Accounting AI and ChatGPT: What It Can (and Can’t) Do for Your Books

Used correctly, ChatGPT is a decent research assistant and a fast first draft generator. It can:

• Explain a standard or a term in plain language before you go verify it against CPA Canada guidance
• Draft a first pass at an internal policy, a client email, or a board memo
• Summarize a report you already produced and trust
• Brainstorm categories or approaches before a human makes the final call

What it should not do is touch your books directly. You cannot put something into ChatGPT and expect a perfect answer: real accounting work has to be built up, decision by decision, on a system that understands your data. That distinction, between a chatbot answering a question and a system that actually processes your transactions, is the one most of the “AI accounting” marketing glosses over.

A single invoice carries somewhere between 80 and 100 small decisions: which vendor is this, is the GST recoverable, does this belong to the current job or a different one, is the coding consistent with last month. A general-purpose chatbot has none of that context built in, and it will still hand you a confident-sounding answer even when it is wrong.

The hard part about building tools like this is the accounting logic, not the software. You cannot get 80% of the way there and call it done: the last stretch, where the tool actually has to be right on real transactions, is the whole job.

That is why well-funded AI startups that know the software but not the accounting keep stalling out, while firms that pair accountants with people who can build software actually ship something usable. It is also why the accuracy bar for a machine is brutally higher than for a person: a controller will forgive a junior’s typo far more readily than a tool’s.

Rather than plugging client files into a general chatbot, we build narrow, purpose-specific tools around a well-defined workflow, with a human still checking the exceptions. Our in-house accounts payable agent runs through roughly a dozen decision points on invoice processing and keeps error rates below what a human clerk produces on the same volume, with a person only pulled in for genuinely new vendors or flagged anomalies. It was built by accountants who learned to code, not by software engineers guessing at accounting rules, and in our experience that order matters: the accounting logic has to come first, with the software built around it.

The lesson that surprised us most was not technical. We built a better tool and, for a while, almost nobody used it. Staff kept entering invoices by hand even though the automated version was faster and more accurate, simply out of habit. Adoption only took hold once we mirrored the existing workflow rather than asking people to change how they worked, and started flagging every manual entry to ask why. As one of our team members put it during that rollout, adoption is the only metric that actually matters. A tool nobody uses does not save you anything.

General AI tools introduce failure modes that a human process did not have. Prompt injection is a real example: a vendor invoice with hidden white-on-white text can attempt to instruct an AI reader to change payment details. Fake-vendor phishing is also getting more sophisticated as it targets automated approval flows rather than a person who might notice something odd. That is precisely why we keep human authorization on new vendors and unusual amounts even inside automated processes. Removing every human checkpoint is not the goal. Removing the repetitive, low-judgment ones is.