What AI Actually Changes in an Accounting Practice

Separating the signal from the sales pitch

Every vendor deck now says AI, which makes the real question easy to lose: what changes in the daily operation of a small Canadian firm. The honest answer is narrower than the pitches and larger than the scepticism — specific tasks compress dramatically while the shape of the practice barely moves.

A disclosure before the analysis: SpidNums, whose blog this is, does not ship AI features. That distance is useful here — this piece has nothing AI-flavoured to sell you.

Where AI genuinely helps today

The reliable wins are language and volume tasks: drafting client emails and engagement descriptions for human editing, summarising long documents, extracting fields from piles of slips and invoices, suggesting transaction categories for a bookkeeper to confirm. Work where a fast, imperfect first draft beats a blank page.

Notice what these share: low stakes per item, a human checkpoint built in, and volume that made the manual version tedious. That is the current honest frontier. Claims much beyond it — unsupervised filings, autonomous client advice — describe the demo, not the engagement file.

What AI does not change

The accountability structure. A professional still signs, professional standards and CRA obligations still apply, the client still trusts a person, and the deadline still falls on the same date whether a human or a model drafted the working paper. AI compresses effort; it does not absorb responsibility.

It also does not fix operational problems. A firm that cannot say who owns which deadline will not be saved by faster drafting — it will produce confusion at higher speed. Operational discipline is a prerequisite for AI leverage, not a casualty of it.

The staffing shift: review replaces production

As first drafts get cheap, staff time shifts from producing work to reviewing it — and reviewing machine output is a different skill from reviewing a junior's, because the errors are confident, fluent and randomly placed. Firms will need explicit review steps where apprenticeship once caught mistakes implicitly.

This has a training consequence worth taking seriously: if juniors no longer produce the tedious first drafts, firms must design other ways for them to learn what good looks like. The tedium was, quietly, the curriculum.

Confidentiality, PIPEDA and client data in prompts

The sharpest near-term risk is not bad output — it is staff pasting client financials into consumer AI tools with no firm policy in place. Client data carries confidentiality and privacy obligations, including under PIPEDA, that do not pause because the tool is impressive.

The baseline is a written policy before broad use: which tools are approved, what data classes may enter them, what the vendor's terms say about retention and training, and where the data is processed. The same questions a firm should ask any software vendor — asked before the habit forms, not after.

How to adopt AI deliberately

Pick one bounded workflow, write down the review step, run it for a quarter, then decide with evidence. Firms that adopt this way accumulate judgement; firms that adopt by letting each person freelance accumulate risk.

• Start where errors are cheap and visible: internal drafts, summaries, categorisation suggestions.
• Name a human owner for every AI-assisted output — the reviewer, not the tool, is accountable.
• Write the tool policy before the third employee starts using the second tool.
• Reassess quarterly: this landscape shifts faster than annual planning cycles.

The boring foundation AI does not replace

Whatever drafting looks like in five years, the operational spine of a practice stays: every client known, every deadline owned, every engagement documented. AI raises the throughput of the work; the spine decides whether the throughput lands anywhere.

That spine is the layer SpidNums builds — client records, service cadences, deadline feeds, task boards, e-signed engagement letters — with no AI in it today, deliberately. If AI halves the hours a T2 takes, the T2 still has a due date, an owner and a client who needs to be told. The firms best placed for whatever arrives next are the ones whose system of record already tells them, on any morning, exactly what is due and who has it.