Reconciliation Automation: Matching at Scale and Handling Breaks

Reconciliation automation compares two records of the same events and lists what does not agree. The automation removes the matching step and nothing else. Investigating a break still costs a person time, and where walking the match list was also how that person built context, the minutes per break rise to swallow the saving. The payback comes from presenting what the automation compared.

There is a sentence that appears in almost every automation proposal, and it is nearly always false: this step takes a person three hours a day, so removing the step returns three hours a day. Automating a step returns the minutes spent on that step and nothing else, and where the step was quietly producing something in addition to its output, the removal transfers cost rather than deleting it. The side effect is usually context: what a person knows about the day before anybody asks them a question about it. A business case that counted only the step is then wrong by exactly the size of the side effect, and it will be wrong in the direction that flatters the proposal. Sumeru Bank Limited, an invented lender, ran exactly that case, and every figure below belongs to its one deployment.

What does a reconciliation actually compare, and in what order?

The domestic version is the familiar one. A handset shows a bank statement and a folded receipt sits in a pocket. A reconciliationComparing two records of the same events and listing the items the two do not agree about. is nothing grander than holding those two side by side and listing the lines that do not agree. Two records, both claiming to describe the same events, produced by different people or different systems, and the question is only ever which items appear in one and not the other, or appear in both with different particulars.

At Sumeru Bank Limited the two records are the disbursal instructions raised by the intake chain and the entries the core banking ledger holds against them. The run happens once every working day and covers 430 items a day, counting lines on the reconciliation list rather than one line for each loan the bank decided. A list length is not a decision count, and reading the 430 as the number of loans decided is the first mistake available.

The order is not decoration, so follow the run left to right. The run is five steps, and only one of them is the step people mean when they say the reconciliation was automated.

How does matching work at scale, and what do the tiers cost?

Nobody writes one rule for 430 items. A bank writes an ordered set of rules, tried in turn, each one looser than the one before it, and every item that pairs at an earlier tier is taken off the table before the next tier runs. Each rule in that set is a matching tierOne rule in an ordered set of matching rules, each tried in turn and each looser than the one before it., and this bank runs three of them.

Tier 1 pairs on reference, amount and date all agreeing exactly. Tier 2 keeps the reference exact but allows the two dates to sit a day or two apart. A payment raised late in the evening on one system lands the following morning on the other. Tier 3 drops the reference to a close resemblance and leans on the amount and the date instead. Somebody typed a reference by hand and dropped a character, and Tier 3 is the rule that still catches it.

The trade is the whole of the design decision. Every tier added takes items off the break list. On a dashboard that looks like pure gain. The price paid for it is the possibility of a false matchTwo items paired by a loose rule that are not in fact the same event, so both disappear from the break list without anybody checking., meaning two items that a loose rule paired and that are not the same event. A break is visible and a false match is not, so loosening a tier trades a problem somebody will look at for a problem nobody will. The third tier is therefore where the argument always happens, and a bank that keeps widening the tolerance to shrink its break count is buying a smaller list rather than a cleaner one.

What is a break, and which kinds does this bank see?

A breakAn item the two records disagree about, which no matching rule could pair and which a person has to investigate and clear. is simply an item no tier could pair. On this run, of 430 items a day, 421 pair automatically, being 97.9 per cent, and 9 do not, being 2.1 per cent. Nine sounds like nothing until the cost of one is known.

The nine are not nine of the same thing, and this is where most reporting goes wrong. The nine split by kind: 5 are a timing breakA genuine item that both records hold, recorded on different days, so it will clear itself once the later day arrives., 3 are an amount break, and 1 is a duplicate referenceThe same reference appearing twice in one record, so a matching rule cannot tell which of the two the other record means.. Five plus three plus one is nine. Only the 3 amount breaks are an error in the sense a controller means, so a report that calls all 9 exceptions describes the majority of its own list wrongly.

The three answers are not alike, so look at what each kind actually asks a person to do. A timing break wants confirming and dating, and it will clear itself when the later day arrives. An amount break wants tracing back to which system holds the right figure and then correcting one of them. A duplicate reference wants disambiguating. The machine could not tell which of two identical references the other record meant, and there is nothing wrong with either record.

