If it were easy, everybody would do it. Hard is what makes it great.
One of my favorite quotes is from Tom Hanks in the classic movie, A League of Their Own: “If it were easy, everybody would do it. Hard is what makes it great.”
At boodleAI, we embrace this philosophy. It reminds us that what we do is hard and that achieving our goal is made all the sweeter because we expect (maybe even want) it to be hard.
Our goal has been to create AI technology that enables our customers to achieve their fundraising, sales, and marketing goals. We’ve had hundreds of customers use our AI generated predictive insights to streamline work, improve campaigns, and increase revenue. We take pride in being on the cutting edge of AI, data, and predictive analytics.
Like everyone else, we’ve been fascinated and amazed by how quickly ChatGPT, OpenAI’s chatbot providing consumer access to its large language models (LLMs), has been adopted. However, as amazing as ChatGPT (and its underlying GPT LLMs) are, they only know what they know (although that doesn’t stop them from sometimes confidently asserting things they don’t know as fact).
So, we asked, what if GPT knew everything that boodleAI has learned over the last 5 years?
What if GPT added to its knowledge all of boodleAI’s 35 billion proprietary, proven predictive insights about 250 million adult Americans on everything from interests and affinities to communication preferences to philanthropic giving? What if GPT could add customized predictions based on a specific vertical or an organization’s internal data? What if GPT could resolve identities based on contact information and then use those identities to create insights, and profiles, and audiences? What you would then have is a Predictive AI Assistant.
Thus, the idea for boodleGPT was born.
boodleGPT will be what’s called a “Middle Layer Model”. We’re building boodleGPT by taking a Foundation Model (one of OpenAI’s GPT LLMs) and then optimizing that Foundation Model with boodleAI’s proprietary data and AI technologies to allow it to do something it could not otherwise do: generate responses that predict the behavior of virtually the entire US adult population.
Now comes a huge leap of faith.
I’m using the future tense on purpose. We haven’t finished building boodleGPT … yet. However, rather than build boodleGPT in stealth, we’ve decided to do the opposite and be radically transparent about what we’re trying to do and how we’re trying to do it. We’re doing so for two reasons: First, we know there’s immense interest in the development and use of LLMs like GPT, particularly for specific verticals like philanthropy, and we thought we could contribute meaningfully to that discussion. Second, we wanted to invite the community to provide us feedback and suggestions as we build boodleGPT. (As they say, it takes a village.)
There’s obviously risks in taking this open approach. First, we could fail spectacularly and that fail would be an embarrassing public fail rather than a quiet private one. Second, we may be providing potential competitors a roadmap to duplicate or even surpass our efforts to bring a Middle Layer AI that generates predictions about people’s behavior to market.
We’re willing to take those risks. At boodleAI, we’re no strangers to failing. In fact, a lot of what we do fails at first (that’s part of being a startup), but we have the humility to learn from our fails and the passion to press on despite them so that failing doesn’t turn into failure. We also believe that the fusion of Generative AI like GPT and Predictive AI like boodleAI will create a Middle Layer AI capable of helping so many nonprofits and businesses that we don’t have to be the first, only, or even dominant AI to market to be successful. For those companies building similar Middle Layer AIs or contemplating doing so, we’d welcome a collaboration. The true competition isn’t each other – it’s the problems facing our community that AI can help solve.
We plan to share our progress openly and often in this blog. Future posts will cover a variety of topics, including:
- What is a Large Language Model?
- What is a Middle Layer Model?
- What is boodleAI’s proprietary data and technologies that give us an advantage in building a Middle Layer AI that serves as a Predictive AI Assistant?
- How do we help users make the most of this new Predictive AI Assistant?
- What are the challenges of combining Generative AI and Predictive AI and how can they be overcome?
- How do we address privacy concerns, the handling of PII, and create an AI that is not just effective but also ethical?
So the big question. How can this actually help me?
At this point, you may be thinking, this is intriguing but what can boodleGPT actually do? What problems does it solve that ChatGPT can’t? Why would a nonprofit or business want to use boodleGPT?
Well, instead of telling you, we’d rather just show you.
(Note: the following demo is nonprofit focused, but we could have just as easily created a commercial example)
First, we used ChatGPT to create a demo dataset that is identical to a subset of what boodleAI’s proprietary insights are but does not contain any real data. The dataset includes 9 people who live in either Austin, TX or Washington, DC and the following predictive insights:
- Marital status
- Number of children
- City of residence
- Wealth score
- Estimated home value
- Estimated annual household income
- 5 year giving capacity
- Veteran status
- Affinity for veteran causes score
- Affinity for children’s causes score
- Top 3 interests
- Preferred communication channel
- Donation history to veteran causes
- Donation history to children’s causes
We also added giving history for the 9 people to an imaginary nonprofit called KidsNonprofit.
We then asked ChatGPT Plus to remember the above data, which replicates what GPT would know about these people if we optimized GPT with boodleAI’s proprietary data (but obviously only for the 9 imaginary people and for the 20 categories of insights above rather than all 250 million adult Americans across 120 categories of insights).
This allows us to then ask ChatGPT Plus questions and generate responses as if it were a boodleGPT (these are screenshots of the actual responses generated).
We can ask it to summarize the donation history to our nonprofit:
We then ask the giving potential of donors to our nonprofit:
Then we ask it to analyze the untapped giving potential of donors (an otherwise time-intensive task):
We can ask it to predict ask amounts/gift sizes:
And we can ask what’s the best way to communicate with each donor:
Let’s focus on one donor and have it generate a donor profile based on boodleAI’s predictive insights:
We can even ask it to draft a letter that incorporates the data provided about the donor:
We can ask suggestions about when would be best to send a donation request:
Now, let’s turn from donor cultivation to donor prospecting:
Let’s ask (nicely) for a donor profile:
Then a suggested ask amount:
An AI fundraising assistant should be able to provide recommendations on how to approach a new donor:
Then let’s give specific guidance on a letter we want to send:
The above gives you a glimpse of just some of what boodleGPT will be able to do once completed.
Note: At boodleAI, we’re extremely mindful of privacy and the sensitivities around handling Personally Identifiable Information (PII) that can identify an individual. boodleGPT will implement substantial systems to ensure that PII is handled appropriately and that compliance with privacy laws is absolute. However, our goal isn’t to merely avoid doing what’s wrong, we want to do what’s right. In a future blog post, we’ll discuss Ethical AI and boodleGPT.