Share the full article
In the hubbub surrounding artificial intelligence and its role in social impact initiatives, the discourse usually focuses on bias, and a fear that algorithms and automated workflows will perpetuate, and potentially exacerbate, human bias. This is certainly a risk if the models are trained on biased historical data.
But this singular refrain misses a larger point: for many algorithms to work with a high degree of accuracy they must be trained on a large amount of data that simulates both desired and undesired outcomes. Impact investors and other mission-driven organizations, who have long operated with the intention to minimize undesirable outcomes, may find it harder to build accurate AI models due to scarcer data on adverse outcomes.
Many impact firms are now heading into 2025 with an explicit goal of “weaving AI into” their operations. We at Accessity and Radiant Data hope that our collaboration on a recent AI-supported automation project can illuminate the trade-offs inherent in building machine models for mission-related decisions.
Accessity (formerly Accion San Diego) is a 30-year-old community development financial institution based in San Diego, California. Like other CDFIs across the US, we operate with a federal mandate to extend capital to customers that traditional banks overlook or deem too risky. In late 2023, Accessity received a grant from the Mastercard Impact Fund through the Mastercard Strive USA program to leverage technological solutions such as machine learning and AI to streamline our underwriting process.
Radiant Data serves as a fractional Chief Data Officer for organizations like Accessity that are working to advance economic equity. We deliver rich customer insights to financial services providers and impact investors by bringing together expertise in data science, finance, automation, and machine learning. We believe AI holds great promise for the social sector, but as technologists we must take a creative, non-traditional approach during deployment to uphold impact mandates.
Nonprofits and other impact-first organizations stand to benefit the most from advances in AI. Often, they operate with limited financial and human resources, so introducing automation for repetitive and administrative tasks has great potential to free up staff to focus on critical tasks that humans excel at, like relationship-building, community outreach, mentorship, program deployment, and many others.
Machine learning models in action
Machine learning is a particular sub-discipline of AI that involves the development of statistical models and algorithms that allow computers to make decisions and perform tasks without explicit instructions. For many small to medium organizations, machine learning is a solid first step into the wider world of AI. At Accessity, we view machine learning as a supportive tool to empower our staff with efficiency and enhanced data perspectives, ensuring that AI augments human intervention, rather than replacing it.
There are many different compelling machine learning use cases, but they mostly involve collecting and standardizing a large amount of historical data (where the outcomes have already been documented) and training a model to “learn” which combination and weighting of different factors leads to different outcomes. Once a model is trained to spot patterns, the algorithm can be used to predict future outcomes when fed brand new data. The higher its predictive capability, the more its predictions can be trusted.
If the only goal is achieving high accuracy, however, the act of building a predictive machine learning model may demand data practices that run counter to an organization’s mission.
A solution for equitable capital deployment
A good case in point is the partnership forged between Accessity and Radiant Data earlier this year, with a goal of building a predictive credit risk model to automate underwriting and increase the disbursement of small business loans into underserved communities.
Together, Accessity and Radiant Data built an automated credit model that can predict (with 87% accuracy) the likelihood of a new loan repaying in full. We can understand the significance of 87% accuracy when we reflect on Accessity’s historical performance: as of 2023, our historical gross loss rate was 4.5%.
While Accessity’s low rate of bad debt is a positive indicator of our strong underwriting history, it did present challenges for the model. Because of our overwhelmingly successful repayment history, we had fewer negative outcomes for the model to train on, resulting in some variability in the accuracy for predicting bad debt.
In contrast, we had ample historical data that had resulted in a “paid in full” outcome, so the model quickly learned how to identify good debt. Ironically, had Accessity’s historical data been closer to a 50/50 split between good and bad outcomes, we would have had a higher likelihood of building a model that could predict both with a high degree of accuracy.
Creative solutions for data limitations
How can we overcome this limitation – especially when in the real world, effective and well-run organizations are likely to have a much lower proportion of actual adverse outcomes data? Here are some strategies we recommend:
- Add proxies for bad debt: Accessity had over 1,000 transactions that were still outstanding and several repayment cycles away from having a definitive outcome. However, some showed early signs of slippage: missed repayments, poor financial performance for a few quarters in a row, increasing leverage, and other factors. If the underwriting team felt they were at high risk for default, they were included in the dataset as bad debt proxies. This proxy approach reflects Accessity’s commitment to innovation in leveraging technology while maintaining the integrity of our mission to serve.
- Consider imputing missing values to correct data gaps. For example, if revenue had not been reported for a particular applicant, consider substituting a similar value to the prior period’s revenue, or use an average of the revenue that has been reported. With reasonable imputation, more rows of complete data can be added for the negative cases and used for model training.
- Build multiple models. We suggest building two different models instead of trying to optimize the predictive capability of both outcomes in one model. This is the path we took in the Accessity project. To fully explore the conditions that led to bad debt outcomes, we employed good old fashioned exploratory data analysis on the historical data. Through charting, visualization, and rigorous data querying, we discovered thresholds for six different variables.
