Artificial Intelligence, Predictive Intelligence, AI, Machine Learning. Do nonprofits really need to be thinking about all of this? Short answer is: yes!
- What we mean when we talk about Artificial Intelligence
- What AI means for nonprofits
- What steps you can take to prepare
This webinar was hosted and recorded by Salesforce.org on 6/13/18.
Transcript
Andrea Schiller (AS):
Good morning everyone, and welcome to the Introduction to Artificial Intelligence for Nonprofits Webinar. Before we get started, a couple of logistics. As we go through the webinar, please feel free to ask questions in the chat window. We’ll be answering questions throughout the webinar. But any that we don’t get to, we will be answering at the end. Also, please be aware that we’ll be recording the call and sending it out to everyone in the coming days, in case anyone is interested in rewatching or sharing. So with that, let’s jump in. Again, welcome and thank you for joining us. But before we jump into the content, please be aware of our forward-looking statement.
If you can’t read them here, you can find them on our website, but be aware that we may be making some forward-looking statements. So please make sure that all of your purchasing decisions are based on currently available technology. Okay. So we have a great lineup of speakers for you today. My name is Andrea Schiller. I do product marketing here at Salesforce.org. We also have Lisa Rau, the Executive Chairman of Fíonta, to talk to us about artificial intelligence. She will talk about herself and give a little more background. But she is an expert in artificial intelligence, so super excited to have her with us today.
Agenda
We have a ton of great content prepared for you. We’re going to start with a very brief introduction to Salesforce.org. Then I’ll pass it over to Lisa to talk about artificial intelligence, what it means for nonprofits, and how you can prepare.
Our Path to Impact
So, for those of you joining us for the first time, I wanted to give a brief overview of Salesforce.org. Salesforce.org is a philanthropic arm of Salesforce.com, and we have three pillars. The first is technology. We build a purpose-built product for nonprofit, higher-ed institutions in K-12 schools.
We also focus on investment. So all of the profit that we make through selling those technologies, we reinvest in the community to ensure that we have a future-ready workforce. And third, we partner with the community to ensure that we’re building products that nonprofits are using and that we are positively working together to influence our community overall. We’ll be focusing on the technology pillar of Salesforce.org during this presentation.
We All Face Similar Challenges
We initially wanted to build purpose-built mission for nonprofits, because we know that nonprofits all face similar challenges.
40% of organizations don’t collaborate efficiently, internally, 47% of nonprofits can’t get the data that they need, and 89% are required to show impact measures. These challenges lead to silos preventing deep engagements, donor databases not being a complete platform, and difficulty measuring and demonstrating impact.
Become a Connected Nonprofit
So to solve for these challenges, we’ve worked to create our vision for nonprofits of all sizes called Connected Nonprofit. The Connected Nonprofit is not an Excel file. It’s not only a donor database. It’s a complete platform for impact.
What we’ve done is we’ve taken the incredible Salesforce.com technology and built solutions for fundraising program management and engagement to ensure that you’re able to accelerate your impact and achieve your mission.
Customer Success
We have a variety of operations that have seen success with our Connected Nonprofit vision. We’ve seen a reduction in cost-per-dollar rates. We’ve seen an increase in general retention and program efficacy and 88% of Salesforce.org customers say that Salesforce helped them improve their ability to achieve their mission, which is super exciting.
All right. So with that, I’m going to pass it over to Lisa to talk about how we’re building artificial intelligence into this platform, to make it even smarter than ever. So, Lisa, over to you.
Lisa Rau (LR):
Thank You, Andrea. I am so excited to be talking to you all about AI. I think AI is the most interesting subject in the world, because, in order to solve AI, you really have to start thinking about how the brain works, and that’s one of the great mysteries of life.
Academic and Applied Expertise
Anyway, I’m Lisa Rau. And just a word about my background. I have a bachelor’s and master’s in computer science, and I specialized in artificial intelligence. I have a PhD. My actual thesis was about information retrieval and trying to mimic the brain in particular real interesting phenomenon, like tip of the tongue. How is it that you know you know something, even when you can’t remember what it is? That’s what my thesis was about.
I worked in artificial intelligence and basic research, starting at Berkeley, with a group. One of my members of the group had literally wrote the textbook on artificial intelligence, Peter Norvig. Then I went into a research, and I spent about eight years doing basic research in AI and the subdiscipline called natural language processing. And so, I slowly worked my way out to coming around to wanting to make the world a better place. I founded Fíonta in 2001. So that’s a little bit of background.
Fíonta
My company is a premium partner with Salesforce.org. Our number one thing is working with clients on implementing Salesforce, but that’s not really what this is about.
What is Artificial Intelligence?
Anything that seems to require human intelligence
This is actually about what is artificial intelligence, and I am going to explain it so that no matter who you are and what your background is, hopefully, you’ll come away with this with a lot better understanding of the complexities, of what’s possible, where we come from.
