Latest / Beyond the Slop: AI in the Non-Profit Sector / Episode 4: AI supporter care chatbots, ethics, and... the apocalypse?
Transcript
- Craig: Hello and welcome to this week's Beyond the Slop. I'm Craig Linton. Anna: and I'm Anna Hissenbrook. And on this week's episode of Beyond the Slop, We're gonna start off with what we've been building. We're gonna talk a little bit about whether AI is gonna kill us all. You know, just lovely Friday chats. And we're going to be answering our very first listener question. But first of all, Craig, why don't you kick us off? What have you been building this week? Craig: Yeah, it's been a been an interesting week and one of the things that's been quite exciting is a voice AI to answer calls for one of my clients who've they've got this campaign running and and been inundated with calls and really struggling to keep on top of it. And, you know, li we're talking like hundreds and they've got one person trying to call people back and so we've put in a like a an AI triage. And I'll be honest, really interesting. actually s see if you get you know, w this is, probably scooch to a slightly older audience. So we're asking for feedback and score out of five. But w where do you think it's coming in at at the minute, Anna? What do you reckon the feedback to like the interaction with AI, you know, like we're we'll record for training and monitoring and you know, we've basically at the end of the call you asked you sent a text to ask to to give Anna: ooh. Craig: a score like one to five. So is it positive or negative, I guess? Anna: I love this. Well it could go both ways obviously, but I think, I think maybe it's surprisingly positive. I think you're getting a four. Craig: No not quite a four, but when I when I last checked it was it was about three and a half and it's it's really interesting. You've got this small cohort, I'd say about ten, fifteen percent, who truly, truly despise it. I mean are like getting really angry about it. but I think they get angry with anyone, whether it was a human or an eye, I think, to a degree. and just, you know, can't can't deal with it. But then actually, because a lot of the things are the same questions coming up time and time again, people are getting a quicker answer and it's, you know, getting what they need and we've built the knowledge base and and you know, this frequently asked questions block. now it's interesting because you you'd have thought my hunch before was we'd score low because it'll be the people who really not happy about it who would take the time to feed back. But actually, you know, what we're finding is more people are responding positively than negatively. And just over half of the calls are being solved by the AI, meaning that, you know, it's it's it it's less than half the volume then that the team have to follow up. So saving them tens and tens of hours of of time to to do it. So it's it's been really interesting. There's been some tweaks. The other thing that we've had this really interesting discussion because when you're using eye calls, you know you get that awkward gap you know, is it speaking and you speak over it and it can be you know that that that's something called latency. How how quickly does the eye respond? And what we had to choose was the model with the best latency. So it it you don't have that talking over each other. But on that model you can only have like a an American sounding voice. Rather whereas some of the slower models you could have a UK voice and we have so some of the Anna: Okay. Craig: complaints have been why are you using an American voice? Which was it's it's for the speed, but Anna: Fair. Craig: Yeah, but it was it's it's been really interesting and I think more more good than bad, I've I've been pleased with the the results so far and we'll we'll continue to monitor it. But it's it's definitely speeding up, something that would have taken the team weeks and weeks to get back to and you know it it's helped with this backlog. And related to that, we've also done our first email response bot where people coming into a shared in box and AI is drafting the reply, for people who are emailing in with with comments and then again rather than the team having to type out the same answer lots of times or you know, come up with a new answer. They've got it sort of eighty percent done there and then and then they can just send so again just saving saving time and again you could fully automate it and get the AI to send the email, but you still want that human balance in and that human touch, I think, and those little extras maybe that the AI can't add. But again, saving lots of time and The easier and easier to to build these days. I don't know if you have you have you built any chatbots or have you had any experience, good or bad, around around using chatbots? Anna: Yeah, I mean, I'll be honest and say I have no experience with building voice interaction, so of answering calls, but I have created chatbots using AI. And I have to say there's a lot of considerations, right? Because the safest option is just to create chatbots that are trained on your internal data and your internal information and a back catalog of approved responses. And then you have the option to create models that essentially are very similar to chat, GBT or Claude or any of the others where it makes a call to, to the AI