Digital Public Infrastructure: The Layers Finance Uses

Digital public infrastructure is a set of shared layers many parties reach and no single one holds privately: a layer for identity, one for payments, one through which a person directs that data held about them be shared, and one for issued documents. Each does once, for everybody, work every lender used to repeat for every person. Shared layers change what a lender can see and what seeing costs, and decide nothing.

Almost every explanation of this subject opens with a list of names. One system, one scheme, one application, and then an argument about which of them counts. The name moves, the operator behind it moves, and the arrangement behind that moves again. So an account built on names goes wrong within a couple of years, and it looks authoritative the whole time it is wrong. Names are the worst available starting point. A better starting point holds still: work somebody has to finish before a loan can happen at all, and the question of who does that work and how many separate times. The same work done once for everybody, and the same work done again by every lender for every person, are not the same work at all once the cost of each and what each one leaves behind are counted. Everything below follows from that one observation.

What makes a layer shared rather than somebody's product?

Start outside finance, with the road in front of a row of shops. Each shopkeeper could lay a strip of surfacing from the main road to his own door, pay for it himself, and maintain it himself, and in plenty of places that is exactly what happens. Or there is one road. All the shops reach it, not one of them holds it as private property, and every one of them runs a completely different business off the same stretch of tarmac. The road sells nothing. The road has no view about what any shop should stock or whom it should serve. The road makes the stocking and the serving possible at a cost no single shop could have carried on its own.

Now take the finance version of the same picture, and take the smallest unit of work there is: establishing that the person in front of a lender is the person they say they are. In the unshared arrangement each lender designs its own way of doing that, pays for the designing, pays again every time it runs, and when it has finished, the next lender is exactly where it started. The work was done. None of it survives outside the lender that did it.

Three properties turn something into infrastructure rather than a product, and all three have to hold at once: many parties reach it, no single user holds it as private property, and each of those users builds a different thing of their own on top of it. Drop the first and it is a private arrangement between two parties. Drop the second and it is a supplier with customers. Drop the third and nothing distinctive is being built on it, so it is a shared cost centre rather than a foundation.

The word that carries the first property is interoperabilityThe property that any party can reach something without first joining a private arrangement with whoever runs it. A public road has it. However well surfaced a driveway happens to be, a driveway with a gate on it does not., and it is worth being precise about it, because it is the property most often claimed and least often present. A thing many parties use is not yet shared if reaching it requires a private agreement with whoever runs it, negotiated separately by each user on terms each of them cannot see. A private agreement of that kind is a supplier relationship wearing the vocabulary of infrastructure.

The part everybody assumes a layer changes is the part it does not. The layer does the work it does and hands back an answer. Whether to lend on that answer, on what terms, for how long and with what recourse, is still entirely the decision of whoever asked. A road that reaches a shop has not stocked the shelves.

Which layers does lending sit on, and why are they named by function here?

Four layers carry most of what lending actually rests on, and each of them does one thing. Take them in the order a loan meets them.

An identity layer establishes that a person is who they say they are. The verb matters more than the noun: against a layer a lender verifies, where without one it investigates. Investigating is open ended, expensive and inconsistent between lenders. Verifying is a question with an answer.

A payments layer moves money between accounts on an instruction, and it moves it in both directions. That is how a loan reaches the borrower at disbursalThe moment the loan amount actually leaves the lender and lands in an account. Everything that has to happen before it is covered separately., and it is equally how each repayment reaches the lender afterwards. A lender that had to build its own way of doing both would be building a small payment system before it could make its first loan.

A consented data layer carries data already held about a person somewhere else to a party that person names, on that person's direction. Nothing about this layer originates data. The layer moves what exists, on a direction, and the direction is the mechanism rather than the paperwork around it.

Somebody has already issued the record, so a document layer presents it to be checked rather than issued again. The saving here is not glamorous and it is very large: the collecting, the copying, the attesting and the re-checking of records that already exist somewhere in a definitive form.

