Finance software is moving from telling operators what happened to telling them what to do, and the shift is real, but it is only real for products that sit on transaction-level data. Everything else is a chatbot with a P&L attached. That is the position of this piece. The evidence is in what the vendors themselves shipped in 2025 and 2026, and in the one prerequisite none of their marketing leads with.
Reporting was the whole category for forty years
Accounting software has always been a reporting machine. QuickBooks, Xero and their predecessors take transactions in and produce statements out: a profit and loss, a balance sheet, an aging report. The user’s job was to read the report and decide. Dashboards, the big idea of the 2010s, were reporting with charts. Cash flow forecasting tools were reporting projected forward. The unit of output never changed: a number, presented, for a human to interpret.
The problem with reporting is that it answers the question the report was designed for, and the operator’s question is almost always a different one. A P&L says profit fell 18 percent. The operator wants to know which six products caused it and whether to cut ad spend or raise prices. Getting from the first to the second has historically meant exports, pivot tables and an afternoon, which is why most sellers do it quarterly or not at all.
What changed in the last two years
Four products, checked on their own sites in September 2026, show the category moving up a level.
Xero’s JAX assistant, included with Xero subscriptions at no current additional charge, answers questions over the user’s own ledger (“show my gross profit trend for the past year”) and describes itself as offering “strategic decision support,” pulling in outside data to answer questions about benchmarks or loan rates. Digits ships Ask Digits on every plan and, on its top tier, an “Agentic Close” that flags anomalies before a human looks. Puzzle sells insights agents that write financial narrative and runway models, and close agents that explain variances. ConnectBooks, an accounting platform for marketplace sellers, has Crunch, an AI CFO in active beta whose page frames the change in one line: reports tell you what happened, and Crunch is meant to explain why and say what to do next.
Three of those four are still reporting, done conversationally. Asking “what was my income last six months” and getting a chart is faster than building the chart, and it is not a recommendation. The variance explanation Puzzle’s close agent produces is a better-written report. The question is which products cross the line into “do this.”
The line is a ranked recommendation
The clearest example on any of the four sites is Crunch’s Q4 storage decision. The seller asks which inventory is worth paying fourth-quarter storage fees to keep. The output is a table: each SKU, its units, projected Q4 storage cost, projected sell-through, and a call, hold, discount 12 percent, liquidate or remove, ranked by cash contribution and treating already-paid cost of goods as sunk. That is not a report. It is a decision, with the reasoning shown, that the operator can accept or override.
The other example on the same page is a profit-decline decomposition: 71 percent of an 18.4 percent decline traced to six SKUs, with ad spend and a fulfillment-fee band change named as the drivers. That is diagnosis rather than recommendation, but it is the step that makes a recommendation possible, and it is the step operators skip because it takes an afternoon.
Whether other vendors follow depends less on their models than on their data, which is the argument of the next section.
Recommendation requires transaction-level truth
A recommendation is only as good as the resolution of the data it reasons over. “Cut ad spend on SKU 4471” requires knowing ad spend per SKU, margin per SKU after fees, return rate per SKU, and the cost layer the next units will carry. A ledger that holds one “Amazon fees” line and a monthly COGS estimate cannot support that recommendation, and an AI asked to produce one from it will produce a confident guess.
This is why the products closest to recommendation are the ones sitting on the most granular books. Crunch runs on settlement-level entries, per-unit FIFO cost and per-channel profit that ConnectBooks already posts, and the company’s own writing says that the AI reads books, it does not fix them. Digits and Puzzle automate the close first and put the question box second, in that order, for the same reason. Xero’s JAX is careful to describe decision support rather than decisions, which for a general ledger serving every industry is the honest scope.
The implication for buyers is unfashionable: the AI layer is the last thing to evaluate. Ask what the books underneath it contain. If cost of goods sold is a monthly plug, if marketplace deposits are booked as revenue, if fees are one line, no recommendation engine can help, and the vendor that promises one is selling the wrapper.
What this does to the accountant
The prediction that recommendation software replaces finance professionals is not supported by the data available. The Bureau of Labor Statistics’ Occupational Outlook Handbook entry for accountants and auditors, updated August 27, 2026, projects 5 percent employment growth from 2025 to 2035, faster than the 3 percent average across occupations, and states that automation is not expected to reduce overall demand but will make advisory and analytical duties more prominent. Read alongside the products above, that is exactly what the software is doing: automating the categorization and the variance explanation so the person’s time moves to the decision.
Every one of the four vendors keeps a human in the loop by design. Puzzle’s guarantee assumes a person still reviews. Digits routes CFO services to an accountant directory. Crunch hands a ranked table to the seller. The recommendation is an input to judgment, not a substitute for it, and the vendors who say otherwise are ahead of their own products.
Where this goes
Three predictions, stated so they can be wrong.
First, within two years every mainstream ledger will ship a conversational query layer and most operators will stop building their own reports. That part is already mostly true.
Second, ranked recommendations will stay concentrated in vertical products, because verticals are where the data is granular enough to support them. A general ledger cannot know what an FBA storage fee band is; software built for marketplace sellers has to. The same will be true of restaurant, construction and healthcare accounting, each with its own vendor that understands the vertical’s unit economics.
Third, the competitive question shifts from “whose AI is smarter” to “whose books are more correct at the transaction level,” and the vendors who spent the last decade on reconciliation accuracy will find that the least glamorous part of their product is the moat.
The operator’s job in the meantime is to get the books to the resolution a recommendation needs, whether or not they ever buy one. A P&L that can say which six products caused the decline is worth having even if a human is the one reading it.