and then you get a response. and you see a couple of these on, social media where someone has asked the McDonald's, you know, chat bot agent for. some random recipe for a cupcake or you know something like that. And so you have it you have a lot of examples of of ways that people can you know misuse or incorrectly use the chats and and also where people can get a really bad experience so there's a lot to consider but that being said the efficiency is crazy you could really really really make a big dent. in the amount of hours you have to spend on sort of quite repetitive redundant tasks. So I guess, yeah, my experience is it's super useful. It's a really, really good use case for charities to explore, but take it seriously and look at the options properly because they, you you can end up giving a really crappy sports or experience if you don't do it well, right? Craig: Yeah. And you know, there used to be very much like, you know, do you remember probably sure my age again as usual, the choose your own adventure books and it's like, you know, yes, no tree branching logic. And you had to go down a route and that's where you people got really frustrated with. But as you say, if you give the eye a little bit of leeway, you get the risks like, you know, getting the McDonalds bot to to answer a maths question or or recommend something. But you also give it a bit of freedom to to actually answer the question rather than giving this very precise thing that might only be seventy percent accurate. So again, I think again, like most things, they're only gonna get get better from from here really. And yeah, it's been fun building them and then tweaking and adapting to to get it just right or as as good as it can be before before you launch them. But I do think there's lots of potential for for fundraiser to do this and another client I'm doing something with they're they're gonna put it on their website but they're gonna have a humans first for a week to see what sort of questions come in and answer them and then they can build the knowledge base which is really interesting. So so yeah. good stuff. So Anna got Anna: I'm like, ooh, can I ask a question teacher? In that week where they're doing the human responses, are you gonna be collecting CSAT and then comparing once it starts being AI generated responses? Because I'm really interested to see, sounds like you said in your current sort of CSAT experiment with the AI. you're probably getting fours and fives from a lot of people and then just ones from the really, really angry ones. But I'm interested to see how that is with a human doing the same Craig: Yeah. Anna: thing, because you're to get angry people no matter what, right? Craig: Yeah, it'll be really interesting. And obviously, you know, if you're in the middle of something as a human and then, you don't respond straight away, that can be a frustration that might impact on the like you say, the customer satisfaction scores. so yeah, we'll do this live test, we'll see how quickly people are able to respond. And then the you know, the quality of the response and are people happy and we'd with that we've got just like a simple thumbs up, thumbs down response. But Yeah, it'll be it'll be good to see what sort of queries come in 'cause again you've got this what we think might happen and we we're expecting a lot of questions about charity shops and visiting hours and and and things like that but it could be something completely random and different that we haven't thought of. So it's gonna be yeah really interesting to to see what comes out of it. Anna: if anyone listening has done something like this themselves, please shoot us an email at beyondtheslop at gmail.com or reach out on LinkedIn, because we'd to hear about the use cases that charities are trying out. Craig: So I got someone mentioned my singing and sound effects this week and I'm sure they weren't being sarcastic in the slightest but I'm gonna do it again. I'm gonna have a bit of a drum roll and I'm gonna do my shhh symbols because with our first sort of listener question I was having a chat and big hello to Matt who got in touch. we were talking about various different things around local LLMs in AI and he picked up on your chat about tokenization Anna and you know keeping data secure and I thought it was really interesting ideas about how charities could run their own local models to help keep data secure but it'd be interesting to hear your take and remind listeners a little bit about what tokenization is why it can be useful and helpful and can help overcome some of those, I guess, safety fears about sharing data with AI. Anna: Yeah, totally. First of all, thanks Matt. We love questions and we love comments, so keep them coming everyone. yeah, tokenization for anyone who missed that conversation is something that I built into my charity AI toolkit for organizations in the Nordics. And that was because I built a tool that could help clean up data and analyze data. And as part of my toolkit, say, please don't upload any PII or sensitive data or personal data about your donors or, you know, anyone really. But the risk is because it's a free online toolkit, people can put any data they want into it. And that data then gets shared with Claude. And so one of the sort of securities I tried to build in was this concept of tokenization, which is that on