Five Evolutions of Using Machine Learning to Quantify Connection and Predict Generosity

Having spent two decades leading nonprofit organizations, I found myself in an all too familiar paradigm that sits between an unquenchable thirst for more philanthropic dollars—from a shrinking pool of donors .

For (too) many years facing this paradigm, I adopted the ideology that to overcome a continual shrinking pool of donors, we simply needed to spend more money to find new donors at any cost—essentially trading the opportunity of building a relationship for a transaction, any transaction. Whether it was through hiring more gift officers or spending more money on multi-channel acquisition, the result of this short-term thinking over time has manifest severe negative consequences upon the nonprofit sector as a whole. The long-term implications of those actions is now brightly evidenced by?annual decreases in the number of households that give to any charity.?While most developed nations are facing a sort of?Generosity Crisis , there is significant reason for hope.

When faced with a difficult circumstance, my mom used to remind me that I could be part of the problem or part of the solution. It was out of love for the philanthropic sector that I decided to flip the giving pyramid by leveraging technology to prioritize relationships over revenue.?

In 2017 after 20 years of fundraising with the idea that generosity came as a result of a person’s discretionary wealth, I adopted a new golden rule that “generosity is the manifestation of connection." From that moment,?I set on a path to leverage advanced technology to quantify connection. After several years of learning about the promise and limitations of machine learning, a hypothesis proved itself by showing that we could leverage AI to predict generosity by quantifying connection.

Over the past five years, I've had the opportunity to drive machine-learning and deep-learning programs for many of the nation’s best nonprofit organizations.?While the learning curve was steep, and we continue to evolve, my partner in all things AI,? Scott Rosenkrans ?and I have experienced many AI revelations, evolutions and epiphanies over the past five years. What we’ve learned during that time has been nothing short of remarkable.

After launching a new AI effort at?DonorSearch?in March 2020, our data science team now calculats more than 35 million predictions a month on $6.5 billion in gift transactions, each using roughly 1,000 datapoints per prediction.?

As the nonprofit sector is about to "cross the chasm" in its wholesale adoption of AI, I felt compelled to share with the world five evolutions of AI that we've experienced while completing over 5.5 trillion calculations. Ultimately, as guardians of the nonprofit sector, I hope these evolutions help your journey.

?___________________

First Evolution: Thou must enrich data - more data IS better

When we started our work in AI, we thought we could simply use first-party client data to make meaningful predictions. We learned quickly that our hypothesis was wholly incomplete and that we needed to find and enrich client data with as much granular external data as possible. The result of this more holistic approach had two benefits.?

First, it helped remove extreme bias of pigeon-holing constituents into 10, 20 or 30 datapoints. People are complicated, and the mere notion that you can accurately predict human behavior using just 20-30 datapoints is insulting from a modern technological perspective whereas using 1,000-plus data points provides a more accurate, more equitable and less biased view of a person. Second, by using a variety of machine-learning algorithms that prioritize experiential data, we were able to boost accuracy of donor predictions by five times over using wealth data alone.

More data is better when in experienced hands but all data is not created equal. Anyone who tells you otherwise does not understand the full power that deep learning holds.

?___________________

Second Evolution:?People are dynamic - the operative word of machine-learning is "learning".

Early in our work in AI/ML, we realized that building a machine learning model that remained static would eventually be no better than some of the antiquated scoring models or basic logistic regression models the nonprofit industry has relied on for the past few decades.?We quickly began to understand that from the moment a model is output, that it will start to degrade based on the new data that would be generated by the constituent.?If our model didn’t evolve over time, the “learning” part of machine-learning would be in name-only.?

The correlation here is that if the Amazon model didn’t “learn” from your behaviors, preferences, shares and purchases over time it would be providing recommendations based on that first book you bought on that platform in 1997. Operationalizing machine-learning models and retraining them continually on new (and more) data over time quickly became our second evolution.

?___________________

Third Evolution: Donors and prospects are NOT the same

For the first few years of predicting donor behavior, our hypothesis was that donors and prospects could be scored via a single predictive model. The unpleasant result of which is that prospects (non-donors) would inevitably score very low compared to their donor counterparts.?This happens every time, 100% of the time.?

For too many years,?Scott Rosenkrans?and I found ourselves in a vicious cycle by trying to put lots of “lipstick on the pig” by normalizing prospects scores and?trying?to convince clients that the “scores weren’t bad, just not as good” because the prospects hadn’t yet made a gift.?This fable is all too familiar to our clients that have built or bought standard predictive models throughout the years, yet always leaves for disappointment by fundraisers and leaders alike.

As it turns out, after interrogating 33,000 machine-learning models, we learned that not only are the extraordinarily low prospect scores unappealing to clients, but worse was that the data science was wrong. We now recognize that donor and prospect behaviors and motivations are quite different and that the mix and rank-order of features within each model are completely different from one another. For example, first time donors to healthcare organizations typically give based on their recent clinical encounters whereas repeat donors to the same organization give more based on their continued relationship with members of the foundation.

When we built our first donor and prospect models independently and viewed them side by side, it was one of the most glaring revelations of my career. We now build fully independent donor and prospect models which allows both groups to score based on their own unique characteristics. To the benefit of the entire nonprofit sector, the significance of this evolution can not be understated enough.?

?___________________

Fourth Evolution:?It's not about more donors - it's about better donors

Using our $6 billion in recent giving transactions, our hypothesis was whether using machine-learning with 1,000+ non-wealth-oriented data points would simply predict donors better, or instead would it be able to predict better donors better??As it turns out, leveraging AI with large quantities of experiential data shed a spotlight that not all donors are created equal.?While average one-time donors have a lifetime value (LTV) of $283, the average LTV of a retained donor is $4,685, a value of 16X greater when looking at LTV.?

By computing and quantifying connection instead of wealth, donors within the top 2 quartiles (scores of 50 and above) of our model scores retain at 70% with the top quartile retaining at 81% compared to an average retention rate of 19-21% for first-time donors.??The significance of this differential from an ROI perspective blew our minds, and should provide a pretty big "ah-ha" moment for you as well.

?___________________

Fifth Evolution:?Trust is the currency of the nonprofit sector - There is no place for black boxes

While our first four evolutions presented themselves by continually questioning, testing and pushing boundaries of what was possible, a grounding principal that became a constant theme was around the idea of responsible AI within the nonprofit sector.?While proprietary algorithms in the private sector are extremely common (and expected), the idea of a “black box” has no place in the nonprofit sector.?Our AI team members have become outspoken advocates on the role of responsible AI within the nonprofit sector, with a strong emphasis on key issues like donor privacy, security, data ethics and data equality.??The risk is that bad players may openly or inadvertently deploy AI that could do long-term harm in the form of weakening trust in the non-profit sector.?

?

I hope that in shedding light on these five evolutions of responsible AI in the nonprofit sector it will accelerate your learning, spark a desire to learn more and to understand that advanced technologies like AI hold tremendous power for good when used responsibly.?

If you would like more information or would like to chat more about any of these topics, please send me a DM. If you would like to join a community of professionals dedicated to the responsible use of AI in fundraising, please join our free group here at?Fundraising.Ai