Users and jobs: the job was sending money home

Start the teardown with the decision question. Imran's is: "What should Mangrove copy from mobile wallets to move cash-on-delivery customers to digital payments, and should it ever build its own wallet?" About 35% of Mangrove's app orders are paid in cash at the door.

Then the sides and their jobs. A mobile wallet has at least five: the sender, the receiver, the agent who swaps cash for balance, the merchant who accepts payments, and the operator who runs the ledger under a regulator's rules.

The most useful finding comes from history. Safaricom and Vodafone piloted M-Pesa in Kenya in October 2005, funded through a UK government challenge fund, to handle loan repayments for customers of a microfinance lender. When it launched fully in March 2007, the job people hired it for was sending money to family. Before M-Pesa, many people sent cash with a friend or a bus driver; a postal money order cost roughly US$3 to US$6 and was slow; and only about 30% of adults in East and Southern Africa had a formal bank account (Mbiti and Weil, 2016). The product found its real job after launch, a reminder that a teardown should look at what people do with a product, which may differ from what it was built for.

bKash, launched in July 2011 as a subsidiary of BRAC Bank, served the same job for a different crowd. The GSMA's 2013 case study describes workers who moved to cities sending money to rural families. UPI, India's Unified Payments Interface, launched as a pilot on 11 April 2016 with 21 banks. Its job was wider: pay anyone with an account at any bank, instantly, including a shop, by scanning a QR code.

How the money actually moves

In agent wallets such as M-Pesa, the balance on your phone is e-money. When you hand an agent $50 in cash, nothing is created: $50 of the agent's e-money balance, called float, moves to your wallet, and your $50 goes into the agent's drawer. When you send money, two numbers change in the operator's ledger. When the receiver cashes out, her e-money moves to the village agent, who gives her cash from his drawer, minus a fee.

The real money behind all that e-money sits in a bank. The GSMA's case study says the central bank required bKash to hold the full value of customer balances in a designated BRAC Bank account that the central bank monitors. So the total never changes when people send money to each other; only the names next to the numbers change.

Now look at the agents. City agents keep receiving cash and giving out float, so they end up with full drawers and no float. Village agents keep giving out cash and receiving float, so they end up with the opposite. Both must rebalance, usually by depositing or buying cash through a bank or a larger agent. This is called liquidity management, and it is an operations product in its own right: alerts when an agent is running low, routes for the people who move cash, and rules about who can rebalance with whom. The details of each operator's rebalancing are not public; the imbalance itself follows from how the ledger works.

UPI works differently, and the difference explains its business model. There is no e-money. Each user's UPI ID is tied to a bank account and a mobile number checked against the bank's records, and money moves from one bank account to another over a network run by the National Payments Corporation of India (NPCI). Any UPI app can reach any bank account, and one app can hold several accounts. So an app like PhonePe or Google Pay is a front end: it shows you the screens, but your money stays in your bank. That makes switching apps easy, which matters for how they earn.

Agents: the branch network nobody had to build

A bank branch costs a lot to open. An agent is an existing shop that adds a phone, a notebook and some float. M-Pesa's agent network grew from about 450 outlets in mid-2007 to more than 18,000 by April 2010 (Mbiti and Weil). By the year to March 2025, Safaricom reported about 299,000 M-Pesa agents in Kenya and about 35.8 million customers who had used M-Pesa in the last month. bKash had about 30,000 agents after 17 months, almost one for every two rural villages according to the GSMA in 2013; it now reports more than 350,000 agents, more than 83 million registered customers and more than 900,000 merchants.

The agents create the product's main loop. More agents near people make cash-in and cash-out easy. Easy cash makes people trust the wallet, so more people use it. More users mean more transactions per agent, so each agent earns more commission, and more shops want to become agents. The same loop explains why a late copycat struggles: the first wallet already has an agent on every street, and its users can send to everyone they know.

The loop has one weak point, and it is physical. If a customer walks to an agent and the agent has no cash, the wallet has failed her, whatever the app looks like. That is why liquidity, agent training and agent fraud controls take as much product work as the customer app. For PMs who build internal tools, the agent's menu is a frontline tool used all day by shop staff, with the same needs as a store scanner: speed, clarity, working on a cheap phone and on patchy signal (Designing for frontline work). Customers of the bank-led wallet used a numbered USSD menu that worked on handsets costing about $15 in 2013.

