What it does well: reading documents
Turning a photographed receipt into a vendor, a date, an amount, and a tax component is the strongest use. It is high volume, it follows patterns, and mistakes are visible to a reviewer who knows the client. For a firm processing thousands of documents a month, this is the largest single block of time that automation can genuinely reclaim.
The important framing is that this is a first pass. Output arrives as a draft to be checked, and evaluated that way it holds up well. Evaluated as a finished ledger, it does not.
What it does well: repetition and recall
Applying the same coding to the same supplier every month is rule-following, and software does not get bored. So is applying the right provincial rate for the right client, or noticing that a line has no supporting document attached.
Retrieval is similarly well suited. Finding every change to a client's ledger after a given date, or tracing a category total back to source documents, is tedious by hand and quick for a machine.
What it does badly: knowing your client
A four hundred dollar charge at a restaurant might be a client meeting at fifty percent, a staff event, or the owner's anniversary dinner that does not belong in the books at all. The document is identical in all three cases. Nothing in the image distinguishes them, so no amount of model quality solves it. Only someone who knows the client and can ask does.
The same applies to business-use percentages, home office claims, shareholder benefits, and whether an expense is capital or current. These are judgment against a specific set of facts, and getting them wrong is exactly the kind of error that surfaces in a review.
What it does badly: knowing when it is wrong
This is the failure mode that matters most for a firm. A tool that misreads a total produces a confident, well-formatted, plausible line. There is no visible difference between a line it got right and a line it invented, which means the error surfaces during review or it does not surface at all.
That single property should drive how you evaluate any tool in this category. A vendor claiming very high accuracy is describing the average case; your exposure is in the tail, and the tail is only caught by a review step you actually perform. Any workflow where machine output reaches a client or a return without a person in between is taking on risk the accuracy figure does not describe.
What it cannot do: hold the responsibility
If a return is wrong, the CRA's counterparty is the taxpayer, and the professional relationship is with your firm. Responsibility does not move to a software vendor because a draft came from a model. This is not a legal technicality, it is the reason the review step exists and why a tool that encourages skipping it is selling you a liability.
The practical consequence is that automation changes the composition of the work rather than the amount of accountability. Less keying, the same sign-off.
How to evaluate a tool on this basis
Ask where the human checkpoint is and whether the product makes it easy to skip. Ask whether you can see the source document from the line without leaving the screen. Ask what the audit trail records when a figure changes and who changed it.
Then run a real month for a messy client, not a demo. Demo files are clean. Your clients photograph receipts in the dark and send them in March.
Common questions
- Will AI bookkeeping replace bookkeepers?
- It removes data entry, which is a real share of the hours but the least valuable share. Coding judgment, client knowledge, reconciliation, and review remain professional work, and the review requirement grows rather than shrinks as more drafting is automated.
- Is it safe to use AI on client financial records?
- It depends entirely on the vendor's handling. Ask where data is stored, who can access it, whether it is used to train models, and how access is logged. Under PIPEDA your firm remains accountable for personal information you hand to a processor.
This article is general information for bookkeeping professionals, not tax advice. Confirm current rules with the Canada Revenue Agency or a licensed practitioner before you rely on them for a client file.