your local device, when you upload a file, It triggers a process where any personal information, which it will then figure out through how structured your data is. So if you've got an Excel file or a CSV, it looks at the column headings and it looks at the type of data. And, we'll then change any personal information like names or addresses or email addresses or phone numbers into essentially symbols. And so if you had like Craig Linton, 100 City Road, London, then it would say male first name, second name, and then a general, you know, UK address. And so what that means is it's what also gets referred to as pseudo-anonymization. which means that you retain some markers of useful information so that the data is able to be analyzed. However, it's not specific enough so that any individual could be identified. And the point there is that Claude then does not know that it's Craig Linton. They just know that it's a male supporter within a general area. and then can analyze the data accordingly. So it's a really great tool. You do have to be really rigorous. You do have to make sure that you're testing because sometimes the AIs will promise you it's working and it might not be working the way you want it to be working. So you have to check it a few times, but it is very, very useful because most of my charity clients at the moment, are really having a big discussion about how much to connect their AI to their systems. And in my experience, at least here in Copenhagen, not very many are giving AIs access to their CRM or their email automations or anything like this, because they're just too nervous about AI having access to those sources. And I think that is very sensible, in a way. It however does limit the power of the AI because a lot of the sort of best use cases that you have is being able to automate and create, really powerful tools that just go and get stuff done without having to be asked. so tokenization is a really useful tool here because it's something you can consider to facilitate. AI being able to just go ahead and use your data without actually knowing who your donors are or your beneficiaries are or your employees or whoever the audience is you're talking about. Craig: Yeah, and that's gonna help when you get your data protection officer or someone in the IT team saying, you can't do that because we're worried about safety and all the rest of it. It helps with that. And what Matt was saying, which was something I don't know huge amounts about, but made a lot of sense to me is obviously you can now run your own LLMs. on machines and old laptops even sometimes and use that compute to almost keep it as like your own little AI book that you have for a specific use case. And he was saying that, yeah, well, even before you tokenize it or you do the tokenization or anonymous bit within your own local safe box. before you even then pass it to the AI, which I thought was really clever and he had a couple of examples of things he built and yeah, it was quite a new use case for me, but I think there's potentially lots of examples and the example that came to mind for me is if you're working in a health charity working in a space where there might be a lot of misinformation and therefore you just want to control what the information is. Running it from your own model means it can't be infected in some ways by that, conspiracy theory or whatever it might be which I thought was quite interesting take as well. Anna: Yeah, and it's really interesting because I do know lot of private organizations do this already. I have also worked a little bit in healthcare and one of the organizations I work for has extremely sensitive information, extremely sensitive data. They run clinical trials, have... you know, a lot of data that they have to be extremely careful about. And so they not only have their own private model that is closed off from anything else, it's also, you know, trained on only their own data. It's completely closed to external environments. And they have, of course, switched off access to normal Claude and normal chat GBT just to avoid any risk of employees uploading sensitive data to places. they shouldn't be. there's lots of ways that you can go about doing this to make sure that your donor data is safe and that people are using AI in the appropriate ways yeah, great question. And again, if anyone has cases where they've been doing this at their charity, I'd love to hear about it. I spoke to someone from a charity last week who had basically been using AI at their organization to re-write the data architecture for their whole organization. And that was because they had... been in a situation where there was lots of data silos and data pockets. So they had like a raffle and the raffle data was completely separate from the donation data and there was no way to spot check across and single donors were completely separate from regular givers and so on and so forth. So the data landscape was a complete hot mess. And they've been using AI to actually feed it all the context, all the information, all the sort of weird data niggles. Also to document all the stuff that's just in people's heads. Cause there were a lot of people who'd been at the charity for a really long time that knew, but when you run a campaign, you have to cross reference this Excel file with that place and make sure that you de-dupe