Your bank has the transactions. Does it understand the business?

Almost every fintech conference in Africa features a panel where data scientists declare that mobile-money and payment-gateway statements have solved small-business credit underwriting. The claim rides a global wave: open-banking mandates, API aggregators and alternative-data scoring have convinced investors on three continents that visibility equals understanding. Because a lender can see every digital shilling hitting a merchant's till, the argument goes, traditional loan officers and cash-flow interviews belong to the past. That claim confuses watching payment velocity with understanding business economics, a blind spot that leads automated lenders to over-lend to noisy pass-through traders while starving profitable, cash-heavy firms.

The blind spot is not new; the technology has merely given it scale. Back in April 2015, a Consultative Group to Assist the Poor financial-diaries study profiled Renato and Hecinta, a couple in northern Mozambique combining seasonal smallholder farming, casual labour and informal trading across several formal and informal savings tools. No single financial statement captured their real economic life, and any lender reading only their harvest receipts would have misjudged both their peak earning capacity and their off-season cash crunches. Eleven years later, an urban Kenyan hardware dealer handing over a 12-month M-PESA statement presents the same analytical trap at digital speed. Inflows mix genuine customer sales with inter-account float sweeps, supplier refunds, borrowed working capital from rival apps and family remittances, while outflows jumble inventory purchases with school fees and clinic bills.

Reconciling a digital receipt before calling it revenue.

Open the cash book of an illustrative Nairobi retail shop whose mobile-money statement shows KSh300,000 in gross monthly inflows. An automated ingestion engine that treats gross credits as top-line turnover will happily approve a loan sized for a KSh300,000 enterprise. Look closer at the rows, however, and KSh60,000 of that inflow is the shopkeeper sweeping float between her own accounts, while another KSh40,000 is the disbursement of a short-term digital loan from a competing lender. Genuine customer sales sit at KSh200,000, meaning the naive ingestion rule overstated revenue by 50 per cent before subtracting a single shilling for stock, rent or the household draw that feeds the owner's family.

The opposite mistake damages growth just as badly. A timber dealer who settles 70 per cent of trade in paper cash, or spreads collections across three different paybills, looks tiny inside a single wallet statement, just as a seasonal grain buyer stocking up before the rains looks dormant during her two-month purchasing window. Mistaking a single payment channel for the whole business loads bad debt onto over-borrowed digital transactors while turning away solvent builders. Both errors flow from the same root: the statement records where money moved, not why.

Fixing the training labels before polishing the algorithm.

Pointing a more complex machine-learning model at corrupted accounting labels only automates the error faster. If a credit model is trained on historical files where gross wallet inflows were mislabelled as business turnover, it will faithfully copy the distortion. Worse, if past good-repayment tags include borrowers who quietly took high-interest loans elsewhere to pay the bank on time, the algorithm learns to reward pyramid borrowing. And a generative assistant asked to summarise relationship-manager notes will spin unverified borrower claims into a polished credit memo that lacks a single verified invoice.

Before spending money on complex alternative-data scoring, a disciplined credit desk puts two simpler filters in place. First, deterministic accounting rules strip out self-transfers between linked national ID or phone numbers and separate known loan-disbursement paybills from real customer sales. Second, loans above a set shilling threshold pair the cleaned statement with a 15-minute stock-cycle verification call or supplier-invoice check, benchmarked against typical margins for that trade sector. Neither filter requires a data scientist. Both require the institutional humility to admit that a transaction log is testimony, not truth.

Auditing the borrowers your data excluded.

Every credit book suffers from survivorship bias. The bank holds detailed repayment logs on the applicants it approved and almost nothing on how rejected applicants, or traders who walked away from rigid 30-day schedules, performed afterwards. A fixed 90-day statement lookback may capture twelve stock turns for a vegetable vendor while missing the six-month harvest cycle of a cereal distributor altogether.

Before expanding automated SME credit limits in the first half of 2027, the chief credit officer and head of SME banking should run a forensic audit across 300 files: 100 performing loans, 100 stressed or defaulted accounts and 100 borderline rejections, reconciling the model's automated inflow estimate against verified supplier invoices, cross-market debt and net cash left after family expenses. If the gap between automated turnover estimates and verified net margin tops 20 per cent at the second-quarter credit committee review, freeze the model expansion and fix the accounting rules first. The banks that win Kenya's SME market in this decade will not be the ones that saw the most transactions. They will be the ones that understood the businesses behind them.