Skip to Content

The Causal AI conference is back in San Francisco for 2024, bigger and better than ever.

Register Interest

AI that SME lenders can bank on

At a glance

  • SME lenders need to evaluate many types of risk efficiently and diligently. 
  • Current incarnations of machine learning heighten risk. These systems:
    • Misperform under normal day-to-day volatility and completely crash during crises.
    • Overfit to noise and are misled by spurious correlations.
    • Lack transparency.
  • Causal AI is the tech solution SME lenders need because it:
    • Makes more accurate predictions, by honing in on causal relationships.
    • Is more adaptable, by modelling deep causal relationships.
    • Is more explainable, giving insight into the true drivers of risk.
    • Facilitates human-machine partnership with underwriters. 
    • Syncs with and optimises loan portfolio strategy.

Small and medium-sized enterprises (SMEs) account for 99% of all businesses. They power economies and, in doing so, generate approximately $850 billion of revenue for banks each year. 

Yet, relative to big corporates and consumers, SMEs are underserved — the global SME finance gap is $5.2 trillion. There are vast untapped business opportunities for any lenders that are up to the challenge of serving SMEs at scale.

Challenges for business lenders

Evaluating and managing risk is at the core of the SME lending value chain. To compete, lenders must be able to evaluate risk accurately, adaptably, explainably, efficiently and with an eye to overall portfolio strategy. 


Accurately predicting loan default risk translates directly into profits, while inaccurate credit decisioning equals loan losses or missed revenue. Thin-file SMEs can be particularly challenging to underwrite accurately.


The world with its changing demographics, consumer behaviour, policies, climate patterns, financial systems — is constantly in flux. Risk models must be able to quickly adapt to new conditions, whether due to major crises or ordinary day-to-day volatility. 


Internal and external auditors, compliance teams and regulators require lending decisions to be transparent, trustworthy and fair. And SME owners expect loan decisions to be explainable. 


Digital lending has revolutionised the customer experience in commercial lending. Time-to-apply, time-to-decision and time-to-cash have been compressed from weeks to minutes. Today’s lenders must be able to keep up. 

Loan portfolio strategy

There has been a recent spike in non-performing commercial loans, described by the World Bank’s Chief Economist as a “quiet financial crisis”. Stimulus packages and loan moratoria may not provide a permanent solution to SMEs’ troubles. Against this backdrop, it’s more critical than ever that lenders manage concentration risk and protect their loan portfolio’s overall health. 

Current incarnation of machine learning is a red flag

Business lenders have recognised that human-only credit decisioning is too inefficient, and have doubled down on investment in AI. But the overwhelming majority aren’t seeing meaningful returns on investment. This is largely due to the limitations of conventional machine learning (ML).

“ML methods can be very inaccurate when used to predict loan performance in out-of-time samples” — this was the finding of an exhaustive study of millions of loan decisions recommended by ML algorithms. The study found that small perturbations in training data and subtle shifts in the macroeconomic environment lead to wild swings in underestimation and overestimation of default risk. 

The problem is that conventional ML systems overfit to noise in complex, commercial environments. When conditions change slightly — as they do all of the time — ML algorithms break down. The problem is exacerbated in crises.  A Bank of England survey found that more than one in three UK banks’ ML models completely failed as a result of data shift during the pandemic; more broadly, banking has been among the industries whose advanced analytics models have been most impacted by COVID-19

Conventional ML also suffers from a lack of transparency, which makes this technology poorly suited to heavily regulated, high-stakes sectors like banking. The more powerful ML algorithms, like deep learning, are unintelligible — audit, compliance and regulators can’t scrutinise them and lenders can’t explain decisions to SME owners. On the other hand, more transparent models, like linear regression, are typically too restrictive to be useful (see Figure 1). 

Figure 1. There is a trade-off between model explainability and accuracy among conventional ML algorithms. Causal AI overcomes this trade-off. 

Causal AI for loan decisioning

Causal AI — a next generation AI technology that can reason about cause and effect — rises up to the challenges of commercial lending, where conventional ML falls short. Picture a lender deploying Causal AI to make SME loan decisions.  

Causal AI zeros in on the true causal signals of loan default risk, ignoring spurious correlations and misleading noise. Predictions are both more accurate and more adaptive to changing circumstances. We found that, during a crisis, a decisioning model built with Causal AI delivered 38% higher annual PnL than state-of-the-art ML systems (see Figure 2). 

Figure 2. Data from the US Small Business Administration. Causal AI outperformed both conventional ML algorithms and US banks’ actual loan decisions during the global financial crisis, in a period when default rates were skyrocketing. 

In addition to making better predictions, Causal AI is highly explainable. The AI discovers a causal model that represents the key drivers of default risk (see Figure 3). This model provides an intuitive summary of what’s driving loan defaults — including the directionality and strengths of those drivers — which can be audited, regulated and explained to applicants.

Figure 3. A toy causal model, illustrating the causes driving loan default risk. 

An auditor can query exactly which business characteristics are responsible for a rejected loan, and they can anticipate how loan default risk would change if, say, inflation rates increase or the US dollar weakens. Counterfactual reasoning — reasoning about alternative possible scenarios — is the mechanism that allows for these explanations. Counterfactual reasoning is a signature feature of Causal AI that conventional technologies simply can’t replicate reliably — indeed there are too many examples of businesses that get into trouble when they assume that it can.

Causal models provide a highly flexible and interactive user interface via which underwriters can share their domain expertise with the AI. If an underwriter has an insight that, say, board composition is predictive of default rates, then they can convey this to the AI. Deep human-machine interaction is a hallmark of companies that see the biggest financial benefits from AI.

Credit decisioning doesn’t take place in a vacuum — it’s critical that new loans contribute to the overall health of the lender’s portfolio. Causal AI implements cutting-edge loan portfolio optimisation strategies, that achieve diversification based on . The AI then autonomously updates new loan decisions to optimise for overall loan portfolio performance. For instance, if the portfolio optimisation strategy indicates that the lender is overexposed to a given sector, the decisioning model autonomously places a higher threshold on lending to SMEs operating in that sector. 

Wider impact

Causal AI meets the key challenges that SME lenders face in today’s competitive environment. 

Decisioning models built with Causal AI understand the true drivers of loan default risk. They are robust under stress,  harvesting 38% greater PnL in periods of market volatility as compared with conventional analytics. Causal models are auditable, explainable to key stakeholders and they empower underwriters to share their knowledge with AI systems.

Lenders who adopt Causal AI will give momentum to a virtuous cycle. Greater efficiency gives lenders more pricing power. Greater decisioning accuracy enables banks to take on new business, especially while their competitors have lost their risk appetite. In turn, this creates bigger and better proprietary data from which to accelerate learning and amplify advantages. Analysts estimate that a bank with $250 billion on its balance sheet could achieve $250 million of annual PnL gain per year through AI-enabled digital lending. Causal AI promises to be a disruptive enabler of these gains.