So I wanted to start out with, what is artificial intelligence? The easiest definition is just anything that seems to require human intelligence that a computer does, is, by definition, artificial intelligence. There is no better definition than that.
Artificial intelligence is actually not one thing. It’s a bunch of subdisciplines. The core areas of artificial intelligence – research and development – these are in no particular order, but I’m just going to go through them to give you a little flavor of what’s up with these areas.
So natural language processing is where computers try to read and understand text. The reason why that’s hard is, a canonical example is, I saw a man on a hill with a telescope. Well,that could mean that the man on the hill has a telescope. It could mean that you saw the man because you were using a telescope. It could mean that the hill happens to have a telescope on it and that’s where the man, and that’s just one pretty and easy example.
It turns out that in order to understand language, you have to understand everything, and context is key. Context really is an AI killer because it’s very hard to have context. The problem solving is a much more tractable problem in artificial intelligence. Some of you may have heard of Deep Blue, which was IBM’s chess-playing computer that beat the chess master for the first time. A computer could actually beat a human in chess. And Google’s AlphaZero, which was its chess-playing entrant, that was phenomenon. It actually played Deep Blue, so we got machines playing machines. But what it did was really cool and scary. It taught itself how to play chess before it beat Deep Blue. So, I’ll probably come back to that.
But problem-solving, again, is a lot more tractable than natural language processing. Robotics falls under AI, and it turns out that even things like walking is incredibly complicated. The interplay of neurons and muscles and all these things.
Vision and image recognition processing, I’m sure many of you have seen how quickly this field has increased in capability. I read recently that China, for example, is now applying facial recognition to all of its citizens on the street and it’s gotten that good.
Speech recognition still has ways to go, and why is that? It’s the same reason natural language processing is so hard. Many of us who’ve gone into those voice response systems may … Especially if you get frustrated, the system just keeps saying, “I don’t understand what you’re saying,” because it’s still not able to clearly recognize the speech, because the speech is tied up to the national language processing. Another area is actually speech generation, which you may think sounds pretty easy. But if you’ve ever used the directions, the directional navigation that speaks to you, recently, I was on a vacation and using one of those devices, and it said, “Turn right on a street,” when it was actually A Street. It could not tell the difference between A Street and a street. I’m sure we’ve all heard those kinds of disfluencies.
Information retrieval and search is also part of AI. Just one look at Google and you can see how seemingly smart Google can be sometimes, about finding the information that you think you want just by having very intelligent retrieval.
Machine learning, we’re going to go into a lot more. Machine learning is intuitively when you’re teaching computers how to learn something. The thing that everyone’s scared of is once machines really can start to teach themselves things, then we have no idea where that’s going to end up and where they’re going to come to. But that’s not for today.
There’s a lot of sub-areas in artificial intelligence. I just read today that they’ve created systems that can clone your own voice. So if you want to have your directions in your own voice coming out of your car, you can record them and the system will then extrapolate from that.
Mapping facial expressions to emotions is another kind of subdiscipline of image recognition, where it tells if you’re happy or sad, or the reverse, having computers that can really generate appropriate facial expressions. These are very hard problems, very easy to understand, but actually very hard to solve.
Artificial intelligence draws from many different disciplines. One of the interesting things about it is its multidisciplinary nature. It lives first and foremost, as a branch of computer science, but it draws on logic. It also can be informed by cognitive science, which is the study of thought and learning and mental organization. It’s also helpful to know something about psychology. If you want to try to model how the brain works, you might be a neuroscientist and look at neurophysiology. It involves philosophy and what is the meaning of thought, what does it mean to even be a human being. Linguistics for all the natural language processing. And, just core kinds of engineering, mechanical engineering, other kinds of engineering.
One way that people often talk about determining whether a computer is intelligent, is this test, the Turing test, named after Alan Turing. The idea is, once a computer passes the Turing test, you have a computer talking to a human being, and the human being cannot tell that they’re actually talking to a computer. We have not solved the Turing test yet. There are many ways to try to trip it up, and humor is a really good one. If you’re ever faced with trying to take the Turing test and you want to determine if it’s a computer you’re talking to, just try to tell it some jokes and see what it does. These are the things that make us human.
Types of Artificial Intelligence systems
So there’s really four different kinds of AI systems that have built over the years. The very first thing that people tried were called expert systems and they were rule-based, and there’s still a lot of these in play. This is the kind of thing anyone can imagine writing a program to do, which is, you just encode all of the knowledge as rules. One of the first applications was in the medical arena, where a patient comes in and you type in their symptoms, and it’s just got a whole bunch of rules that say, like if they have a rash and their throat is swollen, and let’s increase the score for allergy, and it goes through and then it suggests to the doctor all these diagnoses.
