It is fashionable to talk about “AI underwriting” and “instant approvals”. The reality at Credicorp is more mixed and, we think, better. Routine applications are decided automatically: it is faster for the borrower, it is consistent, and the rule is transparent and challengeable. But the routine cases are not where the responsible-lending question gets hard. The question gets hard at the borderline — the trading history that has one anomalous month, the affordability metric that lands just under the threshold, the recent bank statements that show a recovery from a difficult quarter. This article explains how Credicorp handles those borderline applications.
The three lanes
Every application flows through a triage with three outcomes:
- Auto-approve. The trading history, the affordability indicators, the business credit reference data, and the bank-statement signals all sit comfortably above our thresholds. The decision is automatic; the disbursement is same-day; the rule used is recorded on the application record.
- Auto-decline. One or more hard exclusion criteria are present (current insolvency proceeding, sanctions hit, fraud flag, fundamentally insufficient affordability). The decision is automatic; the applicant receives a plain-English explanation naming the reason; a request for a human review is offered.
- Manual review. The borderline lane. Anything where the indicators are mixed, recent, or genuinely ambiguous routes to a credit analyst. This is the lane that gets the time and the attention.
What the manual review covers
The analyst pulls together three sources of evidence: the application detail (what the loan is for, the term, the requested amount); the business credit reference data from Experian Business, Creditsafe and Equifax Business; and the open-banking-pulled bank statements (with the applicant’s consent) covering at least 90 days. They are looking for the things the automated scorecard cannot see — a one-off bad month that has a clear and benign explanation, a recovery trajectory after a difficult quarter, a customer concentration that explains an apparent revenue dip, a major contract about to land. The analyst writes a short structured rationale that is retained on the application file regardless of the outcome.
Vulnerability awareness in the borderline
If the applicant carries a vulnerability flag, the manual review takes that into account explicitly — not as a softer affordability standard (the £ test is the £ test), but as a flag for: extra time on any follow-up questions; the named handler routing; a written summary of the decision in plainer language; and an offer of an explanatory phone call. The decision is still based on whether the borrowing is genuinely affordable for the business — but the borrower’s circumstances inform how that decision is communicated.
Why the explanation matters
An adverse decision at Credicorp is never delivered as just a number. The plain-English explanation names the specific factor (or factors) that drove the decline, gives an indication of what would need to change for a future application, and offers a route to challenge — including the right to request a human review of an automated decision under UK GDPR Article 22, which we honour proactively on every automated decline regardless of whether the applicant invokes it. The explanation goes in the channel the applicant requested, and a follow-up call is available if the applicant prefers to ask questions verbally.
Auditing the borderline
The credit analysts are not a one-person team and their decisions are not final-word. A sample of borderline decisions is peer-reviewed monthly, looking for consistency — same factors, same outcome, regardless of which analyst made the call. The peer-review findings feed back into the scorecard tuning and, occasionally, into a documented scorecard change. The change history is published in the credit-policy appendix so that an applicant who reviews their declined application can see the rule that applied at the time.
The trade-off, named honestly
Manual review takes hours instead of minutes. We could automate the borderline and approve faster. We choose not to because the cost of an algorithmic borderline error — declining a fundamentally creditworthy small business because of an anomalous month, or approving an applicant who cannot afford the repayment — is high in both directions. A human reading the bank statements catches both. The application is slower for the borderline applicants; we think the trade-off is the right one. The How we lend page has the full decisioning policy, and the vulnerability-in-business-finance article covers the vulnerability framework that wraps it.