An agent network is a workforce you rent. Every agent is a small business that has to cover its own float and its own time before it returns you a shilling. Model it like software and the economics read like margin that scales for nearly nothing: add a node, add reach, keep the difference. Price it as the labour force it actually is, with recruitment, training, float, churn and daily management, and the numbers change shape and so does the moat.
The mobile money industry says so in its own accounts. Providers describe the business as first and foremost an operating-expense business, driven by agent commissions, marketing and personnel, and they warn that in the start-up phase an operator should expect to invest six to eight times the revenue the service generates.7 That is a wage bill and a training budget. It is not a server bill.
Your agent network is not software with people bolted on. It is people with software bolted on.
So the decision this piece is built to serve is the one founders get backwards under pressure: when growth stalls, do you expand the network by signing more agents, or do you raise what your existing agents earn first. The install curve says expand. The labour view says the opposite, and it is usually right.
Why three lenses, and why these three
One model would hide the error that matters. So run three that fail in different directions.
A principal-agent lens, to see each agent as a decision-maker whose effort you do not control and cannot fully observe. A base-rate lens, to replace your signup curve with what agent networks actually earn, keep and lose. An agent-based lens, to see coverage as something that emerges and collapses rather than something you deploy. Where the three disagree is where your risk is hiding.
The behavioural layer folds into the base-rate card, because the bias here is founders reading a signup as a sale. The governance question folds into the blind spot. Three cards is the budget. A fourth would buy a footnote and cost a paragraph of argument.
The framework: pricing the network as people
1. The principal-agent read: an agent is paid to decide, not built to obey
Software executes the instruction you deploy. An agent decides, transaction by transaction, whether serving your customer is worth her time and her cash.
Read the industry’s own worked example. An agent earns roughly twenty-five cents for each cash-in she processes, whatever its size. On a good day she serves twenty customers depositing fifty dollars each and clears about five dollars. She is happy to earn two percent on a fifty-dollar transaction, which is a dollar, and unenthused to earn two percent on a nine-dollar one, which is eighteen cents, so she quietly declines the small ones or lets them wait.1 Her effort tracks her commission, not your growth target. That is moral hazard, and it is the defining property of labour that a node does not have.
The margins under that behaviour are thin by design. In Cameroon agents keep only forty to forty-five percent of the commission on a transaction, and many sit just above their break-even point, so a new tax or a commission cut pushes them under and they stop promoting you.2 When the incentive is wrong, the network does not respond to instruction. Safaricom found this out in 2013, when it tried to make agents hold more liquidity by raising commissions on high-value transactions. The behaviour barely moved, and it reverted to the old commission model.3 You cannot patch a workforce. You can only reprice it.
Assumes: agents are independent income-maximisers whose effort is unobservable and responds to the take rate.
Fits because: the same commission that pays for a fifty-dollar deposit pays for a nine-dollar one, so agents ration effort exactly where the model predicts.
Breaks when: the agent has no rival use of time or float, so effort does not track pay (a captive rural monopoly, a salaried till).
Counteracts: the base-rate card, which treats agents as a population rather than a chooser.
May reinforce: the agent-based card, since individual effort choices aggregate into coverage.
2. The base-rate read: forecast active agents and agent income, not signups
Stop projecting from the number of agents you have registered. Anchor on what agent networks actually earn and how many survive.
The reference class is unglamorous. Median monthly agent profit runs at about seventy dollars in Kenya, seventy-eight in Uganda and ninety-five in Tanzania.3, 4 In Tanzania only four percent of agents are outright unprofitable, yet liquidity gaps alone cost the median agent five transactions a day, which is fourteen percent of volume, income that simply evaporates because the till ran out of cash or e-float.4 These are small businesses living close to the line, not endpoints humming at zero cost.
A registered agent who never trades still cost you money to sign and still returns you nothing. The gap between registered and active is the oldest wound in this industry, and it is the exact size of the illusion that agents scale like software. When agent income falls, agents churn, and a network measured in signups reverts to its active base. So build the model the honest way: forecast active agents, forecast agent income per day, then apply the market’s churn rate to both. A signup is a job opening you have not yet filled, and an empty one pays nothing.
Assumes: your network behaves like the observed population of African agent networks, not like your pitch deck.
Fits because: median profits, unprofitability rates and the registered-active gap are stable and well measured across markets.
Breaks when: your model is genuinely new (salaried agents, a captive use case) and the historical class does not apply.
Counteracts: founder optimism that reads a signup curve as a revenue curve.
