Last updated: September 8, 2026
AI in accounting speeds up the close by handling pattern work: matching invoices, flagging outliers, drafting a first-pass flux narrative. It does not make judgment calls like materiality or revenue recognition. Those still need a controller, and most vendor pitches blur that line on purpose.
A controller I worked with last year ran month-end close for a $180M industrial distributor. Nine days, every month, and the worst two of them were never the numbers. They were the narrative. Finance leadership wanted three sentences on why gross margin moved 40 basis points. Three sentences. To get them she had to dig through a week of AP emails, a Slack thread with the warehouse team, and her own scrawled notes from a call with the freight vendor about a rate change nobody had flagged.
She started feeding that stack into an AI tool. Not the ledger. The mess around the ledger.
First draft of the variance narrative in about four minutes. Wrong in a couple spots. She told me that flat out, no hedging. Confidently wrong, actually, on one number it had misread from a scanned invoice. But close enough that editing it took twenty minutes instead of writing it from nothing over an afternoon. That’s the whole trick, and almost nobody explains it that plainly.
Here’s where I sit, so you can weigh what follows accordingly. I run an ERP and business systems consulting group, which means a reader who decides their close needs help is a reader who might end up on my calendar. This also runs on a staffing firm’s site, and KORE1’s accounting and finance staffing practice places a fair number of the controllers who end up doing exactly what I’m about to describe. None of that changes what’s below. Argue with the substance, not the byline.
Every Close Is Two Different Jobs Wearing One Title
AI in accounting means machine learning and language models applied to accounting data and documents to do narrow, specific jobs: matching a bank transaction to an invoice, flagging a transaction that breaks a pattern, drafting a first-pass explanation of why a number moved. It is not a replacement for the judgment calls a controller makes about materiality, estimates, or how a transaction should be classified.
One job is pattern work. Does this invoice match this PO. Does this transaction look like the other four thousand like it, or not. Has this vendor’s remittance amount ever been $4,212 before. Pattern work is exactly the kind of thing a language model is good at, because pattern work has a right answer sitting in the data already, and the model’s whole trick is finding it faster than a person scanning line by line.
The other job is judgment. Is this expense capitalizable or not. Is a 2% dip in gross margin a rounding artifact or the first sign of a supplier problem you need to escalate before the board meeting. Does this estimate for warranty reserve reflect what actually happened this quarter, or a number someone copied forward because nobody had time to revisit it. There is no pattern sitting in the data for questions like that. There’s a person who has to decide, then explain that decision to an auditor eleven months later.
Everything downstream of this article is really just: which pile is this task in, and does the tool in front of you actually know the difference.
| Close task | Pattern or judgment | AI reality today |
|---|---|---|
| Three-way invoice matching | Pattern | Mature. Most AP platforms do this today without a human touching a clean match. |
| Anomaly and duplicate-payment detection | Pattern | Strong. Gets better the longer it watches your specific vendor patterns. |
| First-draft flux and variance narrative | Pattern, with judgment editing | Genuinely useful as a first draft. Needs a human close read before it goes anywhere. |
| Bank and subledger reconciliation | Pattern | Mature for clean feeds. Manual banks and paper still break the automation. |
| Accrual estimates and reserve balances | Judgment | Weak, and should stay weak. A model has no accountability for the number it suggests. |
| Revenue recognition classification | Judgment | Assistive at best. ASC 606 edge cases need a person who owns the answer. |
| Materiality calls on flagged variances | Judgment | Not a model’s job. It can surface the variance. It cannot decide if it matters. |
What the Data Actually Supports
Ardent Partners’ 2025 AP benchmarking work puts real numbers on the pattern-work side. Best-in-class AP teams process an invoice for $2.78, against an industry average of $9.40, and their touchless processing rate runs 49.2% compared to an industry average of 32.6%. That gap is mostly automation maturity, and a meaningful chunk of the recent movement in it is AI-assisted matching and exception handling.
