The infrastructure behind ATMs

Patrick McKenzie (patio11)

Programming note: Happy New Year! I’ve left my full-time employment and, while I am not entirely certain what I’ll do for my next adventure, I intend to continue writing Bits about Money and potentially increase the number of cycles I spend on it. I’m also moving it to a reader-supported model. More details on Friday but if you already know that is something you want to support, then here you go. It is also now on its own domain (

ATMs are a fascinating example of a pattern we see a lot in finance: an internal operations improvement which was built into a business which eventually begat an infrastructure layer that may be a much bigger business. And for all their ubiquity, almost no one, even people professionally involved in finance, understand how they work.

The first automated teller machines, which debuted in the late 1960s, were, as the name suggests, strictly cost-saving devices for bank branches. Branches exist as sales offices but have incidental cash-management functions. The denser depositors are around a branch, the more transactions happen during peak windows like e.g. the morning commute and lunchtime. The more transactions you need to support in a window, the more tellers you need to employ. Tellers are both surprisingly inexpensive relative to the degree of trust placed in them but surprisingly costly relative to occupations like e.g. cashiers which look outwardly similar. Banks have long wanted to control the costs of the teller base.

The original thesis behind the ATM was that you could move the most routine teller transactions, like cash withdrawals and balance inquiries, to a machine, and then reserve the teller for higher-complexity routine transactions like cashing checks. The machines gradually gained more features as they achieved ubiquity.

Interestingly, teller employment is actually up substantially since the introduction of ATMs. Secular demand for retail banking grew with the economy and the larger number of branches has compensated for reduced numbers of tellers per branch. See Bessen 2016.

The basic ATM transaction

What happens when you withdraw from an ATM? To explain it, first we have to dispose with a very common misconception about money.

Many people believe that cash, the paper printed by the government, is coextensive with money, and that bank deposits kept at private sector banks represent a claim on money. Neither of these is true. Instead, bank deposits are themselves money, created by a public/private partnership. They are the dominant form of money. An economist would tell you about the M1 versus M2 money supply here. As a sometimes product person in financial services, I’d just handwave and say “About 90% of what the entire world thinks of money is useful money and 10% is legacy paper certificates that banks will happily exchange for useful money.”

And so the most common transaction at an ATM is not a withdrawal so much as it is a sale. You are buying some paper with a mystical property associated with it, in return for money, and often paying a convenience fee.

This sale is far, far more vulnerable to fraud than most sales in the economy, due to the mystical properties of the paper you are buying. In particular, most transactions in the economy do not rely on real-time confirmation of availability of funds. This transaction generally does, and so it can’t ride the rails used elsewhere in the economy. ("Rails" are the industry lingo for the combination of technical infrastructure, relationships between institutions, and contracts which allow money to move in a fast, predictable fashion.)

Using an ATM owned by your bank is fairly straightforward: the bank can trust its own databases and make the series of offsetting transactions to effect your paper purchasing internally.

Your deposit is a liability of the bank. Your purchase of paper swaps an asset of the bank, the paper with mystical properties, for that liability, partially extinguishing it. If you are charged a fee, the bank partially extinguishes its liability to you and books revenue.

Getting cash from an ATM not owned by your bank, however, required some infrastructure magic.

“Off-premise” ATMs and ATM networks

While originally being concentrated in bank branches and operated just for the owning bank, ATMs quickly became a business in their own right. So-called “off-premise” ATMs were owned and operated by non-bank entities and placed in bars, restaurants, malls, travel corridors, and similar.

Although many readers may not remember this, bank ATMs used to be free as a matter of policy, through approximately the 1990s. Off-premise ATMs were not; they charged a (generally fixed) convenience fee for cash withdrawals, which were overwhelmingly the most common transaction they were used for.

ATMs became a very, very large and lucrative business. The typical transaction rhymes with "I'll sell you $40 worth of paper for $43, with a few minutes of operator attention being required every several hundred transactions." The attractiveness of this math for operators caused millions of ATMs to be deployed worldwide.