May reinforce: the principal-agent card, because falling income is what triggers the effort withdrawal it describes.
3. The agent-based read: coverage emerges, and it has a viability threshold
Coverage is not deployed. It emerges from thousands of local decisions: where an agent sets up, whether she rebalances her float, whether she stays. That makes the network a complex system with a threshold, and the threshold is geographic.
Frontier agents face lower transaction volumes and higher costs, especially liquidity management, because they sit far from a bank branch or rebalancing point. So viability is easy where density and demand already exist and close to impossible in the thin last mile, and providers rationally cluster agents where it is easy and leave the edges bare.6 Liquidity management alone consumes twenty to thirty percent of an agent’s total costs.5 This is a living distribution, not a server rack, and it does the load-bearing work: in thirty markets there were ten times more active agents than bank branches, and agents conduct ninety-five percent of physical cash-in and cash-out.6
The control law is stated plainly in the operator playbooks. Sign too many agents and none of them clear enough revenue to bother, so drop-out spikes. Sign too few and liquidity management breaks. The instruction is to match the growth of the agent network to the growth of the customer base, not to the growth of your ambition.5
Defensibility inverts here too. If your agents are non-exclusive, the money you spend identifying, training and equipping them becomes a public good your competitor free-rides on the moment their customer walks into the same shop.9 Agent count is not a moat. The moat is exclusivity, or owning the rail underneath the agents. Wave took the second path: rather than out-signing incumbents, it cut fees to free deposits and withdrawals with a flat one percent transfer, ran a leaner network of about a hundred and fifty thousand agents against three thousand staff, and raised a hundred and thirty-seven million dollars of development-finance debt to deepen that rail.8
Assumes: coverage and liquidity are emergent from many local agent choices, with a density-and-demand threshold below which a zone cannot sustain agents.
Fits because: viability is observably geographic and drop-out rises when agents are packed into thin markets.
Breaks when: a rail change removes the cash step (app-to-app transfers, interoperable float) so density stops mattering.
Counteracts: the deploy-more-nodes reflex the software framing produces.
May reinforce: the base-rate card, since sub-threshold zones are where the churn statistics come from.
The levers, from a dashboard edit to a liquidity rail
Cheapest and most reversible first.
Change what you count. Move active agents, defined as traded in the last thirty days, and agent income per day to the top of the board deck. Delete registered-agent count from the headline. This is a dashboard edit and it reprices every decision below it.
Find your break-even in transactions per day. For each zone, work out how many daily transactions an agent needs to clear costs, then compare it to what your agents there actually do. This is arithmetic you already have the data for.
Reprice the take rate in one zone. Raise commission or add an income floor in a single district and watch active density move before you touch the rest.
Fix liquidity where volume already exists. Rebalancing support, float credit and super-agents lift the fourteen percent of transactions that liquidity gaps currently lose you, at zero new signups.
Own the rail, last. Liquidity software, master agents and a push for exclusivity are the expensive, hard-to-reverse move. Earn the right to it with the four cheaper levers first.
What to change before you fund the next zone
Do now, T+0 to T+14. Swap the north-star metric to active agents and agent income per day. Pull the churn base rate for your market. Compute break-even transactions per day per agent, by zone. Reversible, and it dominates in every scenario.
Hedge, T+14 to T+28. Pilot a take-rate floor plus liquidity support in one high-density zone, instrumenting agent income before and after. It is cheap tail insurance: contained, reversible, and it tells you whether repricing beats expansion before you bet the round on it.
Defer and trigger. Expansion capital into new zones is the irreversible spend, so pre-commit the trigger and do not move before it fires. The trigger: add agents in a zone only once the agents already there clear break-even transactions per day and their income is rising. Freeze net-new agent spend in any zone where your existing agents sit below break-even. Never expand on a signup curve.
What usually happens to an agent network next
Match the reference class on structure, not on sector. The right comparison is any rented labour network where each worker must clear a local break-even and can quit, be poached or be shared: distribution reps, delivery riders, informal sales forces. Not other apps.
The base rate for that class is unkind. Agent income is flat to falling in mature markets, churn tracks income down, registered counts drift far above active ones, and viability stays stubbornly geographic. Now adjust for your present state: your zone density, your fee level, whether your agents are exclusive, and how far they sit from liquidity. Then subtract the counterfactual before you credit yourself with growth, because a share of your signups was always going to sit dormant no matter what you did, exactly as it does for everyone else.