Another kind of artificial intelligence systems are our search are heuristic or algorithmic based, and some of you may not be comfortable with words like heuristics and algorithms. They’re just too multisyllabic. But heuristics are really just ways of naturally describing something that you can do, like a perfect example, to find a way out of a maze, keep your hand on the left wall. That’s a heuristic, which we know actually works to get you to the center of a maze assuming it’s a legitimate maze. An algorithm on the other hand is just like following a recipe. It’s just a series of predefined instructions.
And so, when you build an AI system using algorithms, you’re coming up with heuristics to try to simulate intelligent behavior. I think search falls into this category. Search is one of those things, as I mentioned before, that really can simulate intelligence. Well how does that work? I’ll just give you a simple example of why Google is so powerful, is when you type in a question into Google and it immediately returns on your list of web results exactly what you were looking for, the reason it does that is it has millions, tens of millions and hundreds of millions of results of other people who have looked for the same thing and then clicked on that same answer. And so, its ability to find a matching plan answer based on all that data simulates intelligence.
Similarly, if you put the encyclopedia into the computer and a really great search algorithm and just asked it a question, it could immediately go to the encyclopedia, find that piece of information, give it back to you. Does it understand the information? Is it intelligent? Or is it just really good at finding information? At some point, we don’t really care.
Another kind of AI system that has been tried and frankly failed, this was actually the area that I started out in, was people got the idea that you could represent human thought because we don’t actually think in words. We talk in words. But there’s this notion of concepts inside our head. People got the bright idea to represent concepts in computer language like this is the very simple representation of what it would like, the concept of a dog eating a bone, where the bone is the object of the ingestion. There was this whole idea of semantic representations. And if you could conceptually model the whole world, then the thought was that you could have a computer that could actually be very smart, and millions of dollars have been … tens of hundreds of millions of dollars … In one case, there was one crazy AI researcher who actually tried to model all of human knowledge this way, thinking that that was going to work. Well, it turns out it doesn’t work like that. It’s much harder than that.
Machine Learning
But what has worked? And why are we all talking about AI now? It’s really about machine learning. Machine learning, it turns out has really come into its own. I just want to give you a flavor for this. I’m going to talk about just two different kinds of machine learning. Well, two different aspects of machine learning. The one that’s used the most, sometimes, referred to as deep learning. And here, what happens is, you take a training set of data. What a training set is, is examples of the information that you want to learn, like let’s use an example of what does a face look like. You compare it to actual faces, and the computer tries to understand the features iteratively. So first, it might combine pixels to form the areas that are darker or lighter. Then it might resolve them into, okay, there’s two things at the top and one thing at the bottom. Then it might figure out, well, those must be the eyes and the mouth. Each layer builds on the other. This can be done in a supervised way or in an unsupervised way by just dealing with patterns and figuring out what is the commonality. If you give it a million examples of faces, it can actually abstract what those qualities are. So that’s kind of interesting.
One of the things I find the most interesting is Bayes’ Law. I swear this is the last slide that is any math-like, but Bayes’ law is actually really powerful and really interesting because it’s really simple. All Bayes’ law is, is calculating the conditional probability. What is that? It means what’s the probability of something happening, given that something else is happening? It’s a very simple formula. I really have it there right now. I just want to give you an intuitive understanding of how Bayes’ law works.
So say you have a donor who’s given three out of the last four years, John. Say you have another donor, Many, who’s given $100 just one out of the last four years. So what’s the probability that Mary is going to give $100 this year? That’s easy math, right? One out of four, 25%. We’re all good with that.
But then we add in some other information. Like, every time, Susan’s the one who asks … When she asked Mary, she gave. So that was one out of one for Susan, and same thing with John. What is the probability then of Mary giving $100 if Susan asks her … It’s now 50% because you’ve added in additional information. It’s the probability of an event based on prior knowledge of conditions that might be related to that event.
Let me give you just another simple example because this is shockingly powerful. So, let’s say cancer is related to age. So you might initially know what your probability is of getting cancer, but once you add in what your age is, that changes the probability. So, Bayes is very different. Instead of saying, “If A then B,” it really is working backwards and saying, “Given B, what are the causes of B?” I would love to go into this even more, but Bayes’ Law is actually underneath a lot of the statistical models.
One of the reasons this is important, I know some of you probably can’t see this that well (Cognitive Bias Codex image). It’s one of my favorite things. It’s called the cognitive bias codex. Why am I showing you the cognitive bias codex? This shows all of the different ways that human beings are very flawed reasoners. We’re really bad at applying probability in a legitimate, objective way. Computers don’t have that problem. So I’ll give you one example of a cognitive bias. It’s called the recency effect, which is like the last thing that you saw. That’s what … It gets a bump up. That’s what you think you’re going to see in the future. We overweight certain factors; we underweight others. We disregard data. We think lightning striking is much more probable than things. This is very interesting, one of the real benefits of AI.
