Every booking reviewed, automatically.
Every ledger posting is reviewed by an agent against thousands of audited postings, the invoice register and the source documents. The feedback comes back as a short, evidence-scored queue with the correction spelled out, and it finds errors a manual review misses while the period is still open.
Working alongside the accounts payable agent, it confirms that each booking is backed by a real document and that the booking matches it on every attribute that matters for compliance: amount, date, VAT, cost center and the other legally required details.
- Function
- Finance operations, month-end close
- Solution
- An AI review agent on top of the ERP, with a dashboard for the reviewer
- Baseline
- Thousands of audited, compliant postings from years of bookkeeping
- Scale
- Hundreds of postings a month across dozens of vendors
The problem.
Every posting is made by an accountant and then checked by someone above them: a senior accountant, a chief accountant or a finance director, depending on the organization. That second look is what keeps the books clean and up to date, and it is the control against wrong bookings, missing documents and, in the worst case, financial misconduct.
Done by eye in the ERP, that check takes hours every month, depends on who is doing it, and still lets errors through to month-end close, when the period is locked and nothing can be corrected. An automated review does the same work quickly and consistently, and flags what needs a person's judgment.
The reviewer's real problem isn't finding something wrong. It is knowing which of several hundred postings deserves their attention today.
What we built.
The reviewer no longer scrolls through several hundred records. They work a short queue, and each item carries the specific correction and how much to trust it.
Pull
Every posting, with its full general-ledger detail, straight from the ERP.
Compare
Each booking against years of audited bookings of the same vendor, the invoice register, the source document itself, and the accountant's own review.
Score
Each finding carries the evidence behind it and a confidence weight. Weak signals become notes, not tasks.
Publish
One dashboard: fix first, safe to approve, raise with the accountant, and whether the month can close.
Cross-checks with the AP agent
Where the accounts payable agent runs too, it pairs each booking with the invoice behind it and flags anything that differs, from the amount to the cost center.
Learns the patterns
Vendor by vendor, it learns how bookings are normally made, so a deviation stands out without a rule being written for it.
At a glance.
The reviewer's queue.
Review queue
Current period, updated 09:30
- Postings checked
- 356
- Change required
- 8
- Ready to approve
- 343
- Notes
- 5
| Vendor | Amount | Review |
|---|---|---|
| Parking operatorThis document number is already booked for the same vendor last month. Either a duplicate, or last month's number copied onto a new bill. | €1,140.00 | Change required95% |
| Benefits providerBooked to the marketing cost center. Every earlier bill from this vendor went to HR, twelve months in a row. | €18.60 | Change required95% |
| Travel agencyBooked to prepaid expenses. The agreed treatment is to expense travel. | €1,972.35 | Change required90% |
| Cleaning servicesThe invoice behind this booking reads €604.00. Two digits transposed at entry; the accounts payable agent holds the document. | €640.00 | Change required90% |
| Office suppliesMatches the invoice register and past practice. | €287.90 | Ready to approve90% |
| New catering vendorNo precedent for this vendor. Worth a look, not a correction request. | €156.20 | Note40% |
- Parking operatorChange required
- Amount
- €1,140.00
- Confidence
- 95%
This document number is already booked for the same vendor last month. Either a duplicate, or last month's number copied onto a new bill.
- Benefits providerChange required
- Amount
- €18.60
- Confidence
- 95%
Booked to the marketing cost center. Every earlier bill from this vendor went to HR, twelve months in a row.
- Travel agencyChange required
- Amount
- €1,972.35
- Confidence
- 90%
Booked to prepaid expenses. The agreed treatment is to expense travel.
- Cleaning servicesChange required
- Amount
- €640.00
- Confidence
- 90%
The invoice behind this booking reads €604.00. Two digits transposed at entry; the accounts payable agent holds the document.
- Office suppliesReady to approve
- Amount
- €287.90
- Confidence
- 90%
Matches the invoice register and past practice.
- New catering vendorNote
- Amount
- €156.20
- Confidence
- 40%
No precedent for this vendor. Worth a look, not a correction request.
Month over month.