Matrix-break flag. If cash-in and cash-out get displaced by app-to-app transfers or interoperable float, the way some markets are moving toward agent-light models, the labour weight in this whole analysis drops and the software framing gains ground. That is a regime change, not a good quarter. Watch for it, and re-run the ensemble the moment the cash step starts to disappear from your own volumes.
What pricing agents as labour still cannot see
This ensemble reads the supply side well and the demand side not at all. It can tell you whether an agent will show up and turn a profit. It cannot tell you whether the customer wanted the transaction in the first place, and a perfectly run agent in a market with no demand is still a loss. It also treats agent income as the master lever, when in some markets the binding constraint is float and liquidity access rather than commission, and no take-rate change fixes a cash-supply problem. And it goes quiet on the governance question of who owns the agent relationship when a master agent sits in the middle.
Name the one action that survives all of that. This week, put active-agent count and agent income per day at the top of your board deck, and stop funding new agents in any zone where the ones you already have are underwater. If you do nothing else with this piece, do that. The network was never a line of code you scaled. It was always a payroll you were pretending not to run.
Sources and notes
- GSMA, “Getting the Agent Commission Model Right,” Mobile for Development. Source for the agent’s per-transaction economics: roughly twenty-five cents per cash-in regardless of value, about five dollars on a twenty-customer day, and the reluctance to spend time on low-value transactions (a dollar on a fifty-dollar transaction versus eighteen cents on a nine-dollar one). gsma.com. Verified by GET: body contains the worked commission example.
- Institute of Development Studies (ICTD), “Why Cameroon’s mobile money agents are struggling to increase revenue.” Source for agents receiving on average forty to forty-five percent of the commission and earning just above break-even, so that a mobile money tax can push them under. ictd.ac. Verified: body contains the 40 to 45 percent share and the break-even language.
- MicroSave Consulting (MSC), “What is driving agent churn in the mature East African markets?” Source for median monthly agent profit of seventy dollars in Kenya and seventy-eight in Uganda, agents complaining of insufficient income, and Safaricom’s 2013 attempt to change agent liquidity behaviour by raising high-value commissions, which had little effect and was reversed. microsave.net. Verified: body contains the $70 and $78 medians and the Safaricom commission episode.
- NextBillion, “Problems and Potential in Tanzania’s Mobile Money Ecosystem: Highlights from Helix’s Agent Network Accelerator survey.” Source for Tanzania median profit of ninety-five dollars per month, four percent of agents unprofitable, and liquidity gaps costing a median of five transactions per day (fourteen percent of median daily transactions). nextbillion.net. Verified: body contains all three figures.
- International Finance Corporation, “Liquidity Management for Mobile Money Providers,” IFC mobile money toolkit, World Bank Group. Source for the control law that too many agents leaves each with insufficient revenue and high drop-out while too few worsens liquidity, the instruction to match agent-network growth to customer growth, and the CGAP finding that liquidity management consumes 20 to 30 percent of an agent’s total expenses. Toolkit PDF. Verified via pdftotext: body contains the drop-out passage and the 20-30% figure.
- CGAP, “Proximity Matters: Improving the Viability of Frontier Agents.” Source for frontier agents facing lower volumes and higher liquidity costs so that viability is geographic, that thirty markets had ten times more active agents than bank branches, and that agents conduct ninety-five percent of physical cash-in and cash-out. cgap.org. Verified: body contains the frontier-cost passage and the 95 percent and ten-times figures.
- GSMA, “New GSMA publication on mobile money profitability,” Mobile for Development. Source for mobile money being first and foremost an operating-expense business driven by agent commissions, marketing and personnel, and the expectation that operators invest six to eight times the revenue generated during the start-up phase. gsma.com. Verified: body contains the “six to eight times” and OPEX passages.
- TechCabal, “Exclusive: Wave raises $137 million debt to expand mobile money” (June 2025). Source for Wave’s free deposits and withdrawals with a flat one percent transfer fee, its network of more than one hundred and fifty thousand agents and three thousand staff, and the one hundred and thirty-seven million dollar development-finance debt raise. techcabal.com. Verified: body contains the fee structure, the agent and staff counts, and the raise.
- CGAP, “Branchless Banking Interoperability and Agent Exclusivity.” Source for the free-rider problem in non-exclusive networks: a provider that identifies, trains and equips agents sees competitors piggyback on that investment, which weakens agent count as a moat. cgap.org. Verified: body contains the “identifying, training, and equipping agents” free-rider passage.