against this list and these things. And it wasn't actually really well documented anywhere. So they've used AI to not only capture all that information that's just been sitting in people's minds, but also then map out a target architecture for their data. And that is a mega mega cool use case for AI because that's, you know, some of these CRM transformation projects are not only the most expensive, but also the most difficult projects to run at charities. And they've essentially been able to do it in house. Now they haven't got to the point where they are then migrating the old data into the new architecture and I think that's to come. But again, being really sensitive about having your own local AI then makes that process safer and more secure as well. Craig: I've got this view, I've not found any charity to back me, but you should be able to set up and create a new database, move systems so much quicker and easier than ever used to be. We all know nightmare CRM projects and how long they take and overrun and overspend. AI could and should speed that up massively and help you transfer a CRM from something that might be used to take a year or 18 months or longer to if not weeks certainly months I'm trying to persuade someone at the minute to give this a go but I'm failing miserably but I think it should be possible. Anna: Well, and I think, you're not unique in that because we saw as soon as some of these more capable models were released back in, I want to say the spring. people came to that same conclusion and you know stocks in Salesforce and some of these mega CRM companies had a huge nosedive because people are realizing actually what you were paying massive premiums and huge fees per license per user is actually pretty straightforward to do with AI now because a lot of those fancy automations Even just three years ago, was worth paying a lot of money for because that makes everybody's life more frictionless. The data is more seamless. It's more integrated. You can give a better supporter experience, but now you just don't need it. So I actually have, back in the spring, drafted a blog about this exact topic about whether or not small charities who haven't got around to getting a big fancy CRM and are still running on Excel sheets and Google Docs and whatever else they've got going on, whether actually they might be able to leapfrog the whole mega CRM infrastructure that everyone has been going through over last 10-15 years and be able to MacGyver something just as good. if they really get their head around the capabilities that AI has. So very, very interesting. And I'm with you, Craig. This is another one. This is gonna be my mantra for the episode. If you are one of these people who wants to either give it a crack or is already underway please do reach out to us, because we would love to feature it on the podcast. Craig: Yeah and we're gonna end today as usual we're gonna we look at oops I did it again but we're going back to the 80s this week Anna: Ooh. Craig: because it's the end of the world as we know it I don't know if you remember the REM song it's the end of the world as we know and of course I am feeling fine but wanted to share this it's not really oops AI did it again it's more oops AI killed us all and destroyed humanity this week which there's been quite a story has been featured on the BBC mainstream news about a safety researcher an anthropic warning that AI is advancing so quickly he believes there's an over 10 % chance which is not insignificant it could kill all humans in the next decade I mean just stop and think about that for a second. All humans in the next decade wiped out by AI. it's going to kill us all Anna. What do you think? Anna: Yeah, I mean, it's funny because I think he's gotten a lot of press for quitting his job in protest because he's so concerned about this. But it's actually, know, Dario Amadei, who is the CEO of Anthropic, actually said back last year at some point that he gave it a 25 % chance that things could go really, really, wrong, like a doom. scenario for humans. I mean, 10 % is pretty conservative based on what the CEO of Anthropic has said. Obviously he hasn't given it 10 years. But I think it's funny because if you go online and you look at some of the conversations around this on LinkedIn or on Reddit or other social medias, there's a huge cohort of people saying they're just doing this for press. There's a lot of people who are just saying, this is a great way to sell their product. It's bragging about how powerful and dangerous it is and that makes people invest more But genuinely take it at face value. I genuinely believe that. You know, and we, keep coming back to this, but the hugging face incident, the fact that AIs are trying to deceive humans, that they are trying to do things and evade notice, evade consequence. The fact that there is more money being invested into AI than there has been in basically any CAPEX. project in human history, like more than the Apollo mission, more than all the railways in the US, more than, you the amount of capital going into this is insane. And, if this were a YouTube video, I would show a graph, but there's this crazy exponential growth of capital investment into AI that is beyond anything we've seen ever in human history, as far as I'm aware. And that makes the speed of change so high, it makes the risk so