Okay. So there are problems with all these methods. Expert systems just took so long to create and they really only work in very circumscribed domains. Then search is still a really wonderful method of simulating intelligence, but these programs get very complicated and they’re hard to understand. They can get large, complex, hard to maintain, hard to extend, because it’s all just programming and code. The semantic representations, I don’t think anyone tries to do this much anymore. We just know too much. We have really limited success. They have some limited success and highly constrained domains, but they’re really brittle, easy to break and very expensive to build. But learning, machine learning has really started to come into its own.
While I’m sure that many machine-learning researchers would like to could take credit for the fact that it’s come into its own because of their amazing new algorithms, and there have been some developments. The real reason that machine learning now kind of works is because of Moore’s law, which is, of course, the law that says that our computing power will double every 18 months and that we’ve held to more or less since computers started.
As computers get faster and memory gets cheaper and we have more and more data, algorithms that didn’t seem to work before now have a lot more to work with, so they can explore more deeply and do more processing and extract what there is to be extracted.
Salesforce really took this innovation. So a lot of Salesforce is, which is what I’m going to talk about next, was that they took the idea of machine learning. Beforehand, machine learning required a lot of preparation of your data set in order to learn it. You had to figure out what the features were that you wanted to feed into the machine-learning algorithm, for example.
Salesforce got the bright idea that you could actually have the computer figure out what the right attributes were on its own. It could actually learn its own model, and that eliminated a big barrier to widespread use of machine learning, because now, it just works automatically. You can just set it running on your data set, and it’s going to learn what there is to learn. Well, that sounds like magic and it kind of is, but I’m going to talk a little bit more about that.
What does Artificial Intelligence Mean for Nonprofits
Sample areas of applicability
So, I want to talk about what … You’re here because, hopefully, you’re nonprofits, and you’re interested, you know, “Where might this be relevant for me?”
I’m going to talk about easy things and then some things that really aren’t ready for primetime. There’s a lot of area of applicability for AI across all of nonprofit business processes. I just listed a few common ones. Of course, many nonprofits need to raise funds, and they need to apply for grants. They have volunteers they want to engage and do well by. They have communities to feed and provide educational resources. They want to do advocacy. They want to engage in marketing campaigns that have a real ROI. They have programs that they’re trying to run. There’s actually artificial intelligence areas across all of these to help you. I’m going to go through a couple of the capabilities that Salesforce has.
So hopefully, most of you have already heard about Einstein. This is the artificial intelligence product suite that Salesforce has branded. What it does is it works on the Salesforce platform. I’ll show you. If you look at this chart here, this really shows how Einstein fits in. So, on the bottom is your data. Nothing works without your data, and your data is really key to all of artificial intelligence. There’s a suite of tools that Salesforce is making available called myEinstein that helps to make predictions. This is complicated stuff and it doesn’t work out of the box. I’m going to talk a little bit more about that just so you can see what it looks like.
Then there’s another product called Einstein Analytics, which you can buy, which provides a lot of forecasting and more business intelligence analysis, using a lot of the statistical tools already built in. So that’s a very powerful product that can help you to make sense out of large quantities of data. But, you actually have some functionality without having to purchase anything, in that, Einstein has been embedded throughout Salesforce’s Clouds, all of the Clouds. Most of you nonprofits have the Nonprofit Success Pack (NPSP). So we’ll talk about the Sales Cloud with the NPSP. That’s where I’m going to focus now.
Einstein functionality available in Nonprofit Success Pack
Einstein works across on the Sales Cloud. All I’ve done here is show you the different areas, where Einstein works across the Sales Cloud. Unfortunately, as those of you who have a Nonprofit Success Pack know, it works with a household model. The one features that have asterisks, those are the ones that the Einstein functionality works out of the box, with the Nonprofit Success Pack. If you don’t have the NPSP, all of these features are available right now. I don’t know what Salesforce’s plans are for making all of them available to work with the NPSP. I’d be surprised if it didn’t come along. But I’m just going to talk about the ones that are available now.
Easy Things – Data entry automation and insights
So, let’s look at data entry automation and insights. Again, this is available now. You can turn it on in your NPSP instance. What this does is it connects your email and your calendar to come up with next activities that you might want to do. So the example here is you are planning to meet with the prospective funder, and it’s automatically creating that as the next step and a list of activities that you’ve done to try to solicit interest with that funder. So that’s pretty cool, to automatically have email and calendar contribute to activity. So, it really reduces the data entry, and it also allows people in your organization to see what’s going on with the solicitation of the funders.
Easy Things – Email Insights with Inbox
Another cool AI functionality is this Email Insights, and this uses some of the natural language processing functionality. What we see here … I don’t know if you’ve ever had this. Well I have this every day, right? You’re about to reply to something that’s going on. Before you can do that, you have to read an email thread, and I start from the bottom up. These things can get long, right? It can be 20 messages. You have to read all of them to figure out, okay, what’s going on in this situation? Well, what Email Insights does is it actually is able to match emails to types of events.
