AI for accountants in Canada: where it helps a firm today
Four jobs, and they are different products: reading client paperwork into the return, capturing bookkeeping receipts, researching tax law, and drafting the writing around the work. None of them signs a return or files one. The test for every tool is the same: can you see where each answer came from, and where your clients' records went to get it.
Four jobs, four kinds of product
“AI for accountants” is sold as one thing and bought as four. They solve different problems for different people in the firm, and a tool that is excellent at one is usually no help with the others.
- Reading client paperwork into the return. The envelope a client drops off in April: medical receipts, rental records, self-employment logs, donation receipts, statements. Software reads it, sorts it by what each document is, pulls out every figure and enters it into the firm’s tax software. This is where most tax season hours go, and it is the job KeyGlance does.
- Capturing bookkeeping receipts. Receipt capture for a business’s books, feeding QuickBooks Online or Xero through the year. Mature, widely used, and a different job from a personal return: it categorises transactions for a ledger rather than filling lines on a T1.
- Researching tax law. Tools that answer questions about the Income Tax Act and case law, with citations. Blue J, founded in Toronto, is the best known in Canada; several American products have added Canadian legislation.
- Drafting the writing. General assistants that draft the client email, the engagement letter, the memo. Useful, cheap, and the place to be most careful about what client information is pasted in.
What none of them does
No AI product signs a return, takes responsibility for one, or should file one. The certified T1 programs still prepare and transmit the return, and a person with a designation still reviews it. The useful question is never “can AI do the accounting”; it is which hours in the firm are spent on work that does not need an accountant’s judgement, and whether software can take those hours without hiding anything from the person who signs.
The two questions to ask every tool
Can you see where each answer came from? A figure on a return has to be traceable to a document, and a research answer has to be traceable to a source. A tool that hands over a number without showing where it read it has moved the typing into checking, and the checking is now harder because there is nothing to check against.
Where did your clients’ records go to get it? Ask where documents are stored, where they are read, whether the provider keeps them, and whether they are used to train a model. A firm remains accountable under PIPEDA for personal information it passes to a service provider, including when that provider processes it outside Canada, and clients should be told. KeyGlance’s own answer: documents are stored in Canada and read in the United States by Anthropic’s model under a zero data retention agreement, so nothing is kept by Anthropic or used for training, and setup asks the firm to confirm it will tell its clients.
Where the hours actually are
In a Canadian practice, the slip half of a personal return has largely stopped being data entry: Auto-fill My Return brings down what employers, banks and issuers file with the CRA. What it cannot bring down is everything nobody files with the CRA, which is most of what makes a return slow. A family with a rental property, a side business, a year of medical bills and a stack of donation receipts is not a hard return because the tax is complicated. It is hard because someone has to sort and type all of it.
That is the work AI is best placed to take today: high volume, rule-bound, checkable against the document it came from, and done in the weeks when a firm has the least time to spare. See what Auto-fill My Return does not fill for the full list.
What we do about it
KeyGlance takes the client’s whole dropoff, even for the most complex return. It reads every document, links each figure to the spot on the page it came from, and once you have checked them, one click enters them all into TaxCycle, Taxprep Desktop, ProFile, Taxprep Cloud or DT Max on your own Windows machine. It never files anything with the CRA.
Questions we get asked about this
Will AI replace accountants?
Not the part a client pays for. What AI is good at in a firm today is the volume work that sits in front of judgement: reading paperwork, sorting it, typing it, drafting the first version of a letter. The return is still signed by a person who is accountable for it, and no tool on the market changes that.
Is it safe to put client tax documents into an AI tool?
It depends on the tool, and a firm can find out in three questions: where are the documents stored, where are they read, and does the provider keep them or train on them. A firm stays accountable under PIPEDA for personal information it hands to a service provider, so the answers belong in the engagement letter, not in a sales call.
Where should a small firm start with AI?
With the work that takes the most hours for the least judgement. In most Canadian practices that is tax season data entry: the medical receipts, rental records and self-employment paperwork that Auto-fill My Return cannot bring down and someone still types line by line.