Fraud detection tells a similar story from a different angle. The ACFE’s 2024 Report to the Nations found 57% of organizations now use exception reporting and anomaly detection as a standard fraud-analytics technique, up from a much smaller base a few cycles ago. That’s the pattern side of the house maturing in real time, and it’s the part of accounting AI that has the least controversy attached to it. Nobody argues that flagging a weird transaction for review is dangerous. The danger shows up one step later, when someone lets the flag become the decision instead of the trigger for one.
Sixty-two thousand invoices a year for a mid-market distributor. Roughly what one runs through NetSuite in a normal year at that revenue band. At an average handling gap of six or seven dollars an invoice between manual and automated processing, you’re looking at real six-figure money sitting in AP alone before anyone touches the close calendar.
Where a Controller Still Has to Show Up
The AICPA and CIMA’s Future-Ready Finance survey, 1,735 executives across eight regions, found 88% of finance professionals believe AI will be the most transformative trend in the profession over the next two years. Only 8% feel their organization is very well prepared to manage it. Fifty-six percent named generative AI specifically as their biggest skills gap.
That gap is not a training problem you fix with a lunch-and-learn. It’s a judgment problem. A controller who has never had to explain to an auditor why a model’s suggested accrual was wrong is a controller who is going to trust the model’s suggested accrual more than she should, right up until the year the number is material and the explanation is “the tool said so.”
I’ll go further than the survey does. Revenue recognition judgment calls, the honest hard ones under ASC 606, are close to the last place I’d want a language model making the call unsupervised. Not because the model is bad at language. Because a model has no skin in the outcome and no memory of the client conversation from eight months ago that actually explains why this contract’s variable consideration is weird. You do. Not a knock on the tool, to be clear. It’s just a description of what judgment actually is, and judgment is the part with your name on it.
Materiality works the same way. A model can tell you a number moved. It cannot tell you whether your board cares, because caring is contextual, and context is the one thing none of these systems actually have.
Deployment Isn’t the Same Thing as Value
Gartner’s November 2025 survey of 183 CFOs and senior finance leaders found 59% of finance departments now using AI in some form, up from 58% a year earlier. Barely moved. What did move is where it’s landing: knowledge management at 49% adoption, accounts payable automation at 37%, error and anomaly detection at 34%. Ninety-one percent of respondents reported only low or moderate impact so far.
Read that number twice. Nearly six in ten finance departments are using AI. Nine in ten of them are getting mediocre results from it. That gap is not a technology problem. It’s a targeting problem, and it’s the exact same mistake I keep seeing with ERP AI features generally, which I wrote about at length in AI in ERP: What’s Real and What’s Hype. People point the tool at the department, not at the task.
NetSuite’s actual shipping AI tooling is narrower than the keynote slides suggest, and narrower is fine. SuiteAgents handle defined, bounded jobs like drafting a PO description or summarizing a saved search. The NetSuite AI Connector Service, the thing everyone’s been calling the MCP connector since SuiteWorld, lets an outside AI client query and write NetSuite data under your own role’s permission set, which I broke down in more detail in the piece on how that actually works. Neither one closes your books for you. Both are genuinely good at the pattern-work slice of close I described above, and useless outside it.
Nice, when it’s scoped right.
The Days-to-Close Number Everyone Still Quotes Is From 2017
You’ll see it in half the vendor decks selling you close automation: companies close their books in 6.4 days on average, top performers in 4.8. That figure comes from APQC’s General Accounting Open Standards Benchmarking survey, roughly 2,300 organizations. It’s a solid study. It’s also from 2017, before most of the automation and every bit of the current AI wave existed, and people keep quoting it as if AI’s whole job is shaving days off a baseline that predates AI entirely.
That’s the wrong frame. The honest version of what’s shifting isn’t days-to-close, at least not primarily. It’s where the hours inside those days go. A controller who used to spend two of nine close days chasing down flux explanations manually is now spending forty minutes editing a drafted explanation instead. The calendar might not move much. The mix of what she’s doing with the time inside it moves a lot, from typing to reviewing, which is a better use of someone who passed the CPA exam.
