What does your bonus scheme teach your algorithm?

Bankers often talk about machine-learning bias as an accidental mathematical glitch cooked up by data scientists in a back room, and the global governance industry has obliged with ethics boards, fairness toolkits and model charters. In commercial lending, however, algorithmic distortion usually starts somewhere no framework looks: inside the human-resources department and the quarterly sales incentive table. Software does not invent corporate culture on its own. It studies what an institution rewards, reads the historical files created by staff chasing those payouts and scales that behaviour across the balance sheet.

The canonical case predates the AI boom and is the more instructive for it. When United States regulators fined Wells Fargo $185 million in September 2016 after staff secretly opened more than two million unauthorised deposit and credit-card accounts, the root cause was not a rogue algorithm. It was a ruthless cross-selling bonus scheme that punished branch workers who missed daily quotas and rewarded raw volume regardless of customer need. Now drop a modern predictive underwriting model and an automated marketing engine onto a historical database built under that kind of sales pressure.

If an African bank or digital lender pays generous quarterly commissions on gross loan disbursements while credit losses, restructured rollovers and customer complaints surface nine months later, its algorithms will quickly learn to optimise for fast debt traps rather than sustainable lending. The machine is not biased against your customers. It is perfectly faithful to your incentive table.

Separating a booked disbursement from a profitable relationship.

Every loan has two timestamps that sit months apart. Disbursement happens on day one, booking an immediate origination fee and a green tick on the sales dashboard. Funding costs, collection calls, regulatory complaints and bad-debt write-offs trickle in over the next two to four quarters.

Anchoring variable pay to day-one volume pours fuel onto automated lending. A digital credit engine can approve ten thousand loans over a weekend, long before the first cohort shows distress. Worse, when relationship officers or automated renewal rules are rewarded for keeping headline non-performing-loan ratios below a target ceiling, the easiest way to hit the number is evergreening: issuing a fresh top-up or repeated paid extensions to bury an underlying default. The ledger books a performing asset and a fresh fee while the borrower sinks deeper into insolvency. Every one of those transactions then enters the training data as a success.

Auditing the human fingerprints baked into the training data.

Data scientists often treat historical loan files as objective ground truth, yet every row in a bank's database was shaped by past human incentives. If branch officers five years ago quietly inflated self-employed applicants' turnover figures by 20 per cent to push borderline files past the credit committee and hit December bonus targets, a model trained on those files will treat inflated income ratios as normal. If a borrower repaid on time only by selling a dairy cow or borrowing from an informal shylock, the repayment column records a clean outcome the algorithm eagerly tries to repeat.

The same distortion corrupts the back office. Hand a support team a generative assistant while measuring clerks strictly on tickets closed per hour, and staff will use the tool to fire off polished two-paragraph closure emails that never fix the underlying billing error. Deploy an automated collections script told to maximise early recoveries without conduct guardrails, and the software will hound distressed borrowers or auto-close aged complaints because a crude metric rewarded clearing the queue. In each case the technology executed flawlessly. The specification was the defect.

Aligning the paycheck timeline with the loan's survival.

Fixing a skewed algorithm starts with fixing the human scorecard, and with protecting the people who pull the brake. If a credit analyst or branch manager is penalised in her weekly appraisal every time she flags a suspect model approval or escalates a corrupted data field, overrides will dry up and executives will mistake silence for model perfection. A rising override rate is never proof of staff rebellion; it is a signal for the chief risk officer to sample both overridden files and accepted recommendations to see whether the model is drifting or the market has turned.

No bank needs an expensive AI conduct-monitoring suite to fix a broken bonus table, because rewriting the compensation contract so variable pay waits for the evidence to mature costs nothing. Before the remuneration committee signs off 2027 sales incentives, human-resources, finance and risk leads should match bonus vesting windows to each product's risk cycle, holding back a set portion of origination bonuses on working-capital and SME loans until cohort net margin, un-restructured repayment and upheld complaints are audited at six and twelve months. At the next committee sitting, ask management to trace a single loan that paid a sales bonus in the first quarter and defaulted or triggered an upheld conduct complaint in the third, and to show whose bonus pool absorbed the clawback. If the sales team keeps the cash while shareholders and customers eat the loss, your bonus scheme is training your software to fail.