Everything in this category now says AI. The useful question is not whether a product uses it, but what happens between the software deciding something and that decision becoming true in your ledger.
Every product in this category now says AI. The word has stopped carrying information.
What still carries information is the step nobody advertises: what happens between the software deciding something and that decision becoming true in your general ledger.
Some products write straight through. The model categorises, the entry lands, and the ledger looks identical whether a machine guessed or a person thought about it.
Some hold entries for review, so there is a state between proposed and posted.
Some put a human in the loop and charge you for them — Zeni’s dedicated controller, Pilot’s bookkeeper from the Core plan. That is the expensive answer and it is honest about what it is.
You will usually have to ask which one you are buying.
The failure mode is not a model that is wrong often. It is a model that is wrong plausibly.
We wired an agent to our own QuickBooks and asked it to match three months of Shopify Ads receipts to card transactions. It matched the first receipt to the wrong charge, then — being consistent — matched the next receipt to the next wrong charge, and the next. Ninety receipts ended up one slot off.
Every individual match looked reasonable. QuickBooks accepted all of them. The only thing that objected was the credit card reconciliation refusing to balance, weeks later.
A more accurate model would have made that failure rarer and not different in kind. What was missing was a moment where a person said yes, and a link between each receipt and the posting it supported.
Does a posting reach the ledger before a human approves it?
Is the source document attached to the posting itself, or filed near the transaction and matched by date?
Six months from now, can I tell whether an entry was proposed by software or considered by a person?
Most products in this category struggle with the third one, and it is the question an auditor, a lender, or a buyer eventually asks in some form.
Filing. Kick tells you to find a local accountant. Pilot, Fondo, and Zeni price tax as a separate annual line.
The categorisation is what got automated. The judgment and the return are still somebody’s job, and the honest products say so.
Best for: The genuine article at the low end. Auto categorisation, receipt matching, free below 250 transactions a year, and MCP and CLI access from $40 a month.
Watch out: Accrual is on the $100 Plus plan and no tier files your return — they tell you to find a local accountant.
Best for: AI with a human layer bolted firmly on top — daily automated reconciliation, plus a dedicated controller and analyst on every tier.
Watch out: $494 a month billed annually at the pre-revenue tier, and the bundled bank account makes leaving harder than switching software.
Best for: Founders who want the AI tier cheap. Essentials is $99 a month, explicitly software with no human staff.
Watch out: That tier is cash-basis only, and the human review people assume they are buying starts at Core.
Best for: Automated books and month-end close from a ledger its makers describe as AI-native.
Watch out: Public material does not document multi-currency or coverage outside the US. Ask directly if either applies.
Best for: US startups wanting an agent-native ledger rather than an older one with automation layered on.
Watch out: Built around the US market — confirm foreign currency handling before committing.
Best for: Companies that want the automation and also want every posting to wait in a draft state until a person confirms it, with the document attached to the posting itself.
Watch out: Ours, and that confirmation step is not optional. If you want books you never look at, we are deliberately the wrong shape.
What happens after the model decides. Some write straight to the ledger, some hold entries for review, some put a human in the loop and charge for them. That step is rarely on the pricing page and it is the one that determines whether a wrong categorisation is caught in a minute or found in March.
It is accurate enough to be dangerous. In our own books an agent matched three months of Shopify Ads receipts one slot off — the first match was wrong and every match after it inherited the error, ninety in total. Everything looked fine. The credit card reconciliation was the only thing that objected.
Three things. Does a posting reach the ledger before a person approves it? Is the source document attached to the posting or filed near the transaction? And can I tell, six months later, whether an entry was proposed by software or considered by a person? The last one is the hardest for most products to answer.
Not at year-end, and most of these products say so. Kick recommends partnering with a local accountant. Pilot, Fondo, and Zeni price tax separately. The categorisation is what gets automated; the judgment and the filing are still someone's job.
Claims about other products last verified August 25, 2026