Skip to main content

Credit decisions made by models are still your decisions

Lenders and financial services firms were early to automate decisions about customers. Regulators, boards and customers now ask the same question: who answers for the outcome, and can you show it?

AIC.Karoo Lending (Demo)

Decision log: personal loans

Sample data

  • 09:14Held for review
  • 09:21Reviewed
  • 09:30No review needed
  • 09:42Held for review

For

  • Credit risk and model risk teams
  • Compliance and conduct
  • Fraud and collections
  • Boards and risk committees

Where AI decides, and what goes wrong

  • Decides about: Applicants

    Approving or declining credit

    People with thin credit files declined on proxies the model learned, with no reason anyone can give.

  • Decides about: Customers

    Setting limits and prices

    Similar customers offered different terms for reasons nobody can trace.

  • Decides about: Customers

    Flagging transactions or customers as fraud

    Accounts frozen at the worst moment, with no person looking until the customer complains.

  • Decides about: Debtors

    Prioritising accounts for collection

    Customers in hardship pushed hardest because they are most likely to respond.

A declined customer asks for the reason

The National Credit Act gives a consumer the right to ask. The answer has to come from somewhere.

  1. 1

    The request arrives

    A customer declined online asks for the dominant reason their application was refused, as the National Credit Act allows.

  2. 2

    The reason is on the log

    The decision log holds the reason the model gave when it decided, not one reconstructed weeks later.

  3. 3

    A person reviews it

    The analyst who answers for the model reviews the decline. If it is overturned, the override and its reason are recorded.

  4. 4

    The board can see the pattern

    Override rates and open model risks reach the board in a sealed pack, so oversight is evidence rather than assurance.

Rules that already apply

A summary, not legal advice about your organisation.

POPIA section 71

A customer declined by an automated decision that profiles them must have safeguards, including the chance to make representations. Read the guide

What it asks of youHuman intervention and a way for the customer to be heard, with enough about the logic to make it real.

National Credit Act, section 62

A consumer refused credit may ask the credit provider for the dominant reason for the refusal.

What it asks of youA reason you can state for each decline, which an unexplained model cannot give.

Joint Standard 1 of 2023

The Prudential Authority and FSCA standard on IT governance and risk management for financial institutions.

What it asks of youGovernance and risk management over the technology, models included, that the business depends on.

EU AI Act

Lists AI used to evaluate creditworthiness or set a credit score for individuals as high-risk, from 2 December 2027, where its output is used in the EU. Read the guide

What it asks of youHuman oversight, logging, data governance and a risk management system.

Questions

We already run model risk management. What does AIC add?

Model risk asks whether a model is sound. AIC asks whether a named person answers for each decision it makes about a customer, and keeps the evidence that they did. The two meet on the risk register, where model risks are tied to the system and its owner.

Does every declined application need a person to review it?

Not necessarily. The law looks for safeguards such as the ability to intervene and to make representations. Many lenders hold a sample, or every decline near a threshold, for review. Which design fits is your decision; the log shows what you actually do.

Can we show our regulator or board without giving them access?

Yes. An evidence pack is a fingerprinted copy of the record for a period. Anyone holding it can check it matches the record AIC holds, without an account.