One of these problems has a statutory deadline attached and one does not, which is exactly why firms mishandle the second.
Key takeaways
There are two different jobs here and automating them the same way is why firms feel permanently behind in February. Job one is the slips your firm prepares and files for business clients — T4s for their employees, T5s for their shareholders. Job two is the slips your firm collects from personal-tax clients so a return can be prepared. Job one has a hard statutory date and a penalty. Job two has neither, which is precisely why it slips.
For job one, automation means a client-by-client readiness board driven by the deadline. For job two, it means an expected-slip list per client, matched against what has actually arrived, with the gaps chased automatically. Both are ordinary workflow automation; the AI layer reads inbound documents and closes items.
Subsection 205(1) of the Income Tax Regulations requires returns under that Part to be filed with the Minister without notice or demand and, unless otherwise specifically provided, on or before the last day of February in each year, in respect of the preceding calendar year. That is the sentence behind every February scramble in Canadian practice.
Which returns? Section 200 of the Regulations requires an information return in prescribed form from every person who makes a payment described in subsection 153(1) of the Act — the remuneration side. Section 201 covers payments of dividends, interest and similar amounts to residents — the investment-income side. Those are the T4 and T5 families in their statutory form.
Subsection 205(3) of the Regulations prescribes the types of information return for the purposes of the late-filing penalty, and the list expressly includes “Statement of Remuneration Paid T4” and “Statement of Investment Income T5”. The penalty itself is in subsection 162(7.01) of the Income Tax Act: for a person who fails to file one or more prescribed information returns when required, a penalty equal to the greater of $100 and, where the number of returns is fewer than 51, $10 multiplied by the number of days the failure continues, to a maximum of 100 days. The bands rise with volume.
Read the provision itself before relying on the shape of it, because the bands and the prescribed list have been amended more than once. The point for a practice is structural rather than arithmetical: the exposure is per-filing and time-based, so the risk is concentrated in the small number of clients who send payroll data on the twenty-sixth of February.
Build one row per filing client with four states — data not received, data received and unverified, draft slips prepared, filed. Drive reminders off the state and the calendar rather than off a person’s memory, escalating as the last week of February approaches. The AI contribution is reading whatever the client sends (a payroll report, a spreadsheet, a scanned summary) and proposing which state it moves the client into, together with the discrepancies it noticed against last year’s figures.
What it must not do is file. Preparing and filing an information return on a client’s behalf is professional work with a name attached to it, and the review step is not optional because the draft was machine-assembled.
The reason personal-tax clients send slips in dribs is that nobody has told them what the complete set looks like. Generate that list per client from the prior-year return — every slip type that appeared last year is a candidate this year — and then treat the list exactly as an outstanding-items register: inbound documents are classified, matched to an expected slip, and the item closes. What remains open is the chase list, and the chase can run on a schedule.
Two refinements pay for themselves. First, flag the slips that will not exist this year (an account closed, an employer left) so the list does not chase ghosts — that requires one question to the client, asked once, in January. Second, hold the return rather than file early where a late-arriving slip is likely; an adjustment costs more than a week of patience.
A firm in Manitoba handles T4 and T5 preparation for roughly seventy owner-managed corporations and personal returns for many of the same families. It rebuilds February around two boards. The filing board shows every corporate client in one of four states with the last-day-of-February date fixed at the top. The collection board shows each personal client’s expected slips, generated from last year, with received items ticked automatically as documents arrive in the portal.
The manager’s week changes shape: instead of reconstructing who has sent what, she works the two shortest columns — clients in “data not received” inside ten days of the deadline, and personal clients with more than two slips outstanding. Track the count of filings completed in the final three days of February; that number falling is the whole point.
Can AI prepare the slips themselves?
It can assemble a draft from the client’s payroll data and flag differences against the prior year. A person reviews and files. Never describe the tool to a client as doing the filing.
What if the client sends payroll data in a different format every year?
That is the strongest single argument for document extraction here — and also an argument for asking, once, for a consistent export.
Does this replace our tax software?
No. This is the collection and readiness layer that sits in front of it. The compliance software does the return; the workflow decides whether it can be started.
A 30-minute call is enough to tell you whether AI pays for itself here.
