Reconciliation is one of those processes where the failure mode is invisible until it's painful. A business does reconciliations at $5M in revenue using a spreadsheet, a bank feed, and someone careful in finance. It works. By $25M, the same process is consuming a person's full week at month-end. By $75M, the close is slipping. By $200M, the auditors have questions.
The interesting thing is that the process itself didn't get worse. Volume changed — and a process built for one volume doesn't scale to a different one. Here's the failure pattern we see repeatedly, and what gets built when it's done right.
The volume curve nobody tells you about
Reconciliation effort scales superlinearly with transaction volume. 10× transactions is not 10× effort — it's closer to 15-20× once you factor in exception handling, follow-ups, and the second pass to fix what the first pass missed.
Why: the easy matches (90%+) clear quickly. The remaining 5-10% are exceptions — and exceptions take 30-100× longer to resolve than a clean match. As volume grows, the absolute number of exceptions grows linearly, but you also lose the ability to remember the context of each one. The break point isn't a single number. It's the moment exception handling consumes more time than the matching itself.
The four stages of breakdown
- Stage 1 — The hero. One careful person does reconciliations manually. They're slow but accurate. Knowledge lives in their head.
- Stage 2 — The spreadsheet. The hero builds a spreadsheet to speed things up — VLOOKUPs, conditional formatting, color codes. Now the spreadsheet is the system. The hero is also the spreadsheet maintainer.
- Stage 3 — The team. More people join finance. Each one has slight variations on the spreadsheet. Reconciliation becomes a series of handoffs, each with a slightly different interpretation. Quality drops; everyone blames the data.
- Stage 4 — The audit. Auditors ask how a particular transaction got categorized. No one can answer with confidence. Now you have a compliance problem on top of an efficiency problem.
Why the obvious fixes don't work
"Just use a better spreadsheet." The spreadsheet was never the problem. The problem is that there's no system of record for the reasoning behind a match — only the result.
"Hire more people." Linear staffing against superlinear effort growth means costs explode while output barely keeps up. By the time you've hired the third reconciliation analyst, the close is taking longer than it used to.
"Use the bank's reconciliation tool." Bank tools handle bank-side matching. They don't handle the book-side or the cross-system exceptions that account for most of the manual work.
What modern reconciliation infrastructure does differently
- Treats matching as the easy part. Most modern tools auto-match 90%+ of transactions in seconds. The product differentiator is what happens to the remaining 10%.
- Captures reasoning. Every match — auto or manual — carries the rule that produced it. Re-runs are deterministic. Audits don't ask "why" because the why is recorded.
- Surfaces exceptions early. Instead of finding mismatches at month-end, modern systems flag them in near-real-time so they can be resolved as a daily habit, not a monthly fire drill.
- Reversible by default. Categorization went wrong? Undo. The cost of a mistake collapses from "untangle for a week" to "click undo."
How to know you've outgrown your current process
- Month-end close is the most stressful week of every month.
- Two people doing the same reconciliation get different answers.
- The "true" version of a reconciliation lives on someone's laptop.
- You can't answer "why did this transaction get categorized this way?" without a 20-minute investigation.
- Auditors keep asking the same questions because the answers aren't traceable.
Reconciliation isn't broken because finance teams aren't trying hard enough. It's broken because the underlying process was designed for a smaller business, and process design doesn't migrate automatically. The work of operational infrastructure is to make the same process repeatable at 10× the volume without scaling people 10×. Anything less is just postponing the question.