Now the reason the four layers are identified by function rather than by name. The four layers are identified by what they do, and not by the name of any real system, scheme, registry, application or platform. That is a choice and not caution. Names change; operators change; the arrangement sitting behind a name changes without the name changing at all. A reader who has learned four functions can walk into any market, in any year, and ask which thing is doing which job. A reader who has learned four names has learned something with an expiry date printed on it in ink nobody showed them.

What does each layer take off a lender's cost?

Every specific saving below is one shape wearing different clothes. Take the shape first: work that used to happen once per person per lender now happens once and is reached by everybody, and a cost that used to repeat sits in one place instead. Hold that sentence and the four cases are almost boring, which is the point.

The identity layer removes the separate establishing, by each lender, of who somebody is. The document layer removes the collecting and re-checking of records that some issuer has already put into definitive form. The consented data layer removes the assembling of a financial picture by asking for it a part at a time and then trying to work out whether the parts are complete. The payments layer removes the building of a way to move money out and a way to collect it back.

Now the honest part. Something stays standing after all four layers have done their work. The lending decision remains, and no layer touches it. The funding remains behind it, and no layer supplies that either. Everything established earlier about funding still binds exactly as it did: the money that goes out of the door is money the lender itself borrowed, on which it pays a rate for a period, and which it owes back on a date whether or not the loan it funded comes back.

Put a number on where the reach actually stops, using a lender the record does contain. Rukmini Finance Limited, invented, carries assets under management of Rs 18,000 crore for the year and earns 14.50 per cent a year on that book. The interest earned is Rs 2,610 crore. The company has borrowings of Rs 14,400 crore on which it pays 8.50 per cent a year. The interest paid is Rs 1,224 crore. The difference, Rs 1,386 crore, is its net interest incomeWhat a lender has left in a year once the interest it paid is taken away from the interest it earned. The line itself was settled earlier and is used here rather than rebuilt. for the year. Its operating expenses come to Rs 540 crore, and set against that same Rs 18,000 crore lending book they are 3.00 per cent of it for the year.

The Rs 540 crore is the one figure on that list a shared layer can reach. It does not move the 14.50 per cent a year the lender earns, which is a price set in a market against a risk. The 8.50 per cent a year the lender pays is what the market charges that lender for money, and no layer moves it. The two rates and the gap between them are the business. The Rs 540 crore is the cost of running the business, and the running is what the layers touch. For scale, the Rs 540 crore is 20.69 per cent of the year's interest earned of Rs 2,610 crore, so this is not a trivial corner of the accounts. The operating line is also not the engine.

One thing has deliberately not been claimed, and it matters as much. The operating line comes as one number, split by no activity, and no layer's cost stands anywhere beside it, so how much of the Rs 540 crore a layer removes does not follow. A proportion supplied anyway would be the most quotable figure available and the least true.

How Financial Infrastructure Can Expand Access: by what mechanism exactly?

An account of plumbing usually stops being about plumbing at this point and starts making claims about people. Specificity is what keeps it from doing so. The mechanism has exactly two limbs, and both of them are about costs and records rather than about anybody's conduct.

The first limb is the cost of serving one more person. A lender serves somebody when what it costs to find out enough about them, to reach them at all, and to collect from them afterwards is less than what serving them earns. The comparison is a test, it is arithmetic, and it is run on every prospective borrower whether or not anybody writes it down. Now suppose the finding out, the reaching and the collecting all happen through layers that many lenders reach and that nobody rebuilds. The left hand side of the test falls. People who were on the far side of the line are now on the near side of it, and every one of those people is exactly the same person they were the day before.

The everyday version is a vegetable seller who will deliver to the door. He works out what the trip costs him against what the order earns. Three streets away is beyond it, so he does not go. Then the lane is surfaced and a shared handcart route opens, and now three streets away is inside the same arithmetic. Nothing about the household three streets away changed. The cost of getting there did.

The second limb is the record that did not exist. A person can direct that data already held about them be shared with a lender they name. Where previously there was nothing a lender could look at, there is now something assemblable. A cost that falls is no help at all when the problem is that there was never anything to read, so the second limb matters most exactly where the first matters least.

