Why HR not CTOs Will Lead AI Augmentation - with Joshua Wöhle

0:00I think the entire role of HR is about to become absolutely critical and going from a point where it used to be administrative. Actually, it's not too dissimilar from what has happened to the CTO. If you think about the CTO used to be the IT department. Before technology became critical to the company, you had an IT department and a head of IT. And then companies were built where their entire value proposition was around the ability to leverage technology. And so a CTO became a normal role. And I think the same is going to happen from a people or needs to happen from a people perspective. Hi, I'm Joshua Vola.

0:39I am the co-founder and CEO at Mindstone. We are an AI adoption platform helping some of the biggest organizations in the world adopt AI, specifically training their non-technical people on how this technology can really make a difference in their day-to-day. And I am excited to talk a little bit about how I think HR is the function with the highest leverage in the organization to really get the benefits of all this technology today. And it's not the CTO. I've been so excited to have this conversation with you because I have come to know you over the last year. And one of the things that's really cool about your perspective is you bridge these two worlds.

1:25On the one hand, you work with some of probably the most advanced thinkers about AI that I'm aware of. And on the other hand, you run some of the most foundational, fundamental skills-based training programs. And so I'm really excited to talk about some observations you've had and things like that. But I want to start with kind of personal practices because I think it's a really fascinating approach to take. Your weekly habit of identifying opportunities to automate. I want to start there. Would you walk us through how and why do you do that? Yeah, so I'm a big believer in systems, building systems to basically help myself improve and stuff.

2:14And so the, well, one of the systems is exactly what you just said, which is that every Sunday, I always plan my week ahead. So on Sunday afternoon, I always have to make a decision on what is the next piece of my work that I think is highest leverage to either automate or enhance. And I have a slot reserved for every Friday morning. So Sunday, I decide what I want to go and tackle that week. But Friday morning, I have two hours dedicated every week where I then actually get to spend time to go and do that. I don't have to stick to only those two hours. And if I'm honest, the last three weeks, I probably spent two days a week on this, but it does mean that I have at least hours a week to dedicate to that automation.

3:03It's critical. I mean, blocking the time. I'm sure the stuff you're imagining now is kind of even beyond imagination. Can you tell us about a couple of the early opportunities that you identified, maybe early in your journey and what the impact was of those early automation wins? so probably the one that i come back to oh what i've now been using for more than a year and a half is email feedback um it's been extremely high leverage so a lot everyone talks about how you can use ai to write emails and stuff and i'm actually not the biggest proponent of getting it to write emails but i am a really big fan of getting it to critique emails and so i have a in this case it's a cloud project it used to be a gpt on chat gpt which has been trained with how i want to write emails has been equipped with all of the knowledge of what mindstone does and how we pitch and how we can add value and then it's been set up to look at an entire email thread not just an email but an entire thread where the last or the top email in the thread is a draft reply.

4:16So it's got the full context of mindset and it's got the full context of the conversation. And then it starts critiquing the email, pointing out if I missed something, pointing out if I could reframe something, or if I could add some of the data that it has in order to make it more impactful. And I can genuinely point to hundreds of thousands of dollars of business that we won based on that GPT alone. And when I put that claim, it's literal business that I wouldn't even have addressed because I forgot it was further down the email chain and I had totally forgotten that that was a piece of business we were talking about to begin with.

4:52What are some of the things that you've learned based on this? The biggest thing is that the power of iteration, and I think Jeremy talks about this quite a bit as well, which is the idea that with this technology, it's about identifying a use case that could be high leverage and then being okay with the fact that the first time around you use it, it might not be great. You make it a little bit better and a little bit better. And now I'm at the point where I've been using this one for a year and a half. It's pretty damn good. And what's the form factor? It just sends an email to you after that?

5:29Or what's the, how does it, can you talk us through the flow? Yeah, so right now it's very simple. It's a copy paste flow. So I take the entire thread for my Gmail. I just paste the entire thread into a clod. And then clod gives me feedback on how the email could be better. Now, it doesn't rewrite it, and that is key. It doesn't rewrite the side because then you'd have to go around things of like using the exact wording and stuff like that. I prefer that it just tells me, hey, this section over here misses personalization. Or this section over here, you're missing out on a piece of business that you talked about three threads ago.

6:02I forgot one element, which is that I also have set it up that if I have had calls with the people in question, I just upload the transcript of those calls. And so that's additional context it has. Sometimes it'll actually suggest personalization elements, which is like, hey, you should mention this thing that was mentioned three calls ago, which is a good personalization hook for this email. If I can just go back to your Sunday, I find it to be complicated to do that initial identification of things to kind of engage with. And so could you talk a little bit about what is, you know, what prompts you that Sunday?

