I asked Josh how AI was going at his accounting firm. He listed what it does for him, and then summed the whole thing up:
It still feels like we are playing with sticks and rocks over here.
I laughed, and then I thought about how rarely you hear that from anyone with something to sell.
What it actually does for him
Worth being specific, because the honest version is neither breathless nor dismissive.
Cleaning up spreadsheets. Pulling time card data out of a PDF and into Excel. Taking photos of documents a client emails in, with handwriting on them saying which property an expense belongs to, and getting that into a spreadsheet with the date, vendor, amount, and location.
One story I liked: a client needed a bill of sale for a vehicle, but the bank would not release loan funds without one, and the seller would not sign one before an inspection agreement existed. A genuine knot. Claude produced a bill of sale the bank accepted, the funds released, the buyer got to go look at the truck.
That is real value. A person got to buy a vehicle.
And on the whole, in Josh's assessment, it's helpful but not as much as he would like, and it feels primitive.
Both things are true at once. We should be able to hold both.
The verification tax
There is a detail in how his firm uses it that explains a lot about why the gains feel smaller than promised.
When AI converts a document into a spreadsheet, someone checks the total. They know what the number should be, so they verify against it.
That check is cheap here because accounting has a built-in answer key. Totals must reconcile. If the sum is right, the parts are probably right.
Most fields do not have that. If AI drafts you a policy, or summarizes a call, or writes a customer email, there's no total to check against. You have to read the whole thing carefully, which can cost as much as writing it did.
This is the hidden variable in every AI productivity claim, and almost nobody accounts for it. The question is not how fast the tool produces output. It is how expensive verification is in your particular work. Where checking is cheap, gains are real. Where checking is expensive, the gains are mostly a transfer, from producing to proofreading.
The vendor gap
Josh's frustration was pointed and I think fair. He hasn't seen a vendor make the case for AI in accounting specifically. Plenty about custom builds and enterprise plans and what you could do. Very little of "here's the out-of-the-box tool that does this."
He could name a couple of exceptions. A 1099 system. Digits, which has an AI-integrated ledger. Ramp does OCR on invoices and pulls the details into a review screen, which he described as limited AI and very helpful.
Notice that the tools he actually values are narrow. They do one thing, in his workflow, well.
But what he wants does not exist yet. His words: he wants something crawling through the ledger.
Why the honest version is more useful
The overselling has a cost that the people doing it don't pay.
A professional tries the tool because a demo suggested it would transform their work. It does not transform their work. It cleans up a spreadsheet and needs checking. They conclude the technology isn't ready and stop looking for two years, during which it actually does get ready.
The honest pitch would have kept them. Here's a narrow thing it does well, here's what it costs you to verify, here is what it cannot do yet. That earns a second look. Transformation language earns exactly one look.
Where this leaves you
If you're early and disappointed, you aren't doing it wrong. You are in the sticks and rocks phase along with a very capable accountant who has been at it longer than you.
Two things are worth doing anyway.
Find your cheap-verification tasks first. The ones with an answer key, where you already know what right looks like. That is where the gains are real today.
And keep a short list of the thing you actually want, the one that doesn't exist yet. Josh's is an AI crawling his ledger with an approval screen. Knowing precisely what you're waiting for means you'll recognize it the week it arrives, instead of finding out two years later.
What would yours be?
From a conversation with Josh on J&C Unscripted.