Every close-management vendor now demos an AI that matches transactions for you. Almost none of them will say, in writing, what confidence threshold it needs before it stops guessing and asks a human.
Watch enough reconciliation software demos and they start to look identical. A bank feed loads, the software proposes matches against the general ledger, most of the screen turns green, and the presenter moves on to the next slide before anyone can ask the only question that actually matters: what happens to the transactions that don't turn green.
That question is where the marketing stops and the product documentation goes quiet.
What "AI reconciliation" is actually being sold as
Strip the branding and the current generation of AI-assisted reconciliation tools is doing one of four things to every line on a bank statement: matching it exactly, matching it with some tolerance for amount or date drift, flagging it as one of several plausible matches for a human to pick from, or giving up and routing it to an exception queue. The first category was solved a decade ago. The interesting product decisions, and the ones vendors are vaguest about, live in the second and third.
Visual 1 — What gets automated, and where judgement still sits
Match type | What the AI does | Who's actually deciding |
|---|---|---|
Exact match | Amount, date and reference align precisely | The AI, fully — this was never the hard part |
Tolerant match | Amount or date within a set variance | Whoever configured the tolerance, months ago, for a different transaction volume |
Ambiguous match | Multiple plausible candidates surfaced for review | A human, choosing from a shortlist the model generated — the accuracy of the shortlist is the whole product |
No match | Routed to an exception queue | A human, from scratch, usually the least experienced person on the team by the time exceptions pile up |
How to read it: Vendor demos live almost entirely in row one. The product's actual value — and its actual failure modes — live in rows two and three, which is precisely where public documentation gets thin.
A tool that auto-matches 95 percent of transactions and silently mis-matches 2 percent of them is worse than a tool that auto-matches 80 percent and flags the rest, because the 2 percent doesn't announce itself. It shows up as a discrepancy in next quarter's audit, attributed to whoever happened to be reconciling that account.
NetSuite's own release notes are the honest version of this problem
NetSuite's 2026.2 release is a useful, unusually concrete example — not because it's worse than competitors, but because its documentation is detailed enough to show the pattern plainly. The update adds a Suggestions Workspace that surfaces ambiguous matches previously discarded, letting a reconciler compare up to five candidate matches and choose; payment matching that proposes exact matches between bank transactions and open invoices and then automatically creates and applies the payment; and an Intelligent Flux Analysis feature that drafts explanations for account balance variances, reducing manual narrative-writing.
What the release notes do not include, at any point, is a stated accuracy rate, a time-savings percentage, or a specification of where the exception-handling logic actually draws its line. The capabilities are described qualitatively — "recommending," "suggesting," "helping resolve" — language that is honest about what the tool does and silent about how often it's right.
That is not a criticism unique to one vendor. It's close to universal across the category in 2026, and it puts finance teams evaluating these tools in the position of every buyer of an unmeasured product: forced to run their own pilot to learn the number the vendor could have given them upfront.
What to ask a reconciliation-AI vendor that the demo won't tell you
What's the false-match rate on tolerant matches, and how is it measured? Not "high accuracy" — a number, against a defined test set.
When the model is wrong on an ambiguous match, what does the audit trail show? You need to be able to reconstruct why the AI suggested what it suggested, after the fact, not just who clicked accept.
Does the confidence threshold get retuned as transaction volume or vendor mix changes, or was it set once at implementation and left alone?
What is the actual exception rate in a live customer environment at your scale, not the pilot environment used in the sales deck?
What follows from this
Treat the exception queue as the product, not the leftover. The auto-match rate is the easy number to market. The quality of what happens after auto-match fails is the number that determines whether close gets faster or just differently slow.
Ask for the audit trail before you ask for the ROI case. A tool that can't show you why it proposed a match six months from now is a tool you can't defend to an auditor, however fast it made this quarter's close feel.
Budget for retuning, not just implementation. Confidence thresholds set against last year's transaction mix degrade quietly as the business changes — new vendors, new payment rails, new entities — and almost no vendor proactively flags that drift.
The reconciliation software category will keep selling the green screen, because the green screen is what closes deals. The number worth asking for is the one that describes what's happening off-screen, in the rows the demo never lingers on — and until vendors publish it voluntarily, the only way to get it is to build your own test set and measure it yourself.
Sources and method. A FinancyHub original. NetSuite 2026.2 feature descriptions — Suggestions Workspace, payment matching and auto-apply, Intelligent Flux Analysis, and the absence of published accuracy or time-savings figures — from NetSuite's own release documentation. The four-category match framework and the false-match cost argument reflect FinancyHub's independent analysis of publicly available close-automation product documentation across the category as of August 2026, not a formal benchmark study. Journalism, not procurement advice — nothing here is a recommendation to buy, renew or terminate any product. Corrections will be made openly on this article.