What did a day of this cost before the automation?

One person, one working day of 7 hours, being 420 minutes. The work came in two parts and the bank measured both.

The first part was walking the list. Both records printed out or on screen, item by item, ticking off what agreed. The list walk took 180 minutes. The second part was the nine breaks, and they took 45 minutes between them, at 5 minutes each. Add them and the day cost 225 minutes, a little over half of a 420 minute working day.

The 5 minutes a break is the figure everything after it turns on, and nobody wrote down why it was 5. Five minutes is fast for investigating a disagreement between two systems. Five minutes was fast because of where in the day it happened. The person reached the break having just walked all 430 items, so they already knew what sat either side of it, what the day looked like, how noisy that channel had been that morning, and whether three other items carried the same reference stem. None of that was in the break. All of it was in their head, put there by the 180 minutes.

What does a day of this cost after the automation?

The matching now runs in seconds. The 180 minutes of list work is gone completely, and that is the one part of the before picture nobody disputes: there is no list to walk, so nobody walks it. Nine rows arrive on screen saying these nine did not pair.

Then the second measurement came back, and it was not what the proposal expected. A break now takes 25 minutes rather than 5. Nothing about the breaks changed. The same 430 items, the same 9 breaks, the same 5 timing, 3 amount and 1 duplicate reference. The change is that the person now arrives at each break cold, and the 20 extra minutes is the cost of building from scratch the contextWhat a person knows about the day's items before being asked a question about any one of them, which the old list walk supplied for free. that the list walk used to supply as a side effect.

Nine breaks at 25 minutes is 225 minutes.

Why do the two totals land on exactly the same number?

Put the two days beside each other and read the ends. Before: 180 plus 45 is 225. After: 9 times 25 is 225. The automation at this bank saved no minutes at all, and that sentence is a measurement rather than an opinion about automation.

Why does more volume not rescue the saving?

The first thing anybody says when they see 225 against 225 is that the volume is too small. Automation is meant to pay at scale, so run ten times the items and the fixed machine cost spreads while the people do not multiply. The instinct is sound on most automations and wrong on this one, and the reason is worth sitting with.

Take n items a working day at this bank's own break rate of 9 in 430. The list walk is proportional to the items, so it costs 180n over 430. The breaks are proportional too, at 9n over 430 of them, and each one cost 5 minutes. So the before side comes to 180n over 430 plus 5 times 9n over 430, or 225n over 430. Now the after side: the same 9n over 430 breaks, each at 25 minutes, again 225n over 430. The two sides are not merely close at this volume, they are the same expression of n, so the difference is zero at every volume and no amount of scale will open a gap.

Per item it is starker still. Each item on the list costs 225 over 430 of a minute, being about 0.52 minutes, and it costs exactly that before the automation and exactly that after it. Ten times the items gives 4,300 a day, 90 breaks, and 2,250 minutes on both sides. A hundred times gives the same answer a hundred times over. Volume is the first control anybody reaches for on an automation, and at this bank it is precisely the control that changes nothing.

Try the volume buttons and watch what refuses to happen. The numbers under the bars grow by five times and ten times, the axis relabels itself, and the two bars do not move a pixel relative to each other. Because 225 divided by 9 is 25 whatever the multiple is, the crossing sits at 25 minutes a break at 430 items, at 2,150 and at 4,300. The invariance is the point: the only control that changes the answer is the slider on the top, and the slider is not about the matching at all.

What change does produce a saving, and how big is it?

If the per break minutes are the only lever, then the design question is not how to match better. The design question is how to hand a person a break they can start work on. The bank's own answer, and it is not an expensive one, was to change what the break screen shows.

The break screen as first built showed a verdict: this item did not match, here is its reference and amount, please investigate. Everything the automation had done to reach that verdict was thrown away on the way to the screen, so the person had to reconstruct it. The rebuilt screen shows the comparison instead: both records side by side, the tiers that were tried, why each one failed, and the items sitting either side of it in each record. On this bank's measurement that takes a break from 25 minutes to 12.

