The Expert View: AI and credit – smarter decisions, faster

What’s holding back AI in private credit?

 

On 10 June, Business Reporter brought together some of the credit world’s brightest minds to discuss the growing role AI tools are playing in credit decisions.

 

The title of the session captured what every lender wants: smarter decisions, faster. But it soon became clear that the reality of AI tools in credit is more ambiguous, with promising deployments often held back by siloed data systems, a lack of observability and an institutional hesitancy when it comes to new kinds of software product.

 

In broad strokes, many parts of the credit analysis process could potentially be safely handed off to contemporary AI systems. Credit analysis often involves poring through thousands of pages of documents; if an AI model summarised those documents instead, extracting key figures and highlighting notable sections, it could save hundreds of hours that could then be spent on more nuanced work. Beyond data analysis, handing the simplest tier of credit decisions over agentic AI could make the private credit system far more liquid and responsive, with far less human overhead.

 

But those ideas are still far from the reality of how credit decisions are made. Most of the attendees recognised the promise of AI and said they felt pressure from their own organisations to use the technology, but the promised efficiency gains had yet to materialise. So far, AI tools are mostly used for aiding in research rather than initiating decisions, placing a hard limit on how much faster their credit process could run. 

 

Even for analytical tasks, concerns about mistakes or hallucinations have made credit desks wary of turning too much of their process over to a large language model. As one attendee put it, when a five-million-dollar loan request comes through, how much can you trust the AI to tell you everything you need to know? The prospect of an error slipping through is simply too frightening, and AI systems haven’t been used for long enough to gain users’ trust. In the end, the attendee said, he doesn’t feel comfortable until he double-checks the AI’s output, at which point not much time has actually been saved.

 

None of these should be seen as insurmountable obstacles. Instead, the current reservations indicate the specific features and assurances credit desks will need before they can integrate AI into their workflows. There are a number of tech startups already exploring more reliable AI techniques, often implementing deterministic fact-checking systems on top of the more flexible non-deterministic LLMs. 

 

Perhaps the biggest obstacle is bringing all an organisation’s data and expertise into one place, where it can be used as context for an AI model. By now, financial IT professionals have been hearing for 10 years or more that they need to transform legacy systems into a more modern cloud-based data warehouse.

 

In some institutions, this transformation is already underway, but it’s a slow and often politically sensitive process. The promise of AI could break that deadlock, driving home the benefits waiting on the other side of the AI transition – but it could just as easily fall prey to the same inertia and political infighting that has kept it from happening up until now.

 

Credit teams are also eager for observability tools, another area where AI researchers are hard at work. For regulatory purposes, it’s often necessary to explain why a given credit decision was made; if an AI model played even a small part in that decision, it can make that task far more complex. Those tools are also necessary to monitor agentic systems. Otherwise, an AI model could stray far beyond its authorised task without in-house technologists having any idea. 

 

Unfortunately, there are no straightforward tools to give banks the observability they need. AI researchers are making real progress on the observability issue, both in public research and private-sector products, but it remains an unsolved problem for the field.

 

 So far, financial institutions are mostly responding to the challenge of AI by forming committees. But among the attendees (some of whom were serving on those committees), there was real frustration about their inability to clear the very real obstacles to AI deployment. Instead, committees often served to simply reaffirm that those obstacles were still there.

 

“We know in the committee that we don’t really understand AI,” one member said, “but we need to go by the rules that the regulators expect.”

---

To learn more, please visit: www.spglobal.com