Automate the Process Before You Automate the Decision

Automated credit decisioning should enhance decisioning but shouldn’t be the only part in the process. For example, a business can have healthy cash flow, growing revenues and a viable future, and still struggle to access funding because there simply isn’t enough conventional information available for a lender to assess it confidently. That distinction matters. A lack of information is not the same thing as a bad credit risk. Yet for many smaller or younger businesses, the outcome is ultimately the same.
Lenders face a practical challenge of their own. From client feedback, a commercial credit application can absorb between 2 and 4 hours of active manual capture and calculation before an underwriter has even formed a view on the risk, and overall turnaround can stretch across several days. When the amount being requested is relatively small, the economics become difficult, because the work required to assess an application is broadly similar whether a business wants to borrow R100,000 or R10 million.
This is where much of the conversation about AI and automated credit decisioning starts in the wrong place. We are rushing to automate the decision when, in many cases, the bigger opportunity is to improve everything that must happen before that decision can responsibly be made.
Underwriters spend a significant amount of time gathering and working through information before they can apply their expertise to the credit decision itself. The delay a customer feels is rarely caused by the decision. It is caused by everything that has to happen before the decision is possible, which is why automating the decision on its own changes so little.
There is a temptation to put AI into every part of that process simply because the technology is available. That is not where its value lies. Some elements of credit assessment need to remain consistent, explainable and auditable. AI has a valuable role in helping process information, identify patterns and surface insights more efficiently, but faster does not automatically mean better. Speed without explainability can create more risk than it removes.
A system must know when it doesn’t know
As automation becomes more sophisticated, one of the most important capabilities is knowing when human intervention is required. A system can process enormous amounts of information and apply rules consistently, but credit decisions do not always fit neatly into standard scenarios. Information can be incomplete, circumstances can be unusual, and the numbers do not always tell the whole story. A system can produce an answer that looks entirely reasonable and still be wrong, with nothing in the process to flag it. I would rather see technology raise something for review than confidently produce an answer where the evidence is uncertain.
The same principle applies to the debate about whether AI will replace credit underwriters. I don’t believe it will, nor should that be the objective. Underwriters understand context. They question inconsistencies and recognise when the information in front of them needs further interpretation. Technology is good at processing information, applying policy consistently, identifying patterns and surfacing what deserves attention. The strongest credit model is not human versus machine. It is human judgment supported by better, faster and more consistent intelligence.
Consider a growing business with limited credit history but strong recent cash flow. A conventional assessment may struggle with the limited history, while an underwriter able to consider a broader view of the business may reach a different conclusion. The question is therefore not whether the human or the machine should decide. It is where human judgment adds value. Routine cases can be processed more efficiently, while exceptions and borderline cases require an appropriate review path. Every model has blind spots, which is why consequential credit decisions should retain a mechanism for human intervention.
The bigger opportunity is access to finance
This matters particularly for SMEs. Newer businesses frequently have less formal credit history and their financial information may be less structured, which makes them more difficult and more expensive to assess.
The consequence is bigger than a slow application process. A viable SME that waits too long for funding may lose the opportunity it needed the funding for. Stock is not purchased. Equipment is not acquired. A contract cannot be fulfilled. Expansion is postponed. At the other end of the spectrum, poor assessment creates a different risk, where capital becomes tied up in poorly performing debt, impairments increase and lenders have less capacity to fund other businesses.
Better credit assessment therefore has to achieve two things. It needs to help lenders make good decisions faster, while giving them a more complete view on which to base those decisions. If the time and effort involved in assessing an application can be reduced, smaller loans become more economical to consider. If relevant financial and credit information can be viewed more holistically, a thin credit file does not necessarily have to be the end of the conversation.
This is not about finding ways to approve businesses that cannot afford credit. It is about making it easier to distinguish between high risk and not knowing enough yet. Those are not the same thing.
Technology cannot remove risk from lending. Businesses fail, economic conditions change, and strong performance today does not guarantee repayment tomorrow. What technology can do is reduce unnecessary friction, improve the information available to decisionmakers and create greater consistency in how risk is assessed. That may ultimately be the more meaningful opportunity for AI in credit. Not replacing the person making the decision but giving them better information on which to make it.
By Sibusiso Mtiyane – Lead Data Scientist at Mettus