How Is AI Affecting Accounting?

The clearest shift is in data entry and reconciliation. Optical character recognition and large language models can now read an invoice, code it to the right account, and flag exceptions with less error than a person doing the same task by hand. Treewalk built an in-house accounts-payable agent, on a self-hosted automation platform with around a dozen decision points, that handles invoice coding and routes only genuine exceptions (a new vendor, an unusual amount) to a human. It runs across our client base today.

The pattern extends past AP. We use AI to pull audit-request samples and CRA-requested invoices directly out of QuickBooks, roll forward continuity schedules from a one-line instruction, and flag discrepancies in working papers before a reviewer ever opens the file. None of this replaces the reviewer. It removes the hours a reviewer used to spend assembling the file before they could start reviewing it.

Marketing and sales teams want more AI output because more content and more leads are generally good. Accounting doesn’t work the same way: nobody wants more accounting the way they want more marketing. Nobody wants a bigger general ledger or more journal entries. Accounting is a finite, rule-bound function, so AI’s effect is almost entirely on cost and speed, not demand. That means the value shows up as a lower cost to produce the same result, not as a bigger accounting department.

A single invoice carries somewhere close to 80 to 100 small human decisions: which entity, which period, which tax treatment, whether it’s capital or expense, whether the vendor detail even makes sense. You cannot paste a messy transaction into a general-purpose chatbot and get a usable answer. The tools that fail in this space are usually well-funded and know almost nothing about how accounting actually works. The tools that succeed are built by people who understand both the code and the accounting logic underneath it.

A quiet but real finding from our own rollout: the accuracy bar for a machine is brutal. Clients and reviewers are far less tolerant of a computer’s mistake than a person’s, even when the computer is objectively more accurate on average. That changes how conservative you have to build the guardrails.

Two other limits are worth naming plainly:

• OCR still struggles with mixed-format receipts, GST variations, and foreign vendors that spell their own names three different ways across documents. Manual entry sometimes really is faster on messy source data.
• New categories of risk appear that never existed when a person did the work by hand, including prompt injection hidden in invoice text and increasingly convincing fake-vendor phishing. That’s a real reason human sign-off stays in place for new vendors and unusual amounts.

The biggest barrier we’ve seen is not technical. A team member built an AP tool that was faster and less error-prone than the manual process, and staff kept doing it manually anyway, because the tool asked them to change a habit. Adoption only reached its potential once the workflow mirrored what people already did and a manual entry got flagged and questioned every time it happened. If you are evaluating any accounting AI tool, ask less about the model behind it and more about whether it will actually get used on a bad Friday afternoon.

New grads and current CPAs are right to ask whether the designation still matters. It does, but the job is moving. The future accountant looks more like a human orchestrator of digital workers: setting up the automation, defining what “correct” looks like, catching fraud and edge cases, and being accountable for outcomes a machine produced. Interestingly, review work, which relies on judgment over a small amount of data, may automate faster than junior-level data entry, which still requires ingesting large, messy context. That inverts the old assumption that AI comes for junior jobs first.