There is a lot of noise about AI in accounting. Most of it is wrong — or at best, half-true. Modern AI models are genuinely useful for finance teams, but only for a specific set of problems. Outside that set, claims about "AI automation" are usually marketing for things that worked before AI existed, or for things that don't really work yet.
Here's the honest picture, written from the perspective of someone who has built finance products with and without AI for over a decade.
Where AI clearly works
- Transaction categorization. Given a description, an amount, and a counterparty, modern models classify transactions to GL accounts with high accuracy. This was hard with rule-based systems; with LLMs and embeddings, it's effectively a solved problem on most chart-of-accounts schemas.
- Document extraction. Pulling invoice numbers, line items, totals, and vendor info from PDFs and images is now reliable enough to be the default — not a feature to be evaluated, but an expected capability.
- Anomaly detection in bank feeds. Spotting "this transaction is unusual for this account" is well-suited to AI because the patterns are statistical, and humans miss them. This is where bank-feed reconciliation gets dramatically faster.
- Narrative generation. Variance commentary, audit memos, executive summaries from structured financial data — AI handles the first draft well. A human still edits, but the blank page goes away.
- Reconciliation rule suggestion. "Based on how this matched last quarter, here's a rule." Useful as an assistant, not as an authority.
Where AI is overhyped
- Generating journal entries from scratch. AI can suggest entries from descriptions, but the moment the entry has any judgment component — accruals, fair value, impairment — you're back to needing a human accountant. Marketing pitches that imply otherwise are either over-promising or over-engineering.
- Replacing the controller. The controller's job is judgment under uncertainty with accountability. AI is a poor substitute for accountability — it has none.
- "AI close." Some products imply AI can close the books. They mean AI accelerates parts of close. The actual close requires sign-offs, reviews, and final calls that an AI can't make.
- Tax. Tax positions involve interpretation of law in context. AI can summarize positions, but it should not take them. The risk asymmetry is wrong for automation.
How to evaluate AI accounting claims
- Ask what specific narrow problem the AI solves. If the answer is "automation" or "accuracy," the vendor doesn't know themselves.
- Ask what happens when AI is wrong. Good products are designed around AI being wrong sometimes. Bad products pretend it isn't.
- Ask for the human-in-the-loop pattern. AI that nobody can review at the right step is AI you'll regret.
- Beware of accuracy stats without scope. "97% accurate" means nothing without knowing what corpus and what tolerance.
A working model for thinking about AI in finance
The useful frame: AI is best at the high-volume, well-bounded, low-judgment parts of finance — the parts that are repetitive and have clear right answers. Categorization. Extraction. Pattern-spotting. Drafting.
It is worst at the low-volume, weakly-bounded, judgment-heavy parts. Strategy. Policy. Disclosure. Anything with accountability attached.
Most finance work is somewhere in between. The art is knowing the difference and not pretending AI can collapse the distinction.
The honest answer to "should I use AI in my finance stack" is yes, but for the right things. The wrong things — the judgment, the policy, the controller's room — should stay where they are. The right things — the rote, the repetitive, the pattern-heavy — should move to AI as soon as it's reliable for them. That's already most of the way there.