If your AI vendor’s pitch leads entirely with a shorter close calendar, ask them what specifically got faster. If the answer is vague, that’s the tell.
Where to Actually Start
Don’t start with a platform decision. Start with a task inventory, the same way I tell people to approach practical AI for business operations generally. Pull your close checklist and mark every line pattern or judgment. Then:
- Pick the single most annoying pattern task on the list. Usually it’s flux narrative drafting or AP matching exceptions. Not the scariest task, the most annoying one.
- Check what your existing ERP or AP platform already does out of the box before buying anything new. Most companies are paying for capability they’ve never turned on.
- Run it in parallel with the manual process for one close cycle. Don’t cut over on faith.
- Measure minutes saved on that one task, not some vague productivity number nobody can defend in a budget meeting.
- Only then expand to a second task, same method, no shortcuts.
Financial reporting and dashboarding is the natural next stop once the close mechanics are handled, and I wrote a fuller playbook on that specific layer in Financial Reporting and Dashboards in NetSuite if that’s where you’re headed next.
If the honest answer is that your close is slow because the seat doing the judgment work is empty, that’s not an AI problem, and no amount of tooling fixes an unfilled controller role. KORE1 fills accounting and finance searches like that, direct hire or interim, in about 17 days on average, with 92% of placements still in the seat a year later. Their fractional CFO services cover the interim version if you need the judgment before you need the hire. If the crunch is seasonal rather than structural, Jennifer Burdick covers when to bring in contract accountants for year-end close.
Questions I Get Asked on Calls About This
Is my close calendar actually going to get shorter?
Some, eventually, but don’t budget for it. The bigger win in year one is almost always hours reallocated inside the same calendar, not a shorter calendar. Companies that promise a dramatically compressed close from AI alone are usually also selling the software.
What’s the cheapest place to start if I have basically no budget?
Whatever’s already turned on in the tools you’re paying for right now. NetSuite, QuickBooks Advanced, Bill.com. Most of the major platforms shipped some flavor of AI-assisted matching or drafting in the last eighteen months. Check the settings menu before you sign anything new.
Will this replace my staff accountant?
Probably not this year, and if it does, it’s replacing the part of the job that was already the worst part to do. The staff accountant who only does data entry has a real problem, but that problem started before AI showed up. The one who can review a drafted narrative and catch what’s wrong with it becomes more valuable, not less.
How do I know if a vendor’s “AI-powered close” claim is real?
Ask which specific tasks it touches and get a demo on your own messy data, not their clean sample set. If they can’t name the tasks in plain language, things like matching, flagging, or drafting, they’re selling a slide, not a feature.
Should our accrual estimates ever be AI-generated?
A model can suggest a number based on trailing patterns. A person has to own it. Write that ownership down explicitly rather than letting it drift, because “the system suggested it” is not an answer your auditor will accept, and it shouldn’t be one you accept from yourself either.
What’s the actual risk if we get this wrong?
Restating a number after the board already saw it. That’s the risk that should scare you more than a slow close ever did. Slow is annoying. Wrong and confident is the version that costs someone their job.
The Part Worth Remembering
AI in accounting is not one thing. It’s a pile of narrow tools, some of them genuinely good, aimed at a process that has always been two jobs stitched together and called one title. The pattern half is getting faster this year in a way that’s real and measurable. The judgment half isn’t going anywhere, and the companies getting hurt right now aren’t the ones moving too slowly. They’re the ones that stopped telling the difference between the two.
Every business I’ve worked with is underutilizing the tech it already owns. This is just the current version of that same old problem wearing a new acronym.
If you want a second opinion on where your own close checklist splits, or you’ve worked out that the seat that should own the judgment calls doesn’t exist yet, talk to a KORE1 recruiter about filling it, or hit me up on LinkedIn. Send me your close checklist. I’ll tell you which lines you’ve got in the wrong pile.
