AI for Small Accounting Firms: What It Actually Does, and What It Doesn't Yet
Most "AI for accounting" content is written by someone who has never sent a client invoice. This one isn't.
I've had the same conversation with my friend Josh, an accountant at a small firm, spread out over several episodes of our podcast, and the pattern that came out of it is more useful than any single vendor pitch I've read. This post pulls the whole picture together in one place, with links to the longer version of each part.
What it actually does well today
The honest answer, in Josh's words, is that it "still feels like playing with sticks and rocks." That sounds negative and isn't meant to be. Held next to what it actually does for him, it's a fair description of an early but genuinely useful tool.
Concretely: cleaning up spreadsheets, pulling time card data out of a PDF, and turning a photographed receipt with handwriting on it into a structured row with the date, vendor, amount, and which property it belongs to. Real time back, every week, on unglamorous work nobody enjoyed doing by hand.
The clearest single example is document work outside of accounting entirely: taking a stack of paperwork too long to read carefully and pulling out the specific numbers and "gotchas" a normal person would otherwise miss. The same shape applies to lease agreements, loan terms, and vendor contracts that land on a bookkeeper's desk.
Why the gains feel smaller than the marketing promises
There's a hidden cost in every AI productivity claim that almost nobody accounts for: verification.
When AI turns a document into a spreadsheet, someone checks the total, because accounting has a built-in answer key. Totals must reconcile, so checking is cheap. Most other work doesn't have that. If AI drafts a client email or summarizes a call, there's no total to check against, and reading the whole thing carefully can cost as much as writing it did. That distinction is the real reason two firms doing the same task can get very different value out of the same tool: one has a cheap way to verify, and one doesn't.
The actual blocker isn't the model
Ask what would let an AI touch the general ledger, and the honest answer isn't technical. It's trust: whether the tool can be trusted with the data, and whether a person can see what it actually did.
That decomposes into two specific requirements, not vague unease. The tool can't retain client data it was shown, because a careful person will eventually forget to redact something, and a policy that depends on perfect human vigilance forever isn't a policy. And someone has to be able to see the reasoning, not just trust the answer, which is the whole argument for an approve-and-audit pattern: let the AI do the full task, stage it, write down what it did, and require a human to approve before anything commits. Scoped so it can never reach a closed period, sent invoice, or anything already final.
The adoption gap inside every firm
Here's the pattern across nearly every small business I've talked to, not just Josh's: one or two people are quietly using AI, and everyone else is working the way they always have.
It isn't age and it isn't skill. It's time and disposition. Some people poke at unfamiliar tools; others put them down and return to work they already know how to do, because the work is urgent and the tool isn't. Neither is a flaw, but the gap compounds quietly, and most firms have no idea it's happening because nobody is measuring it.
The fix isn't a training program. It's making the one or two people already using it visible, giving everyone else a specific first task instead of a tool, and protecting the time to try it.
What this means if you run a small firm
If you've tried AI and been disappointed, you're probably not doing it wrong. You're in the early phase along with a capable accountant who has been at it longer than most.
Two things are worth doing regardless: find the tasks where you already know what right looks like, because that's where checking is cheap and the gains are real today. And be honest with your team about where the tool can't go yet, rather than either banning it outright or pretending it's ready for the ledger.
If you want a real number instead of a guess for what automation would be worth in your specific practice, the 9-minute UseCase Assessment computes price bands from your own numbers, not an industry average.