Fees, and who pays for "free"

Follow the money. In 2013 the GSMA recorded bKash's prices as: registration and cash-in free, sending money a flat Tk 2 (about 2 US cents) whatever the amount, and cash-out 1.85% of the amount. The pattern is common: putting money in is free, moving it is cheap, and taking it out costs the most, because cash-out pays the agent for keeping cash. Customers notice. A 2018 study reported by The Daily Star found that 53% of the households surveyed named high charges as a major disadvantage of mobile money. In October 2021 it cut the cash-out price to Tk 14.90 per 1,000 (1.49%) for up to Tk 25,000 a month from a customer's chosen "favourite" agent, which also gives each customer a reason to stick to one agent.

M-Pesa's prices push the other way at the small end. Since January 2021, Safaricom's tariff makes transfers of KES 100 and below free; sending KES 101 to 500 costs KES 6. Small payments are where cash is hardest to beat, so the first ones are free. Larger transfers, cash-out and business payments carry the fees. In the year to March 2025, M-Pesa earned KSh 161.1 billion, 44.2% of Safaricom's service revenue in Kenya, with about 38 chargeable transactions per active customer by Safaricom's measure. Credit sits on top: Fuliza, an overdraft on M-Pesa, launched in January 2019. A wallet that sees every payment a person makes can judge small loans, the same pattern you will see in Square's loans in pm-43.

UPI apps face the hardest version of the question. Since 2020, merchants have paid no fee on most UPI payments, and users pay nothing to send. Volume exploded: about 20 billion UPI transactions in August 2025 alone. In July 2025, a month of about 19.5 billion, PhonePe processed about 46% by number and Google Pay about 36%. But PhonePe reported a net loss of about ₹1,727 crore (roughly US$200 million) for the year to March 2025, on revenue from operations of about ₹7,115 crore. Its lending and insurance distribution, while small, roughly tripled. The public summaries we read don't explain fully where its payments revenue comes from; likely sources include merchant payments by card and wallet, bill payments and payment devices for shops (inference). NPCI also plans to cap any app's share of UPI volume at 30%, with the deadline pushed to the end of 2026.

Rules can change the model overnight. On 14 September 2026, India's finance ministry notified that a merchant fee may be charged on UPI and RuPay debit payments above ₹2,000; payments up to ₹2,000 stay free, and NPCI is to decide the rate and which merchants pay. When this lesson was written, those details were not yet set. For a teardown, this is the moat lens in action: when the network is public infrastructure and the price is set by regulation, apps compete on reach and design, and they earn from what they sell on top.

Trust, fraud and why wallets beat banks

A wallet has to be trusted by people who have never had a bank, so trust is a product feature. Three mechanisms show up in all three systems. Identity: M-Pesa checked customers against Kenya's national identity cards at registration, and UPI ties each ID to a bank account and a mobile number the bank has checked. Confirmation: both sides get a message for every transfer, and well-designed flows show the receiver's registered name before you confirm, so a mistyped number is caught. Limits: smaller limits for lightly checked accounts reduce what a thief can take.

Fraud follows the easiest path, so controls change as fraud does. Common patterns include messages claiming money was "sent by mistake, please return it", fake agents, stolen SIM cards and requests that trick people into approving a payment. UPI shows a control being removed because fraud kept using it. A UPI "collect request" let anyone ask you to pay; you approved it with your PIN. NPCI capped these pull requests at ₹2,000 in 2019, fraud continued, and from 1 October 2025 NPCI ended person-to-person collect requests entirely. Every person-to-person payment is now started by the payer, typically by scanning a QR code. A useful feature died because its fraud cost was higher than its value.

Why did wallets win where banks had not? Read through the lenses: distribution (agents in existing shops instead of branches), the account (a phone number instead of forms and a minimum balance), access (menus that work on the cheapest phones), price (cheap or free for the small amounts people send most), and rules (Kenya's regulators gave M-Pesa a special licence, since the banking law did not cover non-banks). The outcome was large. Suri and Jack (2016) estimated that access to M-Pesa lifted about 194,000 Kenyan households, about 2%, out of extreme poverty; a 2019 paper disputes the finding, so treat the number as contested.

The moat follows from the loop. The two agent wallets have network effects (everyone you pay is already there) and agent networks that took years to build. UPI apps have neither on their own, because the network belongs to everyone; their defence is habit, merchant relationships and the products they sell on top.