Entries that sit outside the usual pattern are marked before the books are closed.
AI Anomaly Detection reads invoices, payments, journals and balances, and marks entries outside their usual pattern, whether from a typing slip, a wrong account or something odder. A bookkeeper reviews each mark. The books are never changed by it. It needs a ledger with steady habits, and a business whose entries are all one-off will see many marks that mean nothing.
Most ledger lines are dull. A utility bill posted to travel, an invoice numbered far out of sequence, or a balance that jumps on a Sunday hide among the thousands of normal lines. The reviewer cannot read them all. Checking a sample means hoping it was the right one, which is how a wrong code stays wrong for a whole year.
Most oddities are mistakes, not crime. A decimal in the wrong place, a vendor coded twice, or a depreciation run twice in one month looks as strange to the software as a made-up invoice would. GullySystem sets it up against your own ledger, with thresholds your accountant agrees, and each mark carries a plain reason a reviewer can accept or reject.
The problems this is meant to ease.
What AI anomaly detection does.
What this looks like in a working week.
The people who open this every day.
From first conversation to daily use.
Walk through the close
Sit with the accountant and follow one month end from invoice entry to statements, noting where wrong codes, doubles and odd balances have been found before. Those places get checks first.
Clean the ledger history
Export a year of journals, invoices, payments and balances, merge duplicate vendors and fix the old coding errors that your team already knows about. Dirty history teaches the wrong normal.
Trial on closed months
Run the checks over closed months and compare the marks with the errors your accountant corrected at the time and in later periods. Count real marks, noise and misses.
Check marks with the bookkeepers
Let the bookkeepers and the finance manager work a live list, accept or dismiss each mark with a reason, and change the thresholds that produce too many. Dismissals teach the tool what is normal.
Go live before each close
Switch on the list of open marks as a step in the month end checklist, with alerts to named reviewers, and keep every correction inside the ledger itself. The software edits no entry.
Review after a few closes
After a few closes, the finance manager counts the marks that led to corrections, the marks dismissed and the alerts that were ignored. Thresholds are retuned afterwards.
How you would know it is working.
Record where you stand before it goes live, so the change is measured against your own figures.
- Useful alert rate is the share of marks that reviewers judged worth checking instead of dismissing, read from the first month of the trial, before any threshold was changed.
- Investigation time is the days from a mark appearing to a recorded decision, with the baseline read from how long bookkeepers took to chase similar queries before go-live.
- Confirmed anomalies is the count of marks that reviewers confirmed as real errors or odd entries, compared with the errors your accountant corrected in the closed periods.
- Duplicate payments caught is how many repeated invoices or payments were held before release, counted against duplicates found later in a review of the year before go-live.
- Open marks at sign-off is the number left unresolved when the close is signed, compared with the unresolved queries recorded on the checklist for the last close before go-live.
Where a person stays in charge.
A mark is only a question
A mark is not a finding. Most oddities are honest mistakes, and a reviewer decides whether each one is an error, a normal event or something to raise with a manager.
Corrections stay with people
No correction is made by the software. Every correction is made by a person inside the ledger, where the audit trail records who changed which entry, when, and why the reviewer thought it was wrong.
Irregular books give noisy marks
Irregular trades produce noisy marks. A job-based business with one-off invoices has little pattern to learn, so thresholds start loose and a careful accountant may do better by hand.
Ledger data is sensitive
Journals show staff names and supplier terms. Only the reviewers you name receive marks, access follows roles in your ledger, and your security adviser decides where the data is hosted.
Rules differ by country
Accounting practice differs by country. Your own accountant decides what makes a code wrong or a period closed. Tax, legal and financial questions go to your advisers, not to this tool.
It is one part of a system, not an island.
A module earns its place by what it passes to the next one. These are the connections we set up most often.
- Accounting ledger
- Accounts payable and receivable
- Bank reconciliation
- Inventory valuation
- Month-end close checklist
Products that include it.
How we would put it in.
Questions owners ask about AI anomaly detection.
How is this different from fraud detection?
Fraud checks look for signs of intent. Anomaly checks cover honest error across the whole set of books, such as wrong codes, doubled entries and slips of the keyboard. Every mark is only a question.
How many months of books does it need?
A year of ledger history is a fair start, because it needs to see each month once to learn that December always looks different. Clean the worst of the backlog first.
What if my business has no steady pattern?
Irregular trades make for noisy marks. A job-based business with one-off invoices may do better with a monthly review by a careful accountant, or with thresholds loosened for the first few months of use.
Does it correct the entries it marks?
The software corrects nothing at all. A person makes each correction in the ledger itself, and the audit trail records who changed which entry, when, and why the reviewer thought it was wrong.
Is this accounting or tax advice?
Reporting practice is not the same in every country. Your own accountant settles what counts as a wrong code, never the software. Tax and legal questions go to your own advisers.
What does it connect to?
It reads your accounting ledger, payables, receivables, bank reconciliation and stock valuation, so what it checks depends on what each system can export. A file upload covers any gap.
Who sees the marks, and what about privacy?
Marks go to named reviewers only. Journals and vendor files can hold staff and supplier details, so access follows ledger roles, and your security adviser agrees where the data is kept.
How much effort does it take to start?
The work is mostly agreeing thresholds. Your accountant and finance manager decide what counts as odd for each account, the worst of the backlog is cleaned, and the first months go on tuning.
Businesses that ask for this.
Modules that work with it.
Tell us how you handle AI anomaly detection today.
A spreadsheet, a register or another system: say which, and we will tell you plainly what is worth changing.
- No obligation
- A reply within one business day
- Your details stay private
Start with your requirements
Seven short steps that tell us what you need. An engineer reads them before getting in touch.
Share your requirements