the course of commercial activity, PIPEDA is back in play regardless of where the practice sits.
If an engagement goes wrong, the client’s route is a professional negligence claim against the accountant, assessed against the professional standard of care. Whether the claim is framed in negligence or in contract changes the analysis in ways that matter, and a disclaimer clause has limits.
There is one point in the firm’s favour worth knowing: where the client withheld or misstated information, contributory negligence can reduce the claim. That defence depends entirely on the quality of your file — on being able to show what you asked for and what you were given. A well-kept AI-assisted file helps here; a fast one with no record does not.
Record on the file which steps used a tool and what a person verified. Recompute at least one control figure independently of anything extracted. Keep the source documents, not just the extracted values. Have a second person review anything that reaches a client under the firm’s name. And write down the rule about what may never be pasted into a public tool, because unwritten rules do not survive a busy season.
A five-partner practice uses extraction on personal tax season only. Slips and statements are read and their fields written into the working papers; the preparer reviews and corrects rather than keying. The query list for each client is drafted from the prior-year file and edited by the preparer.
Nothing else changes. Materiality decisions, unusual treatments and every return are reviewed and signed exactly as before. The measurable difference is in elapsed days from documents received to return prepared, and in how many files reach review with a complete record — not in how many returns one person can sign.
Can AI sign a return or a report?
No. A signature is a person accepting responsibility. Nothing about a tool changes who that person is or what they are accepting.
Do we have to tell the client?
Look at your engagement letter and your confidentiality obligations first. Where client data is being processed by a third party, that is a disclosure question your terms may already answer — and if they do not, they should.
What is the single biggest risk?
Speed without a record. A file that was assembled quickly and cannot show what was checked is the file you least want to defend, and the one most likely to have a quiet extraction error inside it.
A 30-minute call is enough to tell you whether AI pays for itself here.