That gives you the thread so that all you need to look at is the red circled items, even if the email isn’t shared with you. So say your major gifts officer is on vacation and a major donor, a prospect, calls in, and you’re on the hot seat, well you want to see what went on. You can see they were on vacation because they’re out of office at the bottom. Then they talk to their superior about giving you a grant, and then they talk to you about their intention in terms of how much money they wanted to give. And now, it’s time to set up a meeting. So that can be pretty powerful in order to allow you to see the history of an email chain, without having to go back and read.It can alert you to when it’s time to actually approach a donor, about something, about an ask that you might be given.
Easy Things – Opportunity Insights
Similarly, with the opportunity insight, so what you see here is that the system has noticed that there was an opportunity that had been scheduled to close that it hasn’t closed, and it goes into the email and it sees that, in fact, there was this next step about Jim wanting to talk about the option. So it wants to alert you that, “Hey, maybe you should contact Jim, because it’s been a few weeks or few days, and he wants to talk.” So that’s kind of a high-tech functionality.
Easy Things – knowledge management for Communities
Another big area that I’m excited about piloting is in knowledge management. A lot of us have lots and lots of knowledge resources. We may have a community, Salesforce community. What Einstein can do there is automatically order and suggest tools based on the history of interest. So what the machine-learning functionality does, is it identifies certain characteristics of users, their history of what they’ve been interested in, what kinds of resources they might be looking at correlating, and it can automatically order these items based on all of that data, by learning what might be most of interest to a specific user based on whatever features of that user you might have in your database. So that’s kind of cool. That helps to keep members engaged.
So, say you have a very active community. Every day, the system can see what’s trending hot and automatically reorder. It’s not just based on click-throughs. It’s based on a variety, really, whatever data you’re collecting within your community, and these knowledge resources can be reordered dynamically, based on the interests of individuals. It can also help to identify potential phrases that you might want to highlight, so that user-generated content can then be categorized. So that you’ll see all of a sudden, all of our users are talking about this weird thing called GDPR. Even if you never program GDPR, as a topic area, it will notice that there’s this surge of interest around this new word, using really fairly easy-to-understand, statistical methods. It’s going to determine that this new word shouldn’t have the frequency that it’s starting to have, and therefore, it’s going to accelerate it.
Easy Things – Marketing and Engagement Insights
Then here’s another capability in the marketing cloud. What we see here are predictions. The predictions, they only work across all of your contacts. So, right now, it’s not going to be able to predict who’s most likely to donate this year, if you have your donors there with your other supporters in your database. So it runs across all of your contacts in your Salesforce database, but what you see here is how it calculates email open rates. How likely our email is going to be click-through, converted. How likely is it that someone’s going to unsubscribe and so on and so forth. That’s very powerful data in terms of the predictions.
Harder Things
The predictions actually are harder. I mentioned before that myEinstein, it’s really a set of APIs and tools. It’s not something that you can use out of the box. But if you have someone who’s a tinkerer and really has a background in this and isn’t afraid to play around, it’s very exciting that Salesforce has provided this set of tools to build your own custom AI applications. But it is hard.
Chatbots also are hard, but they’re coming. I liked it when the chatbots were chatting with the other chatbots. It’s like they got into a … They didn’t even realize they were talking to each other, and who knows what they were saying, right?
But those of us who have input from live channels, when you see the same kinds of inquiries and questions, you can put in a chatbot to automatically respond.
Vision is another area that’s really hard, but there’s a lot of really interesting applications. For example, in where I am, here in Washington DC, we have an organization that counts trees in the city. If you have aerial images, you can actually use image recognition to determine changes in the tree canopy.
Natural language processing, also as I mentioned, is harder, and the Prediction Builder is hard. I’m going to show you the Prediction Builder though, because it’s cool.
So, this is the Prediction Builder. You start out by choosing what you want to predict. So say you want to predict the likelihood someone’s going to become a major donor. You say, “Which object is that?” Well a donor is the supporter, and we want to predict major donor. So you point and click. That’s what I want to predict, because it’s asked you now, “Select a field to predict.”
Then you just have to say, “Okay, what do I want to call that prediction?” You want to say, major donor predicted. Then it gives you the list of how many fields it’s going to use to make that prediction, which is their entire giving history. It goes off and it does a lot of calculations. Then it comes up and it assigns that prediction value to every donor in your database, and that’s how the Prediction Builder will work for that.
So that was just hopefully to whet your appetite … tried to give you some understanding of what is present in Salesforce Einstein and the background of what AI is all about.
Steps You Can Take To Prepare
And say this really interests you, and you’re ready to take the next step, what would you do? Well, you can actually run this Einstein Readiness Report, which I would encourage you all to do. You might be a little disappointed though, because most of you probably heard garbage in, garbage out. Well with AI, it really is critical that you have enough data so that the statistics and the analysis that it’s doing automatically can work, because AI is really only as good as the data you have. And by good, I mean complete, and by complete, I mean each record is complete. Then you actually have complete information across all the records. It’s unduplicated, that it’s consistent. So people are filling in the same values consistently and then large enough and representative enough. Even with that, so many of you may have datasets that are just too small to really be statistically legitimate or are in a state where they’re not able to generate things.