And here is the sentence the entire block was built to protect, so read it slowly: a person who was not served was not served because a cost exceeded a return, or because no route reached where they were, or because nothing existed that could be checked. Those are the three reasons, and they are facts about arithmetic, about geography and about records. Not one of the three is a statement about the person. Any reading that supplies a fourth reason of that kind has left the arithmetic, the geography and the records behind, and gone somewhere nothing can be checked.

A silence sits beside those three reasons, and it is deliberate. How many people any of this reached, in any year and in any market, is a count that has to come from a published series rather than from an argument about mechanism. Where such a series exists it belongs to the Reserve Bank of India, at rbi.org.in, and to its data site at dbie.rbi.org.in, and it is read there on the day it is needed.

What does the payments layer look like written down as numbers?

Everything above is shape. The record supplies figures in exactly one place and supplies exactly two of them, so the arithmetic here is a single division rather than something to be read off.

Setu Payments Limited, invented, carried 1,200 crore transactions in one stated year. Rs 3,60,000 crore of value passed through them. Divide, rather than quoting either total: Rs 3,60,000 crore over 1,200 crore transactions is an average transaction of Rs 300.00/-.

Now the reason that single number is worth more than both totals put together. A system whose average instruction is three hundred rupees is doing something categorically different from a system whose average instruction is three lakh, and almost everything worth knowing about either one follows from that one division rather than from the size of anything. Three hundred rupees an instruction indicates that the volume is enormous relative to the value, that the cost of processing one instruction has to be tiny or the whole thing collapses, that a failure of one instruction is an inconvenience rather than an event, and that a lender building on it is building around many small movements rather than a few large ones. Three lakh an instruction indicates the opposite of every one of those, and the two totals on their own would not have shown which of the two was in view.

The refusal that travels with this block is short and it is absolute. The charge a payment system may make, and any limit on that charge, is set by the Reserve Bank of India, at rbi.org.in, and both of them move. No fee, no interchange, no charge and no cap of any kind is stated here. How a payment system makes its money at all is covered separately.

What does a shared layer not do?

The list below is longer than most readers expect, and the length is the teaching. Every item on it is something people routinely assume a layer has taken care of.

A shared layer does not decide whether to lend. It does not set a price, or a period, or a repayment schedule. No layer creates a repayment, and none can make one arrive. A layer that shares data cannot share data that was never created in the first place, so a thin record stays thin, and that limit is absolute rather than a matter of degree. No layer tells a lender whether the picture it has assembled is complete. And no layer funds anything, the constraint that lender funding has turned on from the beginning.

Suvarna Commercial Bank Limited, invented, takes deposits. Rukmini Finance Limited, invented, does not and funds itself in the market instead. The difference in funding decides a great deal about each of them, and it decides nothing whatever about which layers either of them can reach or on what terms. Both lenders reach the same four layers on the same footing, and that is precisely the work the word shared is doing. What each of them then does with what the layers return is a different business, run at a different cost of money, taking different risks.

Summed in one line: the layers change what a lender can see and what seeing costs, and every decision on the other side of that is still the lender's own, made with its own money, at its own risk of loss.

What runs on a person's consent, and what follows from that?

The consented data layer does not work unless a person actively directs it to. Of the four, it behaves least like plumbing. The direction is not a formality bolted onto a technical arrangement. The direction is the arrangement.

A direction of this kind has four parts, and every one of the four is load bearing. There is a named party, the one the data goes to. There is named data: what goes and, just as importantly, what does not. There is a stated purpose, the only use the receiving party may make of it. And there is a stated period, after which the direction stops. Take away the named party and the direction has no destination. Take away the named data and it has no boundary. Take away the purpose and there is nothing to hold the receiving party to afterwards. Take away the period and a direction given once runs forever. Nobody giving one means that.

A direction that can be withdrawn is part of the mechanism, not an exception carved out of it, for the same reason that a tap that cannot be turned off is not a tap. Data already shared has already been shared. Withdrawal changes what happens next.

