FoundMargin US based company
Margin Notes Request a fit call
· N Khanna

Reading 4,000 line items is the easy half

A machine can sort a year of spend by function in under a minute. It never gets bored on row 3,000, and it hands back a confident number that is wrong in 3 predictable ways. Here is how the work actually splits, and which half you can run yourself this week.

An illustration of a person at a long desk sorting a wide fan of paper slips into 2 piles, a much larger discarded pile and a small kept pile, with an open green cash box beside the small pile and coins going into it.
The whole thing up frontA year of payables and card charges at an illustrative 60 person distributor is about 4,000 lines. A machine groups them by function, decodes the unreadable descriptors and flags 180 candidates in under a minute. Then a person throws out 140 of them. It gets the same 3 things wrong every time: it turns a single payment into an annual cost, it reads an unused seat as a wasted one, and it merges 2 charges because they share a supplier name. The 40 that survived came to about $31,000 a year. The first half is nearly free now. The second half is the whole job.

The first thing I ask for is 12 months of accounts payable and 12 months of card charges, exported to a spreadsheet, no formatting, no summary tab.

It usually arrives as about 4,000 rows.

Nobody in the building has read all of them. I have not read all of them either, and neither has anyone who tells you they have.

That used to be the hard part of this work. It is not any more, and pretending otherwise would be selling you a morning of typing.

What a machine is genuinely good at here

Row 3,000 gets the same attention as row 3. That is the whole trick, and it is not a small one. A person reading a long export gets careless somewhere around the second screen and never notices the moment it happened.

3 jobs it does well, all of them dull:

  • Decoding the descriptors. A card statement gives you PMT* HLXAPP and SQ *NORTHGATE 04. Matching those against a vendor list is mechanical, and it is the step that makes everything after it possible.
  • Sorting by function instead of supplier. Every export arrives sorted by who invoiced you. Sorting by what each thing does is what makes 3 tools doing 1 job visible, and it is a minute of work rather than a morning.
  • Cross-referencing 2 files. An invoice line against a seat count from the admin console, 400 times, without transposing a digit.

None of that is clever. All of it is tedious, and tedium is where a careful person stops being careful without noticing.

The 3 things it gets confidently wrong

The output does not look uncertain. It comes back as a tidy number with a total at the bottom.

1. It turns a single payment into an annual cost. A $3,400 data migration paid once in March looks exactly like a monthly charge that started in March. Annualized, that is $40,800 of savings that never existed.

A single payment read as a recurring one Two bars comparing the same charge. A $3,400 data migration fee paid once in March is reported as $40,800 a year when it is treated as recurring, against the $3,400 it actually was. The gap of $37,400 is savings that were never available. 1 migration fee, paid once in March Same charge, 2 readings Reported Actual $40,800 a year $3,400, once $37,400 of it was never there

2. It reads an unused seat as a wasted seat. 22 accounts not opened in 90 days is a real signal and a bad conclusion. Some belong to a crew that starts in January. 1 is the read-only login your accountant uses at year end. Cancelling those saves $0 and costs you a week in February.

3. It merges 2 charges because they share a supplier name. A supplier sells 6 products. You own 2 of them, and they are not duplicates of each other, they are storage and e-signing. Collapsed into 1 line, that reads as an easy cancellation and it is not.

Every one of those errors runs in the same direction. They all make the number bigger.

How the work actually splits

Two passes over a year of spend, and what each one removes A left to right flow in 3 stages. About 4,000 payable and card lines go through a first pass that decodes descriptors and groups them by function, which flags 180 candidates. A second pass by a person removes 140 of those as one-off payments, seats that are needed, or charges that only look duplicated, and keeps 40 real findings worth about $31,000 a year. 4,000 lines 180 candidates 40 worth acting on 1 minute most of a week about $31,000 a year

The middle box is the cheap one. The arrow after it is the expensive one, and it is where the answer actually gets made.

What no tool is going to do for you

Decide what matters. The $180 a month item on your list is the scheduling tool the warehouse runs on. Nothing in the data says that. The only way to find out is to ask.

Ask the person who runs it. A finding about somebody’s process has to come from them or it does not survive contact with reality, and they will usually tell you a better version of it than you had.

Make the call. Reading that a freight bill differs from the quoted rate takes a second. Getting the money back means a person, a phone, and a carrier who would rather not.

Carry the recommendation. Cancel the wrong thing and somebody’s Tuesday stops working. A confident total does not sign its name to that.

Run the cheap half yourself this week

Most of this is an afternoon.

  • Export 12 months of card charges and 12 months of payables to 1 spreadsheet. No formatting, no pivot tables. The mess is the point.
  • Add 2 columns nothing will give you: what each thing does, in plain words, and whether the charge repeats.
  • Sort by the first column. Duplicates announce themselves.
  • Then go down the list and mark the ones you already know are load bearing. That list is shorter than you think and it is the most valuable thing on the page.

The tooling for the first 3 steps is ordinary now, and I would rather you used it than paid anybody for a morning of sorting.

What this changed about the work

The part of this job that used to take a week takes an afternoon. Knowing which 40 of the 180 are real did not get cheaper at all.

The tooling made the boring half nearly free, which means the only thing left to sell is judgment, and judgment is the thing you can check. Ask whoever hands you a savings number which of the 180 they threw out and why. If they cannot answer that, they did not do the second pass.

Go do the sort. If the number that comes back looks too good, it probably is, and the 3 errors above are where I would look first.

And if you would rather we went through it with you, the audit is 3 weeks and $2,500, and if we cannot identify $7,500 a year you do not pay. My co-founder and I take 3 a month, which is a real limit and also the reason I would rather you tried the free version first.

We run this as a fixed-fee audit. $2,500, and if we find less than $7,500 a year we refund it.

See whether it fits your company