Open the bank page in QuickBooks or Xero and you'll find the word "AI" on almost everything. Intuit announced a "virtual team" of AI agents for QuickBooks in July 2025. Its help pages now call the bookkeeping one "Accounting AI" and promise "done for you" accounting. Xero calls its version, JAX, an AI "financial superagent".
Before you trust any of it with your books, there's a more useful question than "is it good?" It's this: how much of this is actually new, and how much is automation you've had for years with a new label on it?
That isn't pedantry. The old parts and the new parts go wrong in different ways, and they need different checks. If you can't tell which one coded a transaction, you can't tell what kind of mistake to look for.
Four jobs, one label
Xero is unusually transparent here. According to Xero's own help article, every transaction its automatic bank reconciliation handles "has the label Rule, Match, Memory or Prediction, to show the method JAX used." Its US product page lists the same four, as did Insightful Accountant's coverage when the feature launched. Other packages don't necessarily work the same way, and most don't show you. But Xero's four labels are a useful lens for thinking about whatever software you use.
Rule. One of your bank rules was applied: "anything from the landlord's account goes to Rent." This is an instruction you (or your bookkeeper) wrote. It isn't AI in any meaningful sense, and it isn't new. Accounting firms were publishing guides to QuickBooks Online's rules feature well before any of the agent announcements. One detail matters more than it looks: in QuickBooks, a rule "will always supersede" the AI's suggestion. An old rule beats a new suggestion, every time.
Match. The software paired a bank line with an invoice or bill that already exists in your books. A customer pays $1,850; there's an open invoice for $1,850; they get matched. Also not new in principle, though it has improved. QuickBooks now finds "partial and combined matches", such as one payment covering two invoices.
Memory. The software coded a transaction the way you've coded similar ones before. QuickBooks describes its suggestions as based on "your history and the transaction details," and says that when you correct one, "QuickBooks will learn for next time." This is pattern-matching on your own past behavior. It's more capable than a rule, but it's the same idea: do what we did last time.
Prediction. The software coded a transaction based on how other businesses coded similar ones. Xero's version, as Insightful Accountant described it at launch, draws on "similar transactions by all Xero users." QuickBooks says that when you have no similar past transactions, it can recommend categories "based on how other businesses have categorized similar transactions," but only when it's confident. This is the genuinely newer layer, and the one the marketing is really about.
Two more things sit alongside those four. Reading documents (pulling the vendor, date and amount off an invoice or receipt) is an old job too: Dext was called Receipt Bank until 2021, and Xero bought Hubdoc in 2018. The tools read messier documents than they used to, but the task isn't new. And the chat layer, where you ask your accounting software a question in plain English and get an answer, is new.
The real change isn't the technology. It's who presses the button.
Here's what the "is it AI?" framing misses. For years, bookkeeping automation suggested and then waited. A rule proposed a category; a person posted it.
The newer features increasingly act. Xero's automatic reconciliation reconciles transactions on its own "where there's high confidence." Its help article describes it as "a feature you can turn on or off for each bank account in your organisation." It also only looks forward: "JAX reviews all future statement lines that import into the bank account, it doesn't review historical statement lines." Switching it on won't quietly rework last year's books; it starts with the next import.
It launched in beta in November 2025 and Xero now treats it as shipped, though more complex cases, such as splitting one payment across sales and fees, are still "coming soon."
QuickBooks' Accounting AI puts transactions it's "very confident" about into a "Ready to post" list you can post "with a single click." It now also posts some matches by itself: payroll payments run through Intuit's own payroll products, and bills paid through QuickBooks Bill Pay. Those arrive with an "AUTO" badge.
Neither is reckless. Both show you what they did and let you change it. But the question you need to ask has shifted from "is this AI?" to something plainer: did a person look at this line before it went into the books?
Xero, usefully, answers that one on the screen. A line JAX reconciled carries one of the four method labels. A line a person reconciled carries a fifth: "Manual." That one label marks the difference this whole post is about.
One month of bank lines, sorted by who handled them
To see why the distinction matters, take an illustrative month for a 20-person service business: one checking account, one company card. This is an example, not data. The point is the shape, not the numbers.
The lines the rules handle. Rent. The payroll transfer. Software subscriptions. The loan repayment. The same payees for similar amounts every month. A rule codes them and, if you let it, posts them. The catch: if a rule is wrong, it's wrong every month, silently. And because a rule supersedes the suggestion in QuickBooks, a bad rule from two years ago will override a better guess from this year.
The lines matching handles. Customer payments against invoices you raised in the system; vendor payments against bills you entered. Usually easy. The risk is the customer with three open invoices, two of them for the same amount, where "matched" and "matched to the right one" are not the same thing.
The lines memory handles. Fuel, office supplies, the regular vendors you always code the same way. Memory copies your habits, including the bad ones. If your previous bookkeeper coded the accountant's fees to "General expenses," memory will faithfully keep doing it.
The lines left for prediction, or for you. The new vendor. The one-off equipment purchase. The refund. The expense the owner paid personally. The payment that's half loan, half interest. These are exactly the lines where your history is thin, so the software falls back on what an average business would do. QuickBooks' own review signals exist to flag a suggestion "based on very little history that needs your attention before you post." And an average business's coding isn't your accountant's. A $3,000 laptop that "most businesses" put under office expenses may be something your accountant wants recorded as an asset.