Five requirements sit directly underneath everything in this section and in the one after it, and none of them is stated here. Each of the five is set by the Reserve Bank of India, at rbi.org.in, and each of them moves. A value written out here would not go gently out of date, it would be wrong on the day it changed. They are drawn below as five rows with the authority printed inside each and the value column left deliberately empty. The reason for leaving a cell blank is worth saying plainly. A blank cell has never misled anybody, it puts the question in front of the reader instead of an answer of unknown vintage, and it points at the desk where the wording that counts is maintained.

What happens when a layer is unavailable or a record does not match?

Two things go wrong here and they go wrong differently, so keep them apart. A layer that is unavailable means the step which ran through it does not complete. A record that does not match the details presented means the check returns no answer, rather than returning a wrong one. The second distinction is the one people collapse, and collapsing it is how a mismatch turns into an accusation.

The consequence of either is the same and it is small. The transaction does not happen. The person is exactly where they were before, with a step outstanding and, importantly, with nothing attached to them by way of explanation. A record that does not match is a record, not a person, and the difference between those two sentences is the whole of why this block exists. Details are entered by people and by systems, they are transcribed, they are abbreviated, they change when somebody moves or marries, and every one of those produces a mismatch that describes a database and nothing else.

The mechanism is named and the route is named, and nothing beyond them. The grievance route a borrower has, and the time inside which it must be answered, are set by the Reserve Bank of India, at rbi.org.in, and both of them move. The row is drawn below with the authority inside it and nothing written in. Three things do not go stale: a route exists, it is a route rather than a favour, and finding it is the thing to do rather than something to be endured in silence.

How would a practitioner read a lender that sits on these layers?

Looking at a lender rather than at the layers is the position an analyst, a credit officer or somebody in a lending business is actually in. Four questions do most of the work, and none of them requires a single figure this record does not contain.

First, which of this lender's costs sit on shared layers and which sit inside its own walls? The question is about the operating line and nothing else. The useful version asks which activities the lender has stopped doing for itself rather than how much it saved. Almost nobody publishes the second.

Second, what does the lender still do itself that the layers cannot touch? On Rukmini Finance Limited that is the 14.50 per cent a year it earns on assets under management and the 8.50 per cent a year it pays on borrowings of Rs 14,400 crore, and it is also the whole of the decision about whom to lend to. A lender whose distinctiveness lives entirely in things a layer does for everybody has no distinctiveness.

Third, who is standing between the lender and the borrower, and on what footing? Where a lending service providerA partner that brings in a loan, or looks after it afterwards, on a lender's behalf, while never carrying the loan on its own books. The conditions such a partner works under are set by the Reserve Bank of India. brings in the loan or looks after it afterwards, the layers change what that party can do quickly and change nothing about who carries the loan. The requirements on such a party are set by the Reserve Bank of India, at rbi.org.in.

Fourth, what does the lender do when a layer is unavailable? A business with exactly one path through a layer it does not run has a dependency it did not choose. The day that path closes tests the lender rather than the layer.

None of these four questions is answered by knowing the name of any layer, and all four are answerable by somebody who knows only what the four layers do. That is the practical case for learning functions rather than names, and it is why what a credit information companyAn institution that keeps a record of loans made and of how they were repaid, and produces a credit score from it. The limits on what such a company may hold, and the way an entry is corrected, are covered separately. holds is kept firmly separate from the four layers: it is a different thing, built for a different purpose, and reached in a different way.

Why access cannot honestly be put on a slider

A control that moved access would fail on two grounds.

The first ground is editorial and it is the stronger of the two. The obvious control here would move the cost of serving one more person and show people crossing into or out of being served. A control whose consequence is how many people are reached puts somebody's access to credit on a slider and invites a reader to watch them cross a line, and it teaches nothing at all that the sentence on expanding access does not teach in words that can be checked. The slider costs something that sentence does not cost.

The second ground is evidentiary and it is decisive on its own. There is no coverage figure to move, no account count, no cost of serving anybody and no before and after of any kind. Every number such a control moved would therefore have been invented, and an invented number about access is exactly the kind a reader carries away as a fact about the country they live in. A figure a reader can move is a figure a reader believes.

In its place stand the one division the figures do support, worked in plain text above where anybody can check it, and a fixed drawing of the layers with the decision and the funding placed outside them.