We surveyed and interviewed 26 bookkeepers about one month-end close they had just finished, and went through it task by task. All of them run bank feeds, bank rules and a modern ledger. Automation touched 30% of the transactions. Out of 323 answers across sixteen tasks, not once did anyone let the software's output through without reading it. And 22 of the 26 fix an earlier month's error entirely by hand, with no software anywhere in the process.
The interesting part is why. Nobody said the software was not capable enough. What they described, over and over, is software that never gets told anything about the client it is working for, and that only runs inside the ledger while 70% of the month happens somewhere else.
Nobody here is behind on tools. This is the part nobody has built for yet.
Everyone we spoke to runs an automated stack. We asked how much of the month's transactions any of it actually touches. The answer came back at 30%.
The stack is bank feeds and bank rules. After those two it drops off fast: 7 of 26 run receipt capture, 5 run any AI tool, 3 run a reconciliation tool beyond the ledger. Two run nothing but the ledger itself.
Coverage is self-reported, median 30%, range 5% to 100%, n=24. Stack was a multi-select, so the counts do not add up to 26.
For each of sixteen tasks we asked them to pick one of four: I do it by hand, software does it and I check every time, software does it and I just accept it, or it never comes up in my work. That gave us 323 answers out of a possible 416, the rest being tasks that never come up for that person. The third option was picked zero times.
Coding is the most automated task on the list, and this is as good as it gets: 20 of 26 let the software do it and then check every single transaction, 4 still code by hand, and nobody lets it through unread. Everything else is more manual than that.
At the other end, 22 of 26 fix an earlier month's mistake by hand, and not one person has software doing any part of it. It is the only task here with no automation anywhere.
We also asked what would be left if coding became perfect tomorrow. 21 of 26 said review. Automating the draft does not get rid of the reading, because the reading is what the draft creates. Today's tools hand over a half-finished month and a person has to go through all of it to find the parts that are wrong.
We gave them the same twenty tasks twice and asked opposite questions: where does automation work, and where does it fall apart. All 26 answered separately. Not one task ended up on both lists.
Everything on the left is a repeat of something it has already seen. Everything on the right needs something the transaction does not contain: this client's history, this client's exceptions, what the owner actually did, what the number was supposed to be.
Worth being precise about what that means. None of it is unknowable. The history is sitting in last year's ledger, the exceptions are in the bookkeeper's head and in the client's email, and what the owner actually did is a question somebody can answer. It is all information that exists. It just never reaches the software.
A third question puts the same line at the level of a single transaction. We took fourteen categories and asked twice: which ones do you barely look at, and which ones do you check every single time.
Same person, same tool, opposite behaviour, and the only thing that decides it is whether they have seen that vendor before. Trust has nothing to do with the tool. It has to do with whether the transaction has a history.
The same line shows up in why the software leaves a transaction blank in the first place: a vague description, 18 of 26, and a new vendor, 16 of 26.
Every close has a set of entries nobody has automated. They are the same entries in nearly every book and they show up on the same schedule every month. An accrual in January looks like an accrual in February. Both get typed in.
Accruals, depreciation, deferred revenue, owner reclasses and the payroll journal are the top five, and they come back every month in the same books. The most predictable hour of the whole close is still being typed in by hand.
Predictable is the word that matters. These entries do not need judgement about a vendor nobody recognises. They run on a schedule, they follow the same logic every month, and the person typing them already knows what the answer is going to be.
The other standing cost is a new client, and it is paid before any of the above begins.
A new client stays slow for two or three months for exactly the reason the last section lays out. For those months there is no history on this client's vendors, and history is the only thing automation works on.
We asked them to split their close time across seven screens. The slider was locked to 100%, so nobody could inflate one number without taking it off another.
The ledger is 30% of the close. The other 70% is spreadsheets, receipt tools, PDF statements, portals and email. Every product in this category automates inside the ledger.
This 30% is not the same 30% from section 01. That one was the share of transactions automation touches; this one is time spent on a screen. Two questions with nothing to do with each other, and they landed on the same number.
So a tool that made the ledger portion free would still leave 70% of the month standing. Not because that 70% is hard, but because it is spread across six places nobody has connected to each other.
The files are small but they are not simple. A median of 132 transactions coming in across 7.5 accounts and 5 different kinds of document. Bank statements and receipts show up for 24 of the 26, aging and card statements for 21. The work is getting separate sources to agree, not getting through the lines.
These are per-category medians, so they do not add up to 100. Each person's own seven shares were locked to 100. Ledger time ranged from 0% to 69%.
We asked each person to split the close twice: how it looked before their automation arrived, and how it looks now. So everyone is being compared against themselves, not against anyone else.
One of the five went down. Coding dropped 9 points, and it dropped for 15 of the 26 people individually, so the median is not hiding a split. Nothing else went down. Reconciling went up for 10, month-end entries for 12, review for 13.
We asked separately which tasks had grown. 20 of 25 said review and 18 said answering the client's questions. Coding came last, at 7. A slider and a tick-list, two completely different question types, saying the same thing.
Where the saved time went: 11 work fewer hours, 8 took on more clients, 4 spend it on the same client, and 3 say nothing was saved at all.
16 of the 26 use an AI tool at least now and then, and 7 use one every day. The eight who never have, or who tried once and stopped, turned out to be the eight least automated people we spoke to. No exceptions either way.
The top two are the same complaint twice. It did not know the client's rules and exceptions, 13 of 19, and it could not see the client's history, 10 of 19. We also asked it as an open question and coded the answers separately, and missing client context came out on top again, 10 of 21.
One person said it was too slow. One said setup cost more than it saved. Nobody said the model was not smart enough.
Which fits how they are actually using it. 10 of the 21 who use AI open it in a separate window and ask it questions; only 4 have it wired into the books at all. Two told us they deliberately held client data back, because the tool was a personal subscription with no business-level data controls, and then checked everything by hand afterwards. The tool was set up with no access to the client, and then judged on not knowing the client.
"Some AI suggesting what to categorize or adjust in client books, but they don't know how they operate, they don't know the reason behind all transactions. Unless you provided them the details, you still need human assistance to identify the right answer."
S8 · runs bank feeds and rulesWhat is stopping them is not availability and not policy. Firm rules stop one person. Cost stops one. What stops nine is that they used it and watched it get things wrong.
Everything above is what people told us about their own close. What we want to know next is what the same work costs when someone is watching it happen instead of remembering it.
Everyone here has already automated whatever this industry knows how to automate. The feeds are connected, the rules are set, the ledger is modern. They are still typing in accruals by hand, still reading every line the software writes, and still fixing last month's mistakes with no software involved at all.
It would be easy to read that as this work being resistant to automation. That is not what the 26 of them described. Nobody said the software was not capable. The accruals that get retyped every month are the same accruals. The vendors that get checked by hand are unknown for one month and then known forever. The context the tools keep missing is sitting in last year's ledger and in the bookkeeper's own head. And most of the people using AI at all are typing questions into a separate window, because nothing has been connected to the books.
The gap is not that this work cannot be automated. It is that 30% of it sits inside the ledger, that is the part the category has built for, and nobody has built for the rest.
Every public survey in this market is run by a software company. Intuit's reached 725 accountants, another reached 486, and they all measure the same thing: who uses AI, how much, whether it feels worth it. None of them asks where the month actually goes.
There is a reason for that. A ledger vendor cannot publish how much of the close happens outside the ledger.