Automatic reconciliation with AI: banks, statements and price lists without crossing spreadsheets by hand

Reconciliation is the most thankless administrative job there is: somebody spends hours crossing two lists of numbers to discover that almost everything matches. The problem is that "almost" is exactly where the money is, and because the process is tedious it gets done late, gets done on a sample, or stops getting done at all.

In this article we look at which pairs of sources are worth reconciling in a distribution business, why manual crossing fails in predictable ways, how automatic matching works — including the difficult one-to-many cases — and what to measure to know whether the process is healthy.

1. What gets reconciled in a distribution business

1.1 Bank statement against the ERP

The classic case. Every bank movement has to find its counterpart in the system: collections credited, payments issued, fees, transfers between accounts. The typical discrepancies are collections credited that were never applied to the invoice, early-payment discounts applied on one side and not the other, and bank charges nobody records.

Without this reconciliation kept current, the accounting balance and the real balance diverge, and every cash flow projection is built on sand.

1.2 Supplier invoice against the purchase order

The check that you are billed for what you ordered at the price you agreed. It is the case we already covered in depth in automatic invoice validation, and probably the one with the highest direct recovery per unit invested.

1.3 Price list against what was actually invoiced

The most forgotten one, and the one that eats the most margin in silence. The list says one price, the system invoices another because of a badly entered discount, an expired term that stayed active, or a promotion that was never switched off. Nobody complains because the customer is happy, and the difference shows up months later in the profitability analysis without being attributable to anything concrete.

1.4 Third-party settlements

Card settlements, transport agency settlements, sales platform settlements or rep commissions. They all share the same structure: a third party sends you an aggregated summary and you hold the detail. Verifying that the aggregate matches the detail is tedious and so it almost never gets done, which is precisely why it is worth doing.

1.5 Physical inventory against the system

The count against theoretical stock, with the difference valued and attributed by warehouse and category.

2. Why manual crossing fails

2.1 Volume makes 100% unviable

With hundreds of movements a month, complete reconciliation does not fit into the time available. It gets done on a sample, or only on the large amounts. And the systematic discrepancies — the small ones that repeat every month — are exactly the ones that survive sampling, and the ones that add up most over a year.

2.2 The identifiers do not match

The statement says "TRANSF 4471 ANONIMO SA", the ERP says "Invoice A-0001-00004471, Anónima S.A.". They are the same movement and no exact-equality cross will ever find them. A human operator recognizes them because they interpret, and that is exactly the work that eats the hours.

2.3 The relationships are not one-to-one

One customer pays three invoices with a single transfer. Another pays one invoice in two parts. A third pays five invoices minus a credit note. Real reconciliation is many against many, and that is the reason the Excel macros that worked the first month break in the second.

2.4 It gets discovered too late

A discrepancy detected at 90 days no longer gets claimed: the supplier disputes it, the customer does not remember it and the carrier changed contacts. The value of reconciliation decays over time far faster than most people assume.

3. How automatic reconciliation works

3.1 Reading any source

The first step is getting both sources into the same normalized format, whatever shape they arrived in: a PDF from the bank, an Excel file from the carrier, a table from the ERP, a photo of a settlement. Document-reading models now solve this without a prior template per supplier, which was the historical limitation of traditional OCR.

3.2 Matching in layers

The cross runs in successive levels, from greatest to least certainty:

1. Exact: same identifier and same amount. Resolves the bulk of the volume with no intervention.
2. Fuzzy: similar identifiers, names with variants, nearby dates and identical amounts. This is where AI contributes what a fixed rule cannot.
3. Grouped: combinations of N records on one side against M on the other that add up to the same figure. This is the level that resolves consolidated and partial payments.
4. With tolerance: differences from rounding, withholdings or fees within a configured threshold, which get reconciled and recorded as an adjustment.

3.3 Well-defined tolerances

Tolerance is the most important business decision in the process. Too narrow and everything falls into exceptions, which means you gained nothing. Too wide and you start automatically reconciling differences you should be claiming.

The reasonable practice is defining it in two dimensions — percentage and absolute amount — and applying whichever is more restrictive. 0.5% on an invoice of a million is not the same as 0.5% on one of a thousand.

3.4 Exceptions with a cause

What does not reconcile is the real output of the process. Every exception should come out classified: missing in one source, amount difference, quantity difference, duplicate, out of period. With the classification in hand, the operator stops investigating from scratch and moves to resolving already-diagnosed cases.

3.5 Traceability

Every reconciliation is recorded with who approved it, which rule matched it and what adjustment was applied. That is not bureaucracy: it is what stops the audit from asking for the same work twice, and what makes it possible to reconstruct any decision six months later.

4. Implementation and metrics

4.1 Where to start

The order that tends to give the best return is: supplier invoices against purchase orders first, because the recovery is immediate and measurable. Then the bank statement, because it unblocks cash visibility. And third, price lists against what was invoiced, which is the one that surprises people most when the first result comes in.

4.2 The process indicators

• Automatic match rate: the percentage reconciled without intervention. A mature process sits between 85% and 95%.
• Cycle time: days from the close of the period to complete reconciliation.
• Amount in exception and how old it is.
• Effective recovery: money claimed and collected thanks to detected differences. It is the number that justifies everything else.
• Recurring differences by counterparty: the indicator that detects systematic problems rather than one-off errors.

4.3 The mistake of only measuring hours saved

The usual argument for automating reconciliation is freeing up administrative hours, and it is valid but minor. The big value lies in two things the manual process does not deliver: reconciling 100% instead of a sample, and detecting things in time to be able to claim them. An automated process that finds a pattern of over-billing by a supplier pays for several years of the tool in a single finding.

Frequently asked questions

What is the difference between reconciliation and invoice validation? Validation compares an invoice against what was expected, before paying it. Reconciliation crosses two sets of records that already happened to find discrepancies. The first prevents, the second detects.

Do I need to change ERP to automate reconciliation? No. The process reads from the ERP and from the external source, crosses them outside and returns the result. It does not require touching the transactional system.

What happens with the differences the system cannot resolve? They go to an exception flow with the cause already classified. There is always a residue that needs human judgement, and that is the part of the work genuinely worth a person's time.

Is it useful if I reconcile few movements? At low volumes the saving in hours is marginal, but early detection of discrepancies still has value. The decision depends on how much what you cannot see today is costing you.

How nBlock automates your reconciliations

nBlock's Reconciliation block crosses any pair of sources over your operation:

• Reads PDF, Excel or ERP data, with no prior template per supplier.
• Crosses records and amounts with exact, fuzzy and grouped matching, on tolerances you define.
• Detects differences and missing items, classified by cause.
• Leaves complete traceability of every reconciliation and every adjustment.

---

Want to know how much you are overpaying without seeing it? Book a demo and we will run it on your own sources.