Accounting

Nothing failed. Everything was wrong

A green tick means the connection is alive. It says nothing about what travelled down it. One company's missing deposits, quietly rewritten rules and lost reconciliations show how to check.

LPLP
Accounting
In this article

The short version

  • A green tick or a “sync completed” line says the connection is alive. It says nothing about whether the data that travelled was complete or right.
  • Silent faults come in five families: records that never arrive, arrive twice, arrive altered, land in the wrong account, or lose their state, such as a reconciled flag.
  • Bank rules fail quietly too: a rule that was edited, or wrong from the start, can misfile the same payee for months.
  • A reconciliation shows that the book agrees with the bank at one moment. It does not show that the book agrees with the source system.
  • Count and total on both sides of every integration, by day and by account, and treat edits by “System” as a finding.

Every finance team has a dashboard light that reassures it. In the accounting software it is usually a small green tick beside the bank feed, or a line in the integration log that says “sync completed”. The tick means the connection is alive. It says nothing about whether what travelled down the connection was complete, correct or even the right thing.

That distinction sounds pedantic until it costs money. The cases in this piece come from my own work and from other controllers I have worked with. Take one from earlier this year: a mid-sized American services company found that its bank feed had stopped delivering one category of deposit, the daily settlements from its card processor, from one Tuesday in May. The feed did not disconnect. It went on delivering payroll debits, card payments and transfers on schedule, and the green tick stayed green. Only the settlements vanished. By the time anyone looked, a dozen deposits worth well over $100,000 were missing from the books, and a second settlement channel turned out to have failed in the same way. When the provider later repaired a genuine disconnection, the error code it fixed had nothing to do with the gap.

Nobody had done anything wrong. That is the problem.

A taxonomy of quiet

Engineers have a name for this class of fault: silent data corruption. Data is altered, dropped or duplicated somewhere between source and destination without any component raising an alarm. In storage systems it is treated as a serious hazard and hunted with checksums. In accounting, where the stakes are arguably higher, it is mostly hunted by luck and by the month-end close.

The faults fall into a handful of recognizable families:

  • A record exists in the source system and never arrives. The missing settlements above are the textbook case.
  • A record arrives twice, often because a retry succeeded after a timeout that was logged as a failure.
  • A record arrives with a field changed: a date shifted to the posting day rather than the service day, a customer reassigned to a default, an amount rounded.
  • The record is right but lands in the wrong account, usually because an automated rule has been edited, or was wrong from the start.
  • State loss. Something that was true about a record stops being true. A reconciled transaction quietly loses its reconciled status.

The last family is the strangest, and the least discussed. In the same company, the audit trail on one ordinary deposit showed thirteen edits over two days, each attributed not to a person but to “System”, each toggling the transaction between cleared and not cleared. Every bank transaction after a certain date had lost its reconciliation flag. The software’s own reports described the affected entries as “deleted” with a current amount of zero, although they were plainly intact. Anyone reading that report cold would conclude that money had been removed from the books.

The rule that changed

Misclassification deserves particular attention, because modern bookkeeping runs on rules. A mid-market company may have several hundred bank rules, each telling the software that a payee or a description belongs to a particular account. Rules are efficient and invisible, which is a dangerous combination.

In the case above, one rule had posted a payroll provider’s monthly fee correctly to payroll expense for four months, then began posting it somewhere else. Nobody remembered changing it. A second rule pointed at the same wrong account. A third had been sending the employer’s pension match into a liability account, where it accumulated for three months to just over $12,000 before anyone noticed that a balance which should be zero was not. Individually each amount was small. Together they distorted payroll cost, inflated a liability and nudged three months of management accounts in the wrong direction.

I have seen the same pattern wherever an operational platform feeds the ledger. A professional-services automation (PSA) tool such as ConnectWise has synced invoices into QuickBooks perfectly well while carrying the invoice date rather than the service period, so that a month’s recurring revenue straddled two periods in the books. A dealership management system has pushed a day’s sales as a summary, and when one day failed to post, the summary for the next day looked entirely normal. The integration is not broken. It is simply doing something slightly different from what the accountant assumes.

Why the close does not catch it

The conventional defense is the bank reconciliation, and it is a good one. But it has blind spots. A reconciliation proves that the book agrees with the bank at a moment in time. It does not prove that the book agrees with the source system that generated the revenue, nor that the transactions are in the right accounts, nor that last month’s reconciliation is still intact this month. A dropped deposit shows up as a reconciliation difference only if someone reconciles; on accounts nobody has reconciled for a year, the error simply compounds.

There is also a human pattern. When a reconciliation is out by a few hundred dollars, the path of least resistance is an adjusting entry. Software makes this easy: several of the products I have worked with will offer to post the difference automatically. Each such entry converts a symptom of corruption into a permanent, unexplained line in the ledger. Do that for a year and the books balance perfectly, and explain nothing.

Checksums for the ledger

The remedy borrows from the engineers. Rather than trusting that a sync succeeded, one counts and totals independently on both sides of the integration and compares the results, by day, by account and by customer. If the processor says 474 payments settled on a given day and the ledger holds 474 lines summing to the same figure, the batch is probably intact. If it holds 473, or the right count with the wrong total, something has drifted, and the comparison says where to look.

This is the idea behind LedgerDrift, an open-source toolkit I have been developing for some two years and am now piloting with several organizations, from a vehicle dealership to a legal services firm. It does not replace the close. It runs alongside the integration, fingerprints what leaves the source system and what lands in the ledger, and reports the differences in terms an accountant can act on: missing, duplicated, altered, misfiled, or no longer reconciled.

The principle matters more than the tool. Any finance team can apply it with a spreadsheet and some discipline: never accept a sync log as evidence; reconcile counts as well as amounts; check that the rules engine still says what it said last quarter; and treat an audit trail full of edits by “System” as a finding rather than a curiosity.

The green tick will go on glowing. It was only ever telling you that the pipe was connected, never what came out of it.