So, it’s always a good idea to pilot. Pilot before you purchase. If you’re going to use the things that are in the Sales Cloud, pilot before using, thoroughly.
Unintended consequences
There are definitely some downsides, and the main unintended consequences of using AI is that these algorithms operate as black boxes.
I hope you got some feeling for the kind of statistical analysis that goes on behind the scenes. That means that there can be bias that’s introduced, and there’s two main types of bias. The first one is statistical bias. I mentioned if you don’t have enough data and it’s not representative, it’s going to draw some conclusions. But it’s based on such a small sample that all bets are off. And so, you may be training the system to predict something based on your dataset, but that’s not actually reflecting reality. It’s only reflecting the data you have. That really is true when you have relatively rare things, small groups of people.
More of a concern to me is the social bias, which is that we may be … There are artifacts in the analysis, and I think the quintessential example is loaning money. So say you’re an entrepreneurship program and you calculate whether or not you should give someone a microloan, and you train the system up on all the people who have successfully repaid their loan and/or had loans or maybe you get a dataset from somewhere else, and it turns out that, say, people from a particular Zip code were bad loan recipients. It can perpetuate a historical bias because there’s a bias present in the training data.
It may not be that your system … all the system knows … It doesn’t know why. It just knows what the data is telling it. And so, you have to be very careful not to perpetuate. Your goal here is to try to give loans to everyone who qualifies. If the system doesn’t have a legitimate basis for that and you can’t tell because it’s a black box, that’s a problem.
Steps you can take to prevent bias
So what can you do about that? Well, first of all, you can’t just let the model run on its own. You really want to check to look at the results and make sure that that’s comporting with what you think the reality should be. You also want to get input if the decisions that you’re making, based on these capabilities, impact your supporters. You want to get them involved.
Always collect and label new data so that your AI functions will continue to evolve, as you continue to get new data.
I think the most important thing to do is to treat it like a clinical trial, which is, do test the AI against control samples, just like you would for any new functionality that you’re rolling out. That means also determining the error rate. Look at when the system’s right, when it should have been wrong; when it’s wrong, when it should have been right. And trying to understand what are the implications of that, because all systems are going to have error margins. You just have to be deliberative about it and mindful.
Also, just in terms of data hygiene, you want to get rid of old data because again, it may not reflect current day models. So you might want to think about, “How long do we want to keep this data once I decide it’s not useful anymore?” You should get rid of it and no longer have it cluttering up the calculations that these systems are doing, and be respectful of privacy. Just practice good data hygiene overall. The GDPR has really forced a lot of nonprofits to start looking more carefully at their data. Treat all personal data as need-to-know. It’s just a good opportunity if you focus on your data to try to incorporate good data hygiene and re-evaluate privacy and security considerations.
Also, just to be a good steward, if you do have a really large source of data, real curated data is very valuable to the community. The machine learning and AI community has been using the same standardized sets of data for decades, because it’s so much work to come up with really good representative examples of different kinds of behavior. Then train your staff on these points so that no one goes off and makes decisions without following these practices, and speak up, so ask.
In Conclusion
What I really wanted to leave you with is that artificial intelligence is really interesting, and it’s already here. While a lot of what you read about in the press is just statistics and math, a lot of what you read is not like breakthroughs, even though it may be couched in a breathtaking article about how insane artificial intelligence has become. It still is powerful. While it may not have a place in your nonprofit yet, you really want to keep an eye on the developments. As Salesforce continues to roll out more and more of these functions and make them generally available, which is what you’re going to continue to see, because they got these roadmaps coming up. You’re going to want to continue to track the progress and understand where and how some of these capabilities are going to help you to make better decisions, reduce cognitive bias, automate certain time-consuming tasks, provide more benefits to your supporters, and generally, really align your services to your mission.
And know the robot apocalypse is not coming anytime soon. It turns out that human beings are really complicated. No one really understands how the mind does what they do, and we’re continuing to make progress. But it’s very, very slow, and I don’t think we’re going to see robots take over the world anytime soon.
So with that, I was going to turn it back over to Andrea.
AS:
Great. Thank you so much, Lisa, for some really, really fantastic information.
Resources to Get Started
So before we jump into questions, we want to provide some resources for you to get started. If you’re not familiar with it, we have a really great online learning platform called Trailhead. If you are interested in learning more about artificial intelligence, you can take advantage of Trailhead. To learn about AI, we have various trails focused on artificial intelligence and the various tools that you heard about today.
Become a Trailblazer in AI
So here are a variety of the different trails available. We also have something called Trailmixes. We’ll send you Trailmix in the follow-up email with all of the trails that we recommend taking around Einstein.
