Ask whether AI can do your bookkeeping and you will get two confident answers. One says the work is already solved. The other says nothing has really changed. Neither is much use when you are deciding who should keep your records next year.
The useful answer is narrower. AI is genuinely good at a particular shape of work, and genuinely weak at another shape, and the line between them runs straight through the middle of what a bookkeeper does. Knowing where that line falls is what lets you judge a service rather than a sales page.
What it handles well
The strength is pattern work at volume. Given enough history, a model is good at saying this transaction looks like the hundreds before it and belongs in the same place.
- Coding recurring transactions. The software subscription, the payment processor fee, the freight charge from the same carrier. These repeat, they look alike, and a machine applies the rule faster and more consistently than a person having a slow afternoon.
- Matching in reconciliations. Clearing the bank feed against the ledger is comparison at scale, which is exactly what software is for.
- Pulling data off documents. Reading a vendor bill for amount, date and line detail is now reliable enough to trust with a review step behind it.
- Flagging the odd one out. Noticing that an account moved differently this month than it has in the last twelve is a comparison, and the machine does not get bored.
If your records are mostly high-volume and repetitive, this is real. The work genuinely gets faster, and a service built on it can reasonably charge less for that part.
Where it still gets things wrong
The weakness is anything that depends on knowing something the ledger does not contain.
- The first time anything happens. A new revenue stream, a new entity, an unusual contract. Pattern matching needs a pattern, and there is not one yet.
- Cutoff. Whether a cost belongs to the month it was paid in or the month it was incurred in is a judgment about the underlying agreement, not about the transaction description.
- Accruals and deferrals. The same payment is handled differently depending on what was promised and when it was delivered. That lives in the contract and in somebody's memory of the conversation.
- Anything with two defensible answers. Capitalize or expense. One entity or another. Models are confident on both branches, and confidence is not accuracy.
None of this means the tooling is bad. It means the hard part of the work was never the typing.
The part that decides it
Here is the question that matters more than any capability list: when it is wrong, who finds out, and when?
A miscoded transaction costs almost nothing to fix in the month it happens. The same error costs a great deal when it surfaces in a diligence request, a lender's review, or a return that has already been filed, because by then it has been repeated eleven more times and the numbers people trusted were built on it.
Automation changes the rate of errors. It does not, on its own, change who is responsible for catching them. If a service automates the recording and leaves the checking to you, it has not removed the work. It has moved it onto the person least equipped to do it, which is usually the owner.
Speed on the recording step is only a saving if somebody else still owns the review step.
How to tell the difference before you buy
These questions separate a service that has genuinely absorbed the work from one that has automated the easy half.
| Ask this | What a good answer sounds like |
|---|---|
| Who reviews the automated coding, and how often? | A named role, on a stated cadence, not "the system learns" |
| What happens when something new shows up? | It routes to a person, and they tell you about it |
| Who decides cutoff and accruals? | Somebody who has read your contracts |
| If an error reaches the financials, who owns it? | A clear answer, in writing, before you sign |
| What do I have to do each month? | Specific and short, or the saving is not real |
If the answers to the first four are all variations on the software, you are buying a tool and supplying the labor yourself. That can be the right call, especially early on. It is just a different purchase than it looked like.
Where this lands
For most growing companies the honest position is unglamorous. Use automation hard on the repetitive work, because it is good at it and it costs less. Keep a person responsible for the judgment, the exceptions and the review, because that is where the expensive mistakes live and because someone has to be accountable when a number is questioned.
The question worth asking a prospective provider is not whether they use AI. Most of the tooling the industry runs on now has it built in somewhere. The question is which half of the work they are taking responsibility for, and whether they will put that in writing.