In the earliest days, ATM operators would negotiate serially with banks to support them, and so you’d have an ATM which could perhaps work with a regional chain and a large national bank, but your choice of which banks exactly would be different on individual ATMs. This was wildly inefficient for all parties. Banks don’t want to be in the business of negotiating with operators who have only ten ATMs. Operators would prefer to have the addressable market of “everyone” rather than only a subset of the foot traffic in front of their ATM. Customers don’t want to have to search over town for a compatible ATM, particularly in the days before smartphones.

And so so-called interbank networks began forming. One early network was Cirrus, which began as an association of six banks and was eventually bought by MasterCard for the princely sum of $40 million. ($40 million went a lot farther in fintech in the 1980s; these days it would be an oversubscribed Series A.) Cirrus’ pitch to operators was simple: for the price of one negotiation/integration, they could light up customers across multiple banks. Cirrus would continue working to expand the number of banks who worked with them, which would passively increase the reach of any ATM bearing the Cirrus logo. In return for this, operators would pay Cirrus a fee for transactions facilitated.

And what did those transactions look like? They mirror the swift series of offsetting transactions used by other payment rails like bank transfers and credit cards.

As you step up to the ATM, two interesting facts about the world are that your bank owes you some money and that the ATM operator owns some magic paper which is currently in a plastic enclosure very close to you.

As you transact with the ATM, the network it connects to quickly makes a deal with your bank, causing your bank to owe less money to you and more money to the ATM network. It simultaneously makes an offsetting deal with the ATM operator, causing it to owe money to that ATM operator. The ATM operator then causes their machine to release the magical paper.

Note that the settlement of these deals is necessarily much slower than the transaction itself is. The ATM network will often be paid the following business day by your bank, but the ATM operator will probably be paid several days later by the ATM network, using an ACH transfer (in the US) or similar frequently slow rails.

From this fact arises two interesting implications.

The first is that operating ATMs is a very capital intensive business to be in, not because of the cost of the machine (only a few thousand dollars each) but because of the cost of tying up magical paper as inventory and floating actual money frequently to cover customers’ purchases of magical paper which have not settled yet.

The second is that, holy cow: doesn’t this exactly resemble a payment rail in all particulars. Operating payment rails is a much, much better business than selling paper, even at eyepopping markups.

Guess what?

ATM networks factually are payment rails

If you’ve ever used a debit card to purchase something in the U.S. you may have been asked “Debit or credit?” and wondered “Hmm, can’t they tell from looking at it that it is a debit card? Or isn’t it, I don’t know, in the card number or the magnetic stripe or somewhere?”

That was not the question you were being asked.

The question you were being asked, and if a human was asking it they probably didn't understand this either, was “Do you want to run this transaction over credit card rails or debit card rails?”

That's right: the ATM networks long since realized that, since they necessarily have all the features of a payment rail, they should get into the payments business even if it was disconnected from the machines they are most associated with and even if no paper was involved. They just needed to convince businesses to use them.

That choice between which network to use, which you didn't even know you were making, is yours by right. This was established by law and regulation. The law is the Electronic Funds Transfer Act (EFTA), better known in the industry as Regulation E. Reg E governs a titanic portion of bank operations and is extremely customer-friendly. (I, ahem, may have once ghostwritten a few hundred letters helping less savvy consumers convince banks to comply with their responsibilities under Reg E and similar. I have weird hobbies.)

Section 920 of Reg E came later. In 2011, implementing regulations (PDF; see page 2) added a little-known requirement that requires that debit card issuers support at least 2+ card networks with different affiliations. The credit card network (Visa/Mastercard/etc) counts for one of their requirements. The credit card networks have affiliated debit card networks but, due to the diversity-of-affiliations requirement, those can't count for the second network. Instead, that gets solved by unaffiliated (to Visa/Mastercard/etc) debit card networks, which mostly grew out of ATM networks.

These companies, with names like STAR and Accel, are far, far less known among consumers than the global payment brands, but they run extremely similar rails. Depending on your specific issuer, you may find their logos on your physical card, often on the back. This is more common on cards issued by community banks, which view these networks as alliances to make their cards convenient even with a very constrained branching network. Large regional and national banks mostly consider these networks a tick-box compliance requirement and may not even affirmatively make customers aware of their existence.