Einstein Readiness Assessor
Okay, and if we go to the next slide, as Lisa talked about, we have an Einstein Readiness Assessor. So if you’re not sure if you want to use Einstein, Sales Cloud, or if you want to understand if your org is ready to use it, you can run this readiness assessor. The link is here. Again, we will send this out on the email, but definitely be sure to run this assessor to see if Einstein is useful for you.
Don’t leave empty handed! Resources for you!
We also have some other resources not directly tied to artificial intelligence. But if you are interested in meeting real live Salesforce users, please do join local user groups. We have over 80 throughout the world. They’re interested in learning more about the Nonprofit Success Pack. We have a demo there that you can watch, and of course, Trailhead will teach you everything you need to know about Salesforce, in addition to all the great AI capabilities. We also have a variety of different webinars coming up that you can take advantage of as well, including the highlights of Summer ’18 Webinar, some software features today, but we’ll have even more on June 20th, and we’ll also be demoing customizable roll-ups, our newest feature in the Nonprofit Success Pack.
We also have the getting started for importing data, for fundraising on June 18th, and then we have a webinar focused on developing an impact measurement strategy on June 31st. So other good webinars coming, as well. So with that, we will open it up to questions, some very high-level questions that I can answer quickly. One is, is this webinar will be recorded? Yes, it will be recorded, and we’ll be emailing the information out to everyone, so stay tuned there. And second is, is Prediction Builder is available now? So Prediction Builder is in pilot mode. We have some nonprofits that are part of the pilot. It will be generally available in Winter ’19. So keep an eye out for that being generally available, and the team is super excited about that tool.
Questions and Discussion
So now, going into questions. Lisa, we have one for you. Let me pull it up here.
Okay, so, there’s a question around impact measurement. So if you’re able to give us some information around how you see artificial intelligence really helping nonprofits show their impact. Any insight there?
LR:
Yeah, I actually think that if you do a really good job of measuring your activities and the tracking your outcomes, then it can correlate using, like Bayes’ Law. What kinds of activities are going to be contributing more to your impact so that you can double down on those? And that’s what Einstein Analytics actually would really help with. So, say you run 10 different kinds of programs, and within those programs, there are lots of measurements. You take, like, attendance and grades and class participation, and then you track your students as they graduate from school or get a job, how much income they have. You like to whether they’re married, whether they’re going to jail. You know, all kinds of different data that you might be tracking. Einstein Analytics will be able to surface contributors to positive outcomes in a way that you might not be able to, because it’s a very powerful business intelligence engine.
For those of you who haven’t already defined how you’re measuring your impact, I think this system could be good at determining what the core factors might be in terms of how people’s lives have changed as a result of your program. So I’m trying to speak generally to all kinds of nonprofits, but hopefully, that gives you some idea of where it applies.
AS:
Okay, great. Thank you so much. We have a viewer who works for an environmental nonprofit, and she is interested in finding out more about the vision aspects of AI, so the counting tree canopy example that you gave. Are there any resources that you could point her to or any information that she should be reviewing to understand those capabilities?
LR:
Sure. I mean, she’s welcome to shoot me an email, which is on the screen, offline, and I’ll send it out. I think, actually, for image processing in the nonprofit arena, that is one of the best applications for things like desertification for even clarity of the oceans and pollution, and really, any images … Oh, I had one client who was using this for mountaintop removal, for identifying areas where coal miners were polluting. I think that that’s a really excellent method when compared with the human method, especially of analyzing the change in certain features. If you have maps of rivers, you can determine the change in flow. Used to promote global warming issues in lakes and so forth or glaciers, perfect example, to automatically calculate the shrinkage there.
So yeah, hit me up offline and I’ll send you some suggestive articles. But that is a great opportunity for the image processing. All that you need though … and this is the kicker, which I don’t know if it was clear in the presentation. But the kicker is that the image processing in the natural language doesn’t work unless you’re willing and able or have a very large training corpus. An example of what you’re looking for already annotated that this is a river, and this is what this means and so forth. Just like in the olden days, those of you who are my age may remember when speech dictation technology first came out, you had to train it for like a couple of hours by reading things off the screen, for it to build up a model so that it could then take transcription from your own voice. A lot of the image recognition and natural language processing is in that state now, which is that it requires a fair amount of effort to get it trained up before it can do its job.
AS:
Okay, great. So we actually have a question that relates to that specifically. A viewer wants to know if you can go into more details about the size and quality of data required for effective machine learning?
And then we have another question around best practices for data cleanup. So I’m not sure if you’re able to talk about…
LR:
Yep. So, yeah, the size and quality of data in order for machine learning to work, that’s like a trick question, because if you only have like two pieces of information, like a person’s first name and the amount that they gave last year and it’s all complete … Well, that’s a bad example actually. My point is that it really depends on your data. There’s no one answer. It depends on what you’re trying to do with the data. So I’m not a statistician, but I think we all understand the concept of a statistically significant result and that’s where your answer lies, and that’s about as far as I can tell you.
