LND // For Lenders · AI credit decisioning

A credit brain your risk team owns and your regulator accepts.

Credit decisioning is the single most defensible AI investment a lender makes. We build firm-specific scorecards that combine open banking, bureau, alt-data and application signals — with explainable adverse-action notices, live A/B champion-challenger, and full audit trail for the FCA and your risk committee.

BuyerCRO · Head of Credit · Head of Data Science← All lender capabilities
The problem

Why generic scorecards are a compounding disadvantage.

Most lenders under £500m book still run one of three shapes: a bureau-driven scorecard the vendor tunes once a year; a Provenir/Zoral rulebook that the risk team has been meaning to overhaul for two years; or a hand-crafted set of policy rules stitched together in the LMS. All three under-price good risk and over-price bad risk relative to what an open-banking + bureau + alt-data model tuned on the firm's own book can achieve.

The compound cost is enormous. Every basis point of adverse selection costs the loan book. Every falsely-declined good borrower is a customer who will now originate somewhere else. Every borderline case that lands in manual review costs an underwriter's time and slows the funnel. The lenders that have built a firm-specific credit brain in the last three years are extending an increasingly uncatchable per-loan-economics gap.

What good looks like

What a modern credit decisioning system looks like.

A firm-owned brain, not a vendor black box. It should produce:

  • A single score per application, combining bureau, open-banking cashflow patterns, application signals, device/graph signals and (where the product allows) alt-data like HMRC, VAT filings, e-commerce revenue or trade-payment data.
  • An explainable adverse-action notice for every decline — regulator-defensible, borrower-comprehensible, in production not on a Confluence page.
  • A champion-challenger harness the risk team runs live — new models shadow-scored against production, with lift and stability metrics tracked per cohort per product per channel.
  • A monotonic-and-bounded architecture so the risk committee, the auditor and the FCA can all follow the logic.
  • Full retention of model version, training data version and decision inputs per application — for regulator inspection years later.
What KJ Capital ships

What KJ Capital ships.

A bespoke credit decisioning service that sits above your LMS and your bureau/open-banking vendors. We deliver:

  • A production scorecard trained on your own historical book, evaluated against a hold-out cohort with published Gini, KS, PSI and calibration metrics.
  • An adverse-action notice generator with regulator-defensible explanations, versioned and audit-logged.
  • A champion-challenger platform your risk team owns and runs.
  • A model registry, evaluation harness and drift monitor.
  • The full source, deployed in your cloud tenancy. No black box, no lock-in.
Typical outcomes

What lenders actually see after go-live.

Auto-decision rate on clean applications
70–90%
Adverse-selection reduction
8–20%
Manual review load
-40 to -70%
Time from kickoff to production
10–14 weeks
Frequently asked

What buyers ask us about this build.

Q01Do you replace our Provenir / Zoral / LendingMetrics rulebook?+
You choose. Most engagements sit the AI scorecard alongside the existing rulebook and let the risk team migrate gradually. Some rip and replace where the rulebook is beyond salvage.
Q02How is this FCA / Consumer Duty compliant?+
Explainable outputs per decision, monotonic architecture where required, adverse-action notices generated in production, model registry and drift monitoring, and full retention of decision inputs. This is stronger than the paper-based rulebook status quo, not weaker.
Q03What alt-data sources do you support?+
Open banking (TrueLayer, Plaid, Tink, Yapily), HMRC filings, VAT records, Companies House, e-commerce revenue APIs, trade-payment data (Xero, QuickBooks), and product-specific sources like automotive DVLA data or property Rightmove/Zoopla data.
Q04How do you handle model risk management?+
Every model shipped comes with SR 11-7 / PRA SS3/18-shaped documentation, champion-challenger harness, drift monitoring and re-training cadence agreed with the risk committee.
Q05Price?+
Productised Build: from £150k, 10–14 weeks. Continues on the AI Operator retainer once live.
Next step

Two weeks. £15k. A written blueprint and a working proof-of-concept on your loan book.