How to Clean Up Your GRNI Accrual With AI Before It Surprises You at Close

Across the 12,722 live controller, accounting manager, senior and staff accountant, accounting specialist and AP postings in our database as of November 2026, 2,851 mention accruals, about 22%, and only 72 mention goods received not invoiced or anything that sounds like it, which is under 1%. So nearly every finance team knows it books accruals every month and almost nobody writes down who owns the GRNI balance, and I'd bet that gap is exactly why that balance is the one that surprises people at close.

The posting counts come from Audit Friendly's own job data, the 80,000-plus postings we clean and de-duplicate, and I mention them because a job posting only names what a hiring manager thinks defines the work, and GRNI mostly lives in the part nobody defends. That makes it a very good place to point AI, because the task is matching, the matching is tedious as hell, and nobody will miss doing it by hand.

What is GRNI and why does it go wrong?

GRNI, goods received not invoiced, is the liability you book for things the receiving dock has accepted but the vendor hasn't billed yet, so your expenses and inventory land in the right period even though no invoice exists. In an ERP it's usually the balance sitting in an accrued receipts or unvouchered receipts account, and in a healthy company it flows in when someone receives a purchase order line and out when the matching invoice posts.

It goes wrong in boring, predictable ways. Receipts get entered against the wrong PO line, the vendor invoices a different quantity or price, partial deliveries leave a few units hanging for months, and invoices get coded straight to expense so the receipt never clears. Each one leaves a stale balance, and stale balances are where overstated liabilities and misstated margins come from. If you've ever opened the GRNI detail and found items from two fiscal years ago, you already know what I mean, and the nice thing is that most of those items are the same five patterns repeating.

Why is GRNI a good job for AI?

Because the hard part is reading and matching rows, and that is what a language model plus a bit of code does well. You have a receipts listing with PO number, line, quantity, amount and date, and an invoice listing with the same keys, and the question for each open receipt is whether it matches an invoice that posted, whether it's a timing item that clears next month, or whether it's stale and needs a person to chase it. A model can propose a category and a one-line reason for every row in minutes, where a person doing it from a pivot table takes a couple of days and tends to stop at the biggest lines.

The same logic is why I keep pointing people at the closely related tasks in AP, like automating three-way match and vendor statement reconciliation, where 213 of the 1,892 live AP postings we track mention three-way match and only 357 flag automation of any kind, about 19%. GRNI cleanup sits right next to both, and you can reuse most of the same data pulls.

How do you run a GRNI cleanup with AI step by step?

Start with an export, and keep it small enough to check. Pull the open GRNI detail as of the last close, with receipt date, PO, line, vendor, quantity, unit cost, extended amount and any matched invoice reference, then pull the posted invoices for those vendors for the last 90 days. A spreadsheet and a general-purpose model are enough to prove the idea before you buy anything.

Second, ask the model to sort every open receipt into four buckets: matched invoice exists and the receipt should have cleared, timing item expected to clear next period, price or quantity variance, and stale with no activity. Ask for a reason and a confidence level on each row, and tell it to quote the invoice or PO reference it relied on, which lets you spot-check quickly. Third, review only the low-confidence rows, the stale bucket and anything above your materiality threshold. Fourth, post the clearing entries or fix the receipts in the ERP yourself, because nothing here should post on the model's say-so, and write the corrections down so next month's run starts smarter.

The tools matter less than the habit. If you're looking for software that can do this natively, the accounting software directory is where to compare ERPs and AP platforms, and our guide to handing accruals and prepaids to an AI agent covers the audit trail side of the same idea.

What controls do you need so AI doesn't make GRNI worse?

Treat the model like a very fast junior who never gets tired and sometimes sounds sure when it's wrong. Three things keep that safe. First, keep a human sign-off on every entry, and keep the reviewer's name and date attached to the output, since a reconciliation nobody signed is worth nothing to your auditors. Second, freeze the input files, the prompt and the output for each month so you can show exactly what was run, and the COSO internal control framework is a decent yardstick for what documentation a control like this needs. Third, test the model against a month you already closed by hand, and see how many of its calls agree with yours before you rely on it for the live month.

I'd also keep the model away from the judgment calls. It can tell you a receipt is 14 months old with no invoice and a vendor who has since been paid for other POs, and it should not decide whether to write that liability off, because that's a conversation with purchasing, the vendor and possibly your auditors, and somebody with a name has to own the answer. The honest middle here is that AI will clear most of the mechanical rows and the remaining handful will still take real judgment, which is a fine trade when it gives you a couple of days back at close.

What I'd do

I'd run it once as a side experiment on last month's closed GRNI balance, with no pressure and no entries, just to see how many of the rows the model sorts the way you did. I'd pick the vendor with the messiest history, because that's where you learn the most, and I'd write down the five patterns that account for most of your stale balance, since turning those into rules is worth more than any tool you could buy. Then I'd put an owner on the GRNI account by name in the close checklist, since the 72-versus-2,851 gap in the postings suggests most teams haven't. If you want a model for how to structure the broader close, our piece on rolling reconciliations and AI month-end close is the pillar this one hangs off.

Frequently asked questions

What does GRNI stand for in accounting?

GRNI stands for goods received not invoiced. It's the liability recorded for goods or services received before the vendor's invoice arrives, so the cost lands in the correct period.

Can AI reconcile a GRNI account?

AI can do the matching and sorting well, by comparing open receipts to posted invoices and proposing a category for each row, but a person should review the output and post every entry. Judgment calls such as writing off an old liability stay with your team.

How often should you clean up GRNI?

Monthly, as part of close, because stale receipts compound and become much harder to trace once they age past a quarter or a fiscal year.

What causes a GRNI balance to grow?

Receipts entered against the wrong PO line, invoices coded straight to expense, price or quantity mismatches, partial deliveries and vendors who never send the final invoice are the usual causes.

Do I need special software to run GRNI with AI?

No, you can prove the idea with exports from your ERP, a spreadsheet and a general-purpose model, then look at ERP or AP platform features once you know the review step works.

Pick one closed month, export the GRNI detail tonight, and see how many rows you and the model agree on, because that experiment will tell you more about your close than another meeting about it.