The experience of debit card networks is a partial answer to the question of “Can one compete against the major credit card brands among merchants simply by slashing the price of interchange down to nearly nothing?”, in that a) that experiment was run, b) that experiment actually runs today, and c) 99.98% of people who read a newsletter specifically about financial infrastructure probably had no idea this option existed. (From this derive interesting implications in both the durability of network effects and the importance of the global marketing campaigns by credit card brands, underwritten by the materially higher interchange that businesses accepting them pay.)  

Why don’t consumers just use these networks preferentially? Prior to asking that question about society, ask it about yourself: how often are you going to use these networks going forward given a relatively obscure incantation you need to do to access them and basically no direct benefit to you? There, that’s a good first cut at the answer.

Why don’t businesses push people in the direction of these networks to minimize their payments costs? There are some fascinating stories here, which partially rhyme with “There are subtle tradeoffs in e.g. authorization rates for transactions pushed over credit and debit networks to particular issuing banks, and very few businesses have sufficient data and sophistication to route transactions in a way which is optimal at balancing cost and authorization rate.”

The nature of career management at large businesses for subject matter experts tends to create a missing middle of optimizations. There are simple things that a Director of Payments will do as a matter of course, like negotiate with their payment providers. Then there are complex, career making projects that they might attempt to accomplish over years, like turning payments into a profit center via e.g. creating some sort of points system to rival that of an airline or Starbucks. Optimizing network routing ends up not happening because it is very hard, requires substantial sophistication, is available for only a fraction of payments (those on debit cards), and is not nearly as impressive to senior executives as the brand new points system that will be all over the website and annual report.

A quirk of retailer cash management

If we can return to "Debit or credit?", U.S. readers might recall that at grocery stores, Target, and the like, if you choose "debit" you are often prompted for a PIN (which will, somewhat surprisingly, frequently be ignored by your financial institution even if you get it wrong) and offered the option of cash back.

You might be curious why that option exists, given it might cost the store a tiny amount at the margin to provide it to you. And the answer is that it solves a cash management need for the store.

People outside the payments industry often compare the costs of it to the perceived free handling of cash. Handling cash is not actually free; it is subsidized for retail clients and very small businesses by the bank. Larger businesses pay market prices, like they do for other banking services. If you deal with commercially relevant quantities of cash, in addition to your own labor costs associated with it, your bank will assess a fee for depositing it. Larger retailers will negotiate their own bespoke rates, but the right ballpark for small businesses above a nominal amount is 3 bps, or $3 per $10,000 transacted.

And so, getting magical paper out of the till and into your wallet without it first visiting the bank saves the retailer money. It can also, potentially, earn the retailer a small amount of float. Cash in its tills is dead money and may not be deposited until e.g. the end of the week or later, but selling that paper to a customer for real money results in it arriving in their bank account faster than physically walking it to the bank. It can then be deployed to more productive purposes than gathering dust. The least interesting productive use, which is still productive, is earning interest at the bank.

More fun in the ATM business

One of the first uses of Big Data (TM) in the banking business was rationalizing forecasts of ATM demand for cash, to allow operators to minimize their needs for cash inventory and minimize the amount of servicing trips they'd need, particularly for out-of-branch locations.

This is a surprisingly large problem in e.g. Japan, which despite a combined push from the government and financial institutions to encourage cashless transactions has a single Schelling Point monthly payday (the 25th every month) and an extremely common user behavior of withdrawing much of a family's wage base on payday. As a result, most ATMs in Japan have the equivalent of hundreds of thousands of dollars prestaged in them for the 25th every month, which is 50-100X typical inventory levels for e.g. U.S. ATMs.

You might sensibly wonder whether cash is on its way out, whether that is a good thing, etc. That probably deserves its own essay, some other day. (The short version: cash is one of many competing payment rails. Different equilibria will happen in different places and for different use cases, as we see with other payment rails.)

As always, if you're interested in more on this topic (or have other feedback on what you'd like to read about in BAM), drop me an email.

← BAM is now reader-supported
Credit cards as a legacy system →

Want more essays in your inbox?

I write about the intersection of tech and finance, approximately biweekly. It's free.