Volume and value
Counted on the transaction date
Bookings received
All transaction types, by month
Vendor-bill value
Bills and credit notes, EUR
| Month | Bookings | Bill value |
|---|---|---|
| Mar | 274 | €51,380 |
| Apr | 329 | €64,920 |
| May | 302 | €58,910 |
| Jun | 347 | €70,180 |
| Jul | 384 | €81,050 |
| AugIn progress | 356 | €46,905 |
Confidence, not alerts.
Most tools that check your data fail the same way: they produce a flat list of flags, a third of them are wrong, and people stop reading. Here every finding is scored on the strength of the evidence behind it, not on how confident the model sounds.
- A reviewer's own verdict on the record100%A person read the actual record, including fields the data extract cannot see.
- A deterministic house rule95%Policy, no inference: a missing cost center, a rejected document.
- Invoice matched on document number90%Then compared field by field against the source document.
- Invoice matched on vendor, amount and date65%Can pair the wrong month's invoice from a vendor who bills the same figure every month.
- Deviation from past practice40–95%Rises with how many earlier audited bookings support the expectation. One is an anecdote; six is a convention.
- No precedent at all40%A new vendor or a new pattern. Worth a look, not a correction request.
The weights are editorial: they rank how much a reviewer should trust a claim, and they are not probabilities. Anything below 50% is demoted automatically to a gray note. That one threshold keeps the queue honest.
What it caught that the manual review didn't.
- The same invoice booked twice under two different document numbers, one carrying the previous month's number. Both sat in the approval queue as separate, plausible bills.
- An accrual that double-counted a bill already in the ledger, because it credited accounts payable instead of the accrual liability.
- A supplier account that had never been relieved: years of invoices and credit notes, and not one payment ever applied against them in the ERP.
- Bills booked net of VAT against years of the opposite treatment for that vendor, understating the cost every month.
- Documents re-dated so they would pass a validation rule, which moved the cost into the wrong month instead of fixing the underlying problem.
- A two-leg settlement left unbalanced, with a small residue stranded in a transit account and the two legs a day apart.
- Daily bank charges merged into one journal, so the bank balance no longer reconciled on the individual statement dates.
What changed.
A short queue instead of several hundred records
The reviewer sees what to fix first, what is safe to approve, and what to raise with the accountant.
Errors found before the period locks
Problems surface while they can still be corrected, not weeks later at close.
Close readiness, per period
For each fiscal period: items awaiting approval, open findings, and invoices received but not yet booked.
A view of the bookkeeping pace
Volume and value month over month, so a stall or a surge is visible before it becomes a close problem.
What made it work.
Rank the sources of truth, explicitly
A human reviewer who saw the record outranks any pattern the system inferred. Encoding that hierarchy, rather than treating every check as equal, is most of the difference between a useful queue and a noisy one.
Score confidence instead of raising flags
A binary flag forces a choice between missing things and crying wolf. A weight lets a weak signal stay visible without demanding action.
Let the reviewer delete your checks
Two heuristics were retired outright on the reviewer's say-so, including one that sounded reasonable and was simply wrong. Every retirement made the remaining output more trusted.
Where else this applies.
The pattern is general: work that a senior person must check, checks that can be written down, and the evidence to check against already sitting in systems you own. That describes bookkeeping, AP and invoice processing, payroll review, supplier onboarding, contract and order intake, and most shared-service quality assurance.
What it needs is unglamorous: read access to the source system, a written-down version of how you do things, and one reviewer willing to say which findings are worth having. The scoring model does the rest.
More work.
AI automation, finance operations
Autonomous accounts payable
An AI agent that turns a shared folder of invoices and bank statements into a maintained payables ledger, a live dashboard and a payment tracker, with no manual data entry.
Read the case studyData and analytics, operations
Business performance dashboard
Hundreds of thousands of raw transactions turned into a dashboard that shows an owner how the business performs: profit per staffed hour, trends and movers, and the read-out written by an AI agent.
Read the case study
Tell us what you're working on.
A 30-minute conversation about your business, and a clear view of what we would do first. No obligation.