Glossary
AI in Accounting: What It Actually Does, and What It Doesn’t
AI in accounting means using machine learning and large language models to handle the repetitive, rule-bound parts of bookkeeping and reporting: reading invoices, coding transactions, matching payments, flagging anomalies. It is not a system that replaces professional judgment. At Treewalk, we build this technology in-house rather than buying it off the shelf, because Canadian tax rules and a small domestic market mean most American AI tools don’t fit our clients’ files.
We operate on a simple assumption: whatever this technology can do today is the worst it will ever be. That’s the premise behind everything below. The technology keeps improving. The question isn’t whether to use it, it’s how honestly you use it right now.
How does AI work in accounting?
Most working AI in accounting today is narrow and task-specific, not a general assistant you chat with. A tool reads a document, extracts fields, applies rules, and asks a human to confirm anything unusual. The hard part was never the software. A single invoice carries somewhere between 80 and 100 small judgment calls: which vendor, which GST treatment, which account, which period. You cannot paste an invoice into a general chat tool and get a reliable answer. You have to build the accounting logic into the system deliberately, which is why firms with deep accounting knowledge tend to out-build well-funded startups that understand the technology but not the work.
That’s the model behind our own accounts-payable agent, built internally by our accounting team. It runs on a chain of decision steps, keeps a human in the loop for new vendors and anything that looks like an error, and now operates at a lower error rate than manual entry across our client base.
Is AI going to replace accountants?
Not the profession. The transactional layer, yes, largely. In our view, it would be naive for anyone in the accounting profession to assume this work won’t be substantially reshaped by AI. Data entry, reconciliations, and first-pass categorization are already being automated at scale. What survives is the part of the job that requires judgment: knowing when a number looks wrong, deciding how to treat something ambiguous, and being accountable to a client or regulator for the answer.
This plays out differently than in marketing, where more output is usually welcome. Nobody wants more accounting the way businesses want more marketing. Nobody wants a bigger bookkeeping department. So AI in accounting drives cost and headcount down rather than expanding demand for the work itself, which is a very different adoption curve than most software categories.
Which jobs survive AI in accounting?
Three roles hold up well, based on what we’ve seen firsthand building this internally:
The reviewer and judgment-caller.
Someone has to decide whether an anomaly is fraud, an error, or a legitimate exception. That call still needs a person accountable for it.
The relationship and advisory role.
Clients hire a firm for a point of view on their numbers, not just clean data. That trust doesn’t automate.
The orchestrator.
Increasingly, the valuable accountant is someone who manages a set of digital tools and knows when to trust them and when to intervene, more like a supervisor of digital workers than a preparer of schedules.
What is the “30% rule” for AI in accounting?
There’s no single agreed definition here, and we’d be skeptical of anyone who states one as fact. The number people usually reach for is a rough estimate of how much routine transactional work, things like data entry, invoice coding, and basic reconciliations, is reliably automatable with current tools. In our experience, that’s roughly the right order of magnitude for what we’d trust unsupervised today. Automating judgment calls and anything client-facing is a much harder problem, and anyone claiming otherwise is probably overselling.
How we approach it at Treewalk
We learned the hard way that building the tool is the easy part. Our AP agent was faster and made fewer errors than manual entry, and staff kept entering invoices by hand anyway, because changing a habit is harder than changing a system.
We built a monitor that flagged every single manual entry and asked why, and adoption climbed toward full use almost immediately once people knew someone was watching. The lesson wasn’t about the technology at all. It was that a better tool nobody uses is worth nothing.
That’s why we don’t sell automation as a headline feature. We build it to mirror the existing workflow, pilot it with our most skeptical staff first, and only call it done once adoption is real, not theoretical.
What this doesn’t fix on its own
AI introduces risks that didn’t exist when a human did the work by hand: hidden prompt injection text embedded in an invoice, or increasingly convincing fake-vendor phishing designed to fool an automated approval. That’s why human authorization stays in place for new vendors and any amount anomaly, no matter how well the automation performs otherwise. AI lowers the error rate. It doesn’t eliminate the need for a human to own the outcome.
Frequently asked questions
Will AI replace my accountant or CPA firm?
No. It replaces the manual, repetitive parts of bookkeeping, not the judgment, review, and advisory relationship a firm provides. Expect fewer hours billed for data entry and more value placed on interpretation and oversight.
Which accounting tasks can AI already handle reliably?
Invoice coding, transaction matching, basic reconciliations, and flagging anomalies for review. Judgment-heavy work, like assessing an unusual transaction or advising on a structure, still needs a person.
Is it worth building custom AI tools instead of buying software?
It depends on how specific your accounting logic is. Off-the-shelf US tools often assume US tax rules and currency, which don’t map cleanly onto Canadian files. We build our own for that reason.
How is Treewalk different from a firm that just bolted on a chatbot?
We build the automation ourselves, accountants first, and we measure it by whether staff actually use it, not by whether the demo looks impressive.
What should I actually do about AI right now?
Start by identifying which of your own transactional tasks are rule-bound and repetitive. Those are automatable today. Anything requiring judgment still needs a person, and probably will for a while yet.
Where to next
If you want to talk through what AI can realistically automate in your finance function today, our private companies team is a good place to start that conversation.