So again, just for shorthand, there is an answer. It’s not a uniform answer across every database. It really depends on the specifics of what you want to do, and you need to know … You can calculate how much data you need in order to generate a statistically significant result. That’s the kind of analysis you need to do so that you can be sure. But it’s a lot of data unless you’re just looking … Part of it also is intuitive, in that, you know, you have to look at the distribution. If you’re getting a really good distribution. That is, let’s say, you have donations that vary from $1 to $1,000, but you’ve only got 1,000 donations. You’re not going to have a good distribution there. But if you have a million donations between a dollar and a thousand, you start to see a graph and some commonality, like you would if you were flipping a coin over and over again. There comes a point when you feel pretty good that you’ve covered the waterfront in terms of the probability distribution of your data, and that’s the point at which the machine learning really can do its own.
As far as the best practice for cleaning up data, that is a subject that I love. There are so many tools in Salesforce to help you clean up your data. And you 501c3’s… Actually, I’m not sure. I think DemandTools might have eliminated their donation program. But they used to give away their demand tools free, which is also very powerful. You want to de-duplify your data. There’s, again, tools built into Salesforce to allow you to find that. You can run reports on your Salesforce instance. The optimizer report will give you valuable statistics on which fields are incomplete and to the number of values within each field. So, with the data, on the data, so we’ve got metadata there. So with the data on your data, you’ll be able to understand what the issues are. There is no substitute, in many cases, to some kind of manual effort to fix things up. In some cases, it does require human effort to determine what you should do with the data.
The best thing that you can do to get good data is to prevent bad data from getting into your system in the first place. That means making it really easy to automatically detect a duplicate when people enter it, having required fields, putting in rules that require data be in a standardized format. Those are three things that you can do today that will significantly increase the quality of your data moving forward.
AS:
Okay. That’s great. Thank you so much. Okay, we have a question around AI and fundraising. So if you’re using Salesforce for web program management and fundraising, are there pitfalls that you would see in AI, being able to tell the difference?
LR:
Yeah, I think that is a very good question. In this instance, the Einstein functionality, it cannot tell which data is donors and which are program supporters. So depending on what you’re trying to do, it would get confused. I don’t know what the future is for that. One of the things you may have noticed I did not suggest is that you can use this kind of Prediction Builder for the donors. Well, what this example is, so yeah.
I think that is a problem because it doesn’t … Well in this particular case, we’ve categorized our donors. We have a separate field for donors. And so, as long as you can restrict it in that way, you should be okay. Does that make sense? So if we look here on the screen, it’s got supporter. Now, in your case, your supporters are both donors and program participants. If you can restrict that, then you can address that issue. I hope that was clear.
AS:
Okay, great. Thank you. Okay, we have another question around the distinctions between fundraising metrics and impact measurements. So any insight you can provide there?
LR:
Fundraising metrics and impact measurements?
AS:
Mm-hmm (affirmative).
LR:
Well, I think I need a little bit more context for that. So, fundraising metrics typically are how much you received in donations. Impact is usually the results of your program, how it is actually changed the world as a result of the activities you’ve done. Please feel free to ask a follow-up.
AS:
Yep, yeah. That was all the information we got in the questions. So if you asked that question, feel free to throw in a follow-up. If not, we will go to the next one. I see the last question that we have here is, if we wanted to have a deeper discussion on AI and how it could fit into our nonprofit, what’s the recommended next step that you have?
LR:
Well, I would first go through the Trailheads and I would read the documentation that’s available on how these systems work. I think you have to educate yourself about it. For myself, being a scientist, I would certainly try to pilot it. Obviously, I’m happy to talk and have a deeper discussion with you, as well. You can always approach your account executive and see if they can help to connect you.
This stuff is actually really new at Salesforce, and you can tell that because, as Andrea mentioned, some of the functionality isn’t generally available yet. Actually, a lot of the functionality is just coming online now.
But yeah, that’s where I would start. You also can join the Chatter group. There’s an Einstein Power of Us group that those are available to clients, aren’t they, Andrea? Do you know? I don’t think they’re just partner groups. So you can actually join a Power of Us group on Einstein and ask some questions of your peers. It’s always a really good option, as well.
AS:
Yeah, there is a group focused on the Nonprofit Success Pack, so you can join that, and a ton of questions get asked there about all walks of nonprofit operations. So, I would definitely take a look there.
Great. Okay, those are all the questions that are relevant that I have on my end. Now we have five minutes to spare, and I don’t see any other questions coming in. So with that, thank you so, so much, Lisa.
LR:
Good.
AS:
We’ve gotten some great feedback really on how much people enjoyed the presentation.
LR:
Good. Thank you all for tuning in. See you in cyberspace.
AS:
Thank you. Bye.