Nine breaks at 12 minutes is 108 minutes, against a before day of 225. The saving is 117 minutes a day, and every one of those minutes comes from the presentation rather than from the matching. Over 20 working days it is 2,340 minutes a month. Against the assumed working month of 8,400 minutes that is 0.28 of a post, and at the fully loaded Rs 9,00,000 a year this bank assumes for a post it is worth about Rs 2,50,000 a year. Rs 2,50,000 a year is not a large number, but it is a real one, and it is the first real one this project produced.

Notice what the rebuilt screen does not do. The screen does not decide anything, it does not rank the break, and it makes no claim about which record is right. All it does is stop discarding the working the automation had already done. The matching was never the expensive part; discarding the matching's working on the way to the screen was. A duplicate reference is the clearest case of it: the machine knew there were two candidates and could not choose, and the first screen reported only that it had failed, so the person spent twenty minutes discovering the one fact the machine already held.

What did the automation actually buy, if not minutes?

Saving nothing is not the same as being a waste, and the review that said so was clear about it. Four things changed and none of them is a headcount saving.

The first is elapsed timeHow long a run takes from start to finish on the clock, as distinct from the minutes a person spends inside it.. A complete break list used to exist only after somebody had spent 180 minutes producing it, so on any given morning it did not exist yet. Now it exists in seconds. The second is the ability to run more often: a comparison that takes seconds can run twice a day, or hourly, where one that takes three hours of a person's day cannot. The third is that the reconciliation no longer stops when one person is on leave. Nothing about it now depends on that person's stamina. The fourth is the least glamorous and the most useful: what the matching did is now written down as it happens.

What does an automated reconciliation owe an audit trail?

A person walking a list leaves almost nothing behind: a tick, a signature at the bottom, and a memory that fades by the afternoon. A tick and a signature were never good enough, and the only reason they survived is that nobody could do better without doubling the work. A machine has no such excuse. The record is a by-product of the run rather than an extra task laid on top of it.

Four things are worth writing for every item, and they are cheap only if they are written as the run happens. Which two records were compared and for which window. Which tier paired the item, or which tiers were tried and how each one failed. The item as it stood in both records at the moment of comparison, before either record had a chance to change. And who cleared the break, when, and on what basis. The failed tiers are the most valuable of the four. A break with no record of what was tried forces the next person to try every tier again, and trying every tier again is the twenty minutes.

There is a second reason, and it is the one a reviewer cares about. A matching rule is a rule somebody wrote, and rules get widened quietly. If the record shows which tier paired each item, then a tier that starts pairing far more than it used to is visible in a week rather than in an annual review. Without that record, the only evidence that a tolerance moved is a falling break count, and everybody reads a falling break count as good news.

How does somebody outside the project actually use this?

Three people meet this guide in three different rooms, and each of them has a single question worth asking.

An operations head handed an automation proposal asks what else the step being removed is producing. The minute count is usually right. The real question is whether the step is the only place somebody currently learns something. The test is simple and it takes one conversation: ask the person doing the work what they would not know if the step vanished tomorrow. If the answer is nothing, the saving is real. If the answer takes them a while and involves the word usually, the saving has a hole in it the size of the answer.

A controller reading a reconciliation report asks what the break count is made of before asking whether it is high. Nine breaks with five of them timing is a completely different report from nine breaks with five of them amount differences, and a single count hides which one is in hand. The second question is which tier paired the bulk of the items. A rising share pairing at the loosest tier is the one movement that looks like an improvement and is not.

An analyst reading an institution's automation claims asks which of the four things on the honest list is being described. Elapsed time, frequency, resilience and a record are all worth having and none of them shows up as a smaller cost base. When an automation is described in minutes and its benefit is really one of those four, the number in the pack is not wrong so much as measuring the wrong thing, and the way to find out is to ask what happened to the per unit handling time after go-live. If nobody measured it, nobody knows what the project bought.

Sumeru Bank Limited is invented.
Educational material. Not advice on any investment, tax, budget or market position.