Who can stop an automated financial decision?

When a digital lender or commercial bank rejects a working-capital application in milliseconds, the borrower rarely lacks a polite notification. What she lacks is a doorway to anyone in the building with the power to change the outcome. Modern fintech interfaces are built to make onboarding instant and appeals practically impossible, stranding disputed decisions in a dead zone between the smartphone screen and the balance sheet underneath.

That dead zone is no longer merely a service failure. It is becoming the specific thing financial regulators look for. The European Union's AI Act classifies credit scoring as a high-risk use that must carry human oversight; Kenya wrote the same principle into Section 35 of the Data Protection Act six years ago; and Kenya's Office of the Data Protection Commissioner has already shown its willingness to fine digital lenders over how automated collection machinery treats data subjects. Three legal systems, drafted on three continents, have arrived at one operational demand.

The cost of ignoring it now has a price tag. In October 2024, the United States Consumer Financial Protection Bureau ordered Apple and Goldman Sachs to pay more than $89 million in redress and penalties after uncovering a breakdown in Apple Card dispute handling. Cardholders who challenged transactions through Apple's interface assumed they had triggered a formal banking investigation, yet Apple failed to forward tens of thousands of those disputes to Goldman Sachs, and the bank failed to properly investigate the files it did receive.

The app on the glass worked smoothly while the statutory recourse pipe underneath was unhooked. Every African bank and digital credit provider scaling automated underwriting faces the same structural risk when one vendor aggregates mobile-money statements, an internal team builds a scoring model, a credit committee sets a cut-off and a customer-service bot delivers the rejection. Once the underlying source data is wrong, nobody at the call centre holds the keys to reverse the machine.

Separating the statistical score from the credit policy refusal.

Take a Kenyan sole trader applying for a KSh250,000 inventory loan. An ingestion engine scrapes her M-PESA and bank statements to calculate monthly turnover, debt-service ratios and cash volatility. A machine-learning model converts those figures into a probability-of-default score, and a hard-coded policy rule decides whether that score triggers instant payout, manual review or flat rejection. Three different machines made three different decisions, and each can fail independently.

Management teams routinely blur those steps when a creditworthy trader gets turned away. A statistical model can perform exactly as validated while a rigid policy rule rejects the applicant because a single disputed water bill or an uncategorised bulk transfer tripped a hard exclusion filter. Moving a cut-off threshold is a commercial credit-policy decision, not a mathematical law, and it requires a named credit owner and an audit trail showing who was locked out. Bolting a generative language model onto the front end to write sympathetic rejection messages only muddies the water: unless that tool is strictly locked to read the exact deterministic variable that tripped the rule, it will write a fluent, polite and completely invented explanation for why the loan was refused.

Making statutory review operationally real.

Under Section 35 of Kenya's Data Protection Act, 2019, every citizen has the right not to be subjected to a decision based solely on automated processing that produces legal or similarly significant effects, together with the right to ask for human reconsideration after being told a decision was automated. A statutory right on paper means nothing unless the lender builds an internal workstation capable of executing it.

When our sole trader challenges a rejection, pointing out that a flagged default at a credit reference bureau was settled three months earlier, or that a two-week dip in till receipts reflected a shop renovation, a lawful human review does not mean promising her a loan. It means routing her proof to an officer who can inspect the raw input variables, separate the model's risk estimate from the policy rule, fix the corrupted data field at the source and re-run the decision, without forcing her to restart the application across three departments. That is a workflow, a permission set and a staffing decision. No general counsel can conjure it after the complaint arrives.

Mapping human authority to the machine's perimeter.

Accountability breaks down in automated lending because ownership is sliced horizontally across departments while the customer experiences the failure as one vertical wall. When a bad rejection happens, the data team blames the bureau feed, the modelling team points to its validation curve, the credit desk points to the risk appetite statement and customer support points to a helpdesk ticket it has no permissions to override.

Even in a lean fintech where one executive wears several hats, four authorities need names attached. Credit policy thresholds belong to the chief credit officer. Source data accuracy and permanent record correction sit with the data lead. Model validation belongs to risk. The exception queue, and the staff empowered to adjust limits or pause collections, sit with the chief operating officer. Manual exception handling is the price of running an automated loan book lawfully: lenders that slash underwriting headcount to boast a 95 per cent straight-through processing rate inevitably starve the remaining 5 per cent of complex, high-value appeals, and call-centre clerks measured solely on tickets closed per hour will fire off boilerplate rejections to clear their queues.

Before scaling any automated credit engine in 2027, the risk and operations leads should run a live stop-test: inject a synthetic borrower file carrying a deliberate bureau error, then track how many hours and system permissions it takes for a frontline challenge to fix the source record and update the customer's account. If the answer is measured in weeks, the machine has no brake, and the next regulator to ask will not be conducting a survey.