A better crop forecast is not a loan.

Every planting season, agritech vendors walk into commercial bank headquarters in Nairobi, Kampala and Accra carrying high-resolution satellite maps that glow in shades of emerald green. The pitch rides two global currents at once: a boom in earth-observation constellations that has made imagery cheap, and a flood of climate-finance capital hunting for bankable African agriculture. Using multispectral imagery, soil-moisture sensors and machine-learning yield estimators, the argument goes, a bank can now predict a smallholder's maize or tea harvest from space and automate agricultural credit at minimal cost.

That pitch overlooks a basic rule of rural finance. An orbital photograph of a healthy maize field does not pay for fertiliser in March, hire a lorry when feeder roads flood in July or compel a cash-strapped broker to pay on time in October. In a May 2014 study published in the Quarterly Journal of Economics, economists Dean Karlan, Robert Osei, Isaac Osei-Akoto and Christopher Udry ran multi-year field experiments with farmers in northern Ghana, testing cash grants alongside rainfall index insurance. Their findings dismantled the lazy assumption that capital alone is the binding constraint in African agriculture: risk exposure and working-capital timing drove entirely different farm decisions, and farmers increased on-farm investment significantly only when they held credible protection against catastrophic weather. A crop forecast estimates what the soil might produce. A bankable agricultural loan requires financing the input window, insuring the weather gap and locking down the buyer who actually pays out the cash.

Following the shilling across the biological calendar.

Software operates in milliseconds while agriculture runs on an unforgiving biological clock. When a commercial bank takes six weeks to validate a credit file and disburses an input loan two weeks after the long rains have started, the farmer misses the planting window and buys overpriced seed on the spot market. Even if the predictive yield model is 95 per cent accurate in April, the late disbursement guarantees a stunted harvest in September. In farming, an accurate model with slow plumbing is indistinguishable from a wrong one.

Lenders who size loans against gross harvest estimates also confuse crop volume with net repayment capacity. Walk through the seasonal cash book of an illustrative five-acre maize farmer in Trans Nzoia whose satellite-assisted forecast projects KSh200,000 in gross harvest sales. Producing that crop consumes KSh120,000 in certified seed, fertiliser and weeding labour; harvesting, shelling, gunny bags and transport to the depot swallow another KSh30,000. Only KSh50,000 remains before the household pays school fees or clinic bills. A credit algorithm that advances KSh80,000 against a KSh200,000 gross forecast has engineered a guaranteed default, even if the weather cooperates and every cob reaches the store.

Separating the satellite pixel from the insurance payout.

Because weather shocks terrify credit committees, lenders frequently bundle satellite forecasts with index-based crop insurance, assuming the policy transfers climate risk off the balance sheet. In practice, index insurance carries a stubborn structural hazard known as basis risk: the gap between what a remote weather station or satellite pixel records and what happens on a specific two-acre shamba.

In a March 2024 paper in Q Open, researchers Digvijay S. Negi and Bharat Ramaswami showed how basis risk systematically depresses demand for catastrophic rainfall insurance among smallholders. A satellite index may record adequate aggregate rainfall across a ten-kilometre grid over ninety days, triggering no payout even when all that rain fell in two violent deluges that washed away the topsoil and left the crop parched during the critical flowering week. Conversely, when drought triggers a payout that flows straight to the bank to extinguish the principal while leaving the farmer with nothing to replant, the bank rescues its quarterly non-performing-loan ratio while killing the customer relationship, and with it every future season's business.

Comparing orbital models against an anchor buyer's ledger.

Before buying proprietary satellite feeds, a credit committee has to look at how each tool fails on the ground. A simple cap tied to an anchor buyer's historic delivery ledger, say 40 per cent of a farmer's three-year minimum delivered tonnage, works reliably until a farmer side-sells to a roadside broker paying cash at the farm gate. Satellite yield models extend visibility across unvisited farms, yet struggle with cloud cover and intercropping where beans and maize share a row, and remain blind to post-harvest price crashes. Bolting a generative agronomic chatbot onto SMS or WhatsApp creates real conduct exposure: if the model invents a pesticide dosage or recommends the wrong planting depth for a local micro-climate, the lender who promoted the bot owns the ruined crop's reputational damage. Automated scripts releasing digital input vouchers to agro-dealers can be drained overnight by collusive phantom redemptions unless locked behind verified inventory receipts and hard shilling limits.

Compare the total landed cost of that software stack against three years of verified delivery receipts from a dairy cooperative, tea factory or grain offtaker, paired with a tripartite check-off agreement and a seasonal repayment schedule charging zero principal during the growing months. That basic structure gives a credit committee harder repayment evidence than any satellite image, at a tenth of the cost. At the close of the first harvest cycle in the third quarter of 2027, the agribusiness and credit leads should reconcile 500 satellite-scored loans against actual farm-gate cash flows, measuring net borrower cash retained after harvest alongside the exact share of weather losses left uncovered by index triggers. A sharper forecast earns its subscription only when it puts seed in the ground on time and leaves the farmer solvent after the lorry drives away.