Look at where the genuinely new technology lands: on the smallest, most judgment-heavy slice of the month. That's where a wrong answer costs you at year-end. The label "AI" covers the whole bank page, but most of what makes the page feel automatic is the rules and matching you've had all along.
Some of those awkward lines are also where a general chat assistant earns its keep. A composite week inside a 25-person company that uses AI well shows a bookkeeper describing an awkward transaction and asking which category fits, then deciding for herself. That's a different job from letting the software post it.
The number nobody publishes
The figure you actually want is simple: how often is it right?
Neither Intuit nor Xero publishes it, at least not on the product pages, help articles and announcements covered here. Intuit's overview of its AI says it works "with best-in-class accuracy," and gives no number. What the two companies do publish is useful, but different:
- Intuit says its AI-powered bank feed can save businesses "up to 12 hours a month." The footnote explains that this is based on 45% of surveyed customers reporting that saving, in a survey Intuit commissioned in April 2025. That's a time figure from a subset of users, not an accuracy figure.
- When the beta launched, Xero said its goal was for JAX to automatically reconcile "more than 80%" of your bank statement lines. That's a target, not a measurement. Xero's current US product page doesn't repeat any percentage at all.
Neither claim is dishonest. Neither tells you the error rate in your books, with your vendors and your accountant's chart of accounts. You'll find benchmarks of document-reading accuracy online, but they test different documents and different measures from your bank feed. Don't borrow their numbers.
So measure your own. For one month, before you post anything, keep a tally: how many suggestions did you change, and which method made each one? In Xero this is quicker than it sounds, because hovering over a line's label shows you "the reasoning JAX used." You see not just that a line was wrong, but why. That's your accuracy figure. It takes an evening, and it will tell you more than any demo.
Then run the tally again the next month. Both platforms say they learn from corrections. Xero's help article says "When you fix a reconciliation, JAX learns from it," and QuickBooks says it "will learn for next time." The second month tells you whether that's true for your books.
Where each layer goes wrong, and the check for each
- Rules go wrong silently and permanently. Review your rule list every quarter. Delete rules for vendors you no longer use. Be especially wary of any rule set to post automatically.
- Matching goes wrong on look-alike amounts. Spot-check matches for customers and vendors with several open invoices.
- Memory inherits your mistakes. If your history is messy, correct the coding for a few recurring vendors and watch whether the suggestions follow. QuickBooks says it "will learn from your changes." Check that it does.
- Prediction goes wrong on anything unusual. Set a threshold: first-time payees, and anything above a figure you choose, get a human decision. Where it's genuinely unclear, that means your accountant.
- Auto-posting goes wrong at scale. If you turn it on, start with your simplest account. Xero lets you switch automation on or off per bank account, so use that. In QuickBooks, look at the "AUTO" badges in the Posted tab now and then, so you know what went in without you.
- None of it should touch payments. Categorizing a transaction is recoverable. Paying a fake invoice, or accepting a vendor's "new bank details," isn't. AI-forged invoices and payment-detail scams are exactly what a busy person approving a tidy-looking list will miss.
What to ask your accountant
Your accountant is the person who has to stand behind the year-end numbers, so bring them in before you switch on anything that posts by itself. Four questions cover it:
- Which kinds of transaction should the software never code on its own? Capital purchases, owner draws, loan movements and sales tax treatment are the usual candidates.
- Are you comfortable with auto-reconciled or auto-posted lines? And would you like to see which method handled each one? In Xero, every reconciled line carries a label: Rule, Match, Memory or Prediction, or "Manual" when a person did it. In QuickBooks, look for the "RULE" and "AUTO" badges.
- Which bank rules should exist that don't? Your accountant's chart of accounts beats the platform's idea of an average business every time. A rule they approve is worth more than a prediction.
- What shouldn't go into a general chat tool? Bank statements, payroll and customer details all carry obligations. It's worth being clear on the basics of AI and data privacy for a small business before anyone pastes a statement into a chatbot.
What it costs
Less than the marketing suggests, because most of it is already in plans you may be paying for. QuickBooks says the AI suggestions on its banking page are "available in all versions," and its Accounting AI is included with the Essentials, Plus and Advanced plans. Xero's automatic reconciliation is on its Growing plan and above in the US, and the equivalent tiers elsewhere.
So the cost decision is usually a plan tier, not a new tool. Separate tools still have a place: document capture like Dext and Hubdoc, and expense cards like Ramp, Brex and BILL Spend & Expense (which was called Divvy until September 2023). But they're worth adding only once you know which of your lines are actually costing you time. Count first, upgrade second.
This is the kind of workflow review we do with clients: sorting where the hours in the books actually go, and which parts are safe to automate.
The honest bottom line
Most of what's labeled "AI" on your bank page is rules and matching that have been around for years. That's fine; boring automation is the useful kind. The genuinely new parts are predicting from other businesses' habits, posting without asking, and answering questions in plain English. They sit on the smallest and riskiest slice of your month.
So treat the label as a prompt, not a promise. For any line in your books, ask two questions: which method handled this? and did anyone look at it? If you can answer both, you're using the technology well, whatever it's called.
If you'd like a second pair of eyes on how your books are set up, and what's safe to hand to the software, let's talk.