A bookkeeper opens the file on a Tuesday. The overnight categorisation ran, and it did fine. Forty transactions are sitting in the exception queue, and thirty four of them she can clear herself in under an hour.
What Firms Actually Use AI For
The other six need the client. A payment to a name nobody recognises. An invoice with no backup. A transfer that could be an owner draw or could be equipment.
Those six will take nine days.
Uku surveyed accounting firms across eight countries in May and June 2026. It is a small sample, dozens of firms rather than thousands, weighted toward small practices of one to ten people, with roughly six in ten respondents being partners, owners or managers. Treat the exact percentages as directional. The ordering is what matters, and the ordering is not subtle.
Seventy six percent use AI for writing and client communication. Fifty nine percent for research and problem solving. Twenty nine percent for meeting notes and documentation. Twenty four percent for data entry and reconciliation. Fifteen percent for reporting and analysis.
Read that list from the top and one pattern falls out: the further a task sits from the ledger, the more AI is being used on it. The actual bookkeeping is close to the bottom.
That is the opposite of how the category has been sold for three years.
The Task They Most Want Handed Over
The same survey asked which tasks firms would delegate to an autonomous agent. The number one answer, at 68%, was chasing clients for missing documents. Preparing client emails and updates came second at 59%. Flagging unprofitable clients or jobs came third at 44%.
Nobody's top request was categorise my transactions.
This is worth sitting with, because it is practitioners describing their own week rather than a vendor describing a market. The thing they want a machine to take is not the accounting. It is the part of the job that involves asking a person for something and then asking again.
Because the Bottleneck Was Never the Categorising
Look at the Tuesday again. The categorisation was minutes of machine time. Clearing thirty four exceptions was an hour of skilled work, and skilled work is what a bookkeeper is for.
The six that need the client are a different kind of item entirely. They are not slow because they are hard. They are slow because the answer lives in a person's inbox, and that person runs a business, and your email is somewhere below a supplier dispute and a payroll question in their day.
A firm can double the speed of everything on its own side of that line and the month still closes when the client replies. This is the arithmetic that makes so many automation pilots feel underwhelming from the inside. The measured task got faster. The elapsed time did not move, because the elapsed time was mostly waiting.
The consequence for firm economics is specific. Work in progress ages while a query sits open. Staff context switch back into a file they had finished thinking about a week earlier, which is a real cost nobody bills. And in a fixed fee arrangement, every extra round trip comes straight out of the margin on that engagement.
What Firms Will and Will Not Hand Over
The same survey is blunt about the limits. Sixty two percent say trust requires a human to approve anything before it is sent or filed. Fifty three percent want their data kept private and never used to train models. Zero firms reported already fully trusting AI.
Put those next to the delegation list and the position is coherent, not timid. Firms will hand over the chasing. They will not hand over the deciding. And they want the client's file to stay where they put it.
The Shape of the Fix
The principle: reduce the number of times you have to ask the client anything at all. Every question you can answer from documents you already hold is a question that does not go into an email and does not come back in nine days.
That is the half MetaWurks is built for. It ingests what the client has already sent, the statements, invoices, contracts and correspondence, and lets a bookkeeper query all of it in plain English, so an unrecognised payment is checked against what is already on file before anyone drafts a chase email. Role based access controls govern who can open which client's records, audit logs record who opened what and when, and documents ingested into the platform are not used to train models or exposed to other users.
It will not make a client answer faster. It reduces how often you need them to.
The pitch for AI in bookkeeping has been aimed at the ledger for years. The people doing the work keep pointing somewhere else.
Join the Conversation
Of the queries your team sent clients last month, how many could have been answered from documents the firm already had?