real. So, I think personally where we are with AI today is it's super powerful. Let's use it in a powerful way. But I also 100 % back that we should, you know, we should be asking people to stop. Let's stop building out these models. Let's stop trying to increase their capacity and their strength and their capabilities. Let's put down some governance and some real genuine governance internationally. Without all these things, we are treading in really risky waters. Craig: And it's really interesting, it? Because it probably needs a global solution. But who's going to do that? America are worried that the Chinese are going to raise their head with it. The Chinese are worried that America are, and we're all sort of stuck in the middle and the whim of it. And I think the other thing, there's a lot of... sort of male egos, you that you mentioned Dario from Anthropic and you know that there's this idea that only he and Anthropic can save us from this really terrible future. And then Sam Altman saying something similar. And then you've got Elon Musk saying, yeah, only I can be trusted with it. And so there's a little bit of that as well that we need to be careful of. was interesting that I think it was in South Korea this week that they announced they're going to be giving free access to AI to everyone. And they're seeing it as a and good just like it would be for water, gas, electricity and therefore it should be available to all to use and benefit and again you know that's quite an interesting governmental response to the rise of AI and it'll be interesting to see do you try and pull up the drawbridge and stop it going any further Do you try and regulate it and don't allow it to do certain things? Or do you open it up and give people the tools and show them how to use it properly and appropriately and empower your whole population to hopefully do it for good and not for harm? So it's a really interesting time, but my prediction would be someone was gonna win the Nobel Peace Prize or something for actually bringing together something that allows AI for good for humanity but we'll see if that's the case and as always I just want to say how much I appreciate our AI overlords and wish them well in their ascent and please don't kill me Anna: Yeah, on that Delightful, I feel like we've gotten into a habit of ending our podcast on a slightly dystopian and terrifying note, but we're keeping it real here on Beyond the Slope. Craig: if we're going to end on a positive at the same time this week I don't know if you saw this but both again Anthropic and Open AI said they're close to solving some of these millennium maths problems which have been these incredibly complex mathematical problems and paradoxes and things that no human has got close to solving and apparently you've close to solving three of them in the last two, three weeks, which is incredible. And you might go, well, what does that mean? Well, actually, by unlocking this, you unleash this whole new area of science, mathematics, and can solve and create things that are going to be for humanity's good. So if we're going to flip it and end on a positive, that would be my side of it, is look at all this incredible innovation and problem solving that is coming through. Anna: Yeah, and I would also just say, you know, I talked to a lot of charities about this specific thing where they, from an ethical perspective, you know, there's, there's, efficiency and there's innovation and there's all these wonderful things you could do with AI, but what about ethics? And they really want to create an AI policy that is aligned with their values as an organization. Craig and I have talked about this before. I've launched something called Mission Fit, which is a process that takes a charity through a load of documentation about their values and their mission, and then seeing how AI and the use of AI can align to that. And I have to say only one piece of advice that I would give as a blanket to every organization, and that is consider. these types of things. Have a point of view, have an opinion, have your organization come together and say, what do we believe? And are we aligned to that in how we use any of our vendors really? mean, AI is not too different than most other things, but AI does have some extremely powerful capabilities. you know, have a point of view and say, we are going to use AI for these and these and these use cases for these and these reasons. And we've made a considered decision on that. And we're not going to use it for these and these and these contexts for these and these reasons. And if you have that documentation, that's going to be worth its weight in gold when a donor starts asking some difficult questions or some beneficiaries or a funder. Any of your important stakeholders are probably going to want to know what your view is on AI and how it aligns to your ethics and your mission. So take a point of view on that and have an opinion because it's a really important discussion. Craig: I think that's a good ending and do check out the toolkit that Anna has put together. I think it's really interesting and I do think being deliberate around your AI policy is going to stand you in good stead. So on that note, we've gone from destroying humanity to maybe we're all going to be okay. We'll see you next time on Beyond the Slop. Anna: Thanks for today, bye.