- In the instances where all six of these variables exceeded the thresholds, the transaction had more than a 90% likelihood of bad debt. We buttoned up this “bad debt rubric,” and recommended that Accessity use this as a first screen to assess a new incoming loan application. If the new application met all six criteria for bad debt, the transaction would get flagged as such and be passed to an underwriter for review and a final decision. If it passed the first stage, the application would then pass through the main algorithm. This algorithm will then produce a “paid in full” or “bad debt” outcome, and share a percentage likelihood that the prediction is correct.
Figure 1: Workflow of how an Accessity loan applicant moves through the Bad Debt Rubric (gray box) and the Predictive Model (orange box). The model was built as an ensemble of different decision trees, and hence, the final output score is accompanied by a voting tally of % of trees that voted in favor or against the loan. These votes provide context to an underwriter as they make a final decision.
Once operationalized, these models have the potential to cut Accessity’s underwriting time in half. The Credit team is clear that the models are not set up to auto-approve or auto-decline. Every transaction will still be approved or declined by a human. But in the cases a transaction is flagged as bad debt or classified as paid in full with a high degree of certainty, the underwriter can do a quick review and validate the decision.
Accessity’s goal is to increase efficiency, allowing the team more time to focus on complex cases that require deeper analysis and thoughtful decision-making. This blend of AI-driven insights and human intervention ensures that we maintain our mission-driven approach to lending, while leveraging technology for smarter, faster outcomes.
Optimizing machine learning outcomes
This collaboration between Accessity and Radiant Data culminated in a successful, high impact model for a few additional reasons: a) the size of Accessity’s historical dataset, and b) the guardrails we instituted to manage bias and “model drift.”
We started the project by analyzing Accessity’s historical transactions that surpassed 11,000 transactions. After cleaning, whittling down to the most important fields, and filling in for missing values, we were left with 3,000+ final transactions to build the model with.
We were able to work with this final number, but a higher number of transactions is preferred, especially when there is a 19% to 81% imbalance between the two outcome classes. This is an important consideration when mission-driven organizations take on automation or algorithm-building. Many times, it’s a no-go simply because of limited historical data and too few transactions (ex: less than 1,000 transactions).
Therefore we always recommend that mission-driven organizations start off with a low-stakes data suitability assessment before committing time and resources to a longer-term build + deploy engagement.
Managing bias
Lastly, to manage concerns around bias, we followed a three-pronged strategy we at Radiant Data like to refer to as “past-present-future guardrails.”
- Past, or historical analysis. Examine both paid in full and bad debt outcomes and observe any instances of human bias. As a hypothetical example, pretend we observed historically that women-identifying customers were disproportionately denied loans. In Accessity’s case, we did not observe historical human bias, but if we had, we would have moved to the next step.
- Present. Create different models for each group that has been affected by bias. Using the prior example, this means that we would gather all instances of women applicants and build a model exclusively for them. Within the model, women applicants are being assessed only against other women so in a way we are controlling for bias. We would construct another model for the other gender groups, and keep that evaluation separate.
- Future. Build thresholds that are closely monitored after the model is put into production. For every model that is built, it is imperative to build parameters to monitor model drift and calculate imbalances in any bias features as the model ingests new data. For example, if we see over time that the predictive model is recommending higher decline rates for a particular underserved group, we will receive instant notifications that an acceptable threshold has been crossed, and be able to quickly tune the models and course correct. Models should never be unsupervised; close monitoring and frequent tuning will ensure they perform as intended.
What’s next for machine learning and social impact
On the surface, it can often feel that the demands of data-hungry machine learning models are at odds with impact mandates, and left unchecked, are also at risk of perpetuating patterns of historical bias. But it would be a shame if these constraints made mission-driven organizations reluctant to adopt AI-powered solutions. By acknowledging machine learning’s constraints, and adopting creative workarounds, it is entirely possible to build automated solutions that support an organization’s mission and simultaneously invite greater efficiencies.
Together, Accessity and Radiant Data have successfully built algorithms that will enhance the decision-making process, allowing Accessity to deploy capital faster and more efficiently to underserved communities, while remaining true to its mission. As organizations like Accessity embrace AI and machine learning, their core principles remain unchanged: technology serves to enhance the capabilities of employees, not to replace them. This ensures that we continue to prioritize thoughtful, human-centered lending decisions, while improving operational efficiency and impact.
Mar Diteos Rendon is Chief Operating Officer at Accessity. Nicole Jansma is the Director of Program Efficiency, Data & Impact at Accessity. Sachi Shenoy is Co-founder and CEO of Radiant Data. The authors would like to thank the Mastercard Impact Fund and the Mastercard Strive USA program for supporting this important initiative.