6:39do you we've had other guests on the pod that kind of said you know sometimes they'll think of something throughout the week and they'll just make like a note in a trailer board or to do whatever they do and then come back to it i'm curious on the triggers or the methodology to identify high return kind of projects to uh to throw ai at so i do have a similar thing where i just kind of i have ideas that come at various moments and put them aside and then figure out where to go so So I always put them in my task list. So my flow is I have an inbox in my task list. And once a week, I then start triaging my task list to go to different days.

7:18And Sunday is that day for me. Now, as much as I'm a fan of that, I also really do believe in the habit of simply stopping for 15 minutes and thinking, wait, where did all my time go last month? out of everything I did and where all that time went, is there anything kind of big ticket items that I can really think about? Because sometimes like inspiration could happen on small issues, which is like really exciting. But then when I go down and say, okay, how often do I really do this thing? Twice a year? I'm not sure if that's the highest priority issue for me to look at. So just forcing myself through trying to tackle the - Can I just add one there, like little hack that I did?

8:00I was thinking the same way. So I now screenshot my calendar for last week and I just take the image and put it into ChatDBT and I have that kind of do that work for me. That's actually no bad. I wonder if one kind of hack perhaps that I've been thinking about is what's the thing that's keeping you from the work, so to speak? There's one way to think about it, which is what am I spending too much time on, which kind of triggers a certain way of thinking. But I find, especially in organizations, there's a lot of stuff that people feel like is imposed on them that's actually keeping them from doing their job.

8:36Maybe it's the bureaucracy, maybe it's the policies, whatever it might be. And telling folks to look for the things that you feel like are actually keeping you from being able to do your job. A lot of times they go, oh, well, it's this report or it's interfacing with this team or it's always sending this information. I think the interfacing is actually one super important component that people don't use that often. One organization prefers the input to be something different than another part of the organization prefers output. And so you can see like a sales team that wants to have designed a deck.

9:11You know, like the way that they would brief a designer is not necessarily how a designer like to get briefed. And so at Bark, we created like these like communicators that basically interview the salesperson and then, you know, ask them salesperson kind of questions. And then it basically creates a design brief and spits it out in the right format. And so I think that's a super good point, Jeremy. I think. Department to department translation. Okay. So Josh, would I just, just to recap what I just heard from this first, because the bigger question was, what were some early wins in automating your workflows?

9:42and what you just said is, hey, I've built a Claude project to copy paste in an entire thread of a conversation with my draft reply at the top. And basically Claude has been instructed, knowing what you know about this conversation, knowing what you know about our other conversations that I've uploaded and knowing what you know about our business, rate my reply and give me a critique. Right. And you're saying that workflow alone has led to hundreds of thousands of dollars of additional revenue that you wouldn't have. Am I clear on that? You are correct, yeah. Okay, what's another example of an early win that came out of this?

10:18Because I love your comment, believing in the habit of stopping. What else came from the habit of stopping? So from the habit of stopping, so on the early side, it was the movie recommender, a movie and series recommender that came out, which was more on the fun side. So it makes a ton of sense retrospectively, but basically large language models are great with language. And so I think nobody is really happy with Netflix and Amazon and Apple recommendations. It's always, or at least I have yet to find someone who didn't find it very hit miss because they have a whole bunch of other agendas to try and push a particular series or show.

11:01But just using a custom GBT that had, I think 20 or 30 of my favorite movies and TV shows to figure out what is the next Netflix show that I should be watching has been like a hundred percent hit rate. And I'm not even kidding. Like literally insane, by the way. I mean, like I think one of Netflix's call it Coca-Cola formulas is their proprietary algorithm. Supposedly, all you're saying is leveraging an off the shelf LLM and telling it 20 shows you like does a better job. That's that's kind of astounding. It is dramatically better. Like literally you can't even compare the two. Wow. Okay. So one thing I'd love because you have a vision of what's possible that I think is kind of beyond most folks horizon.

11:46You run for folks who don't know, you run one of the largest AI communities in the world and you are seeing and hearing use cases on the edge and the fringe and the kind of bleeding edge of things. one thing i would love for you to do is talk about how you think about call it levels of ai proficiency so to speak meaning i think a lot of folks if they regularly use chi gpt they go oh that's all there is right how do you think about you know uh helping someone understand where they are in their journey and what the kind of levels beyond their awareness might be yeah it's a really really good question and it's an interesting one because it's probably one of the first times that i feel that wherever you are on the ladder you think you're at the top of the ladder which is just a weird situation to be in so you've got the starting point those that still think that it's all hype and that the best choice is to not use it because it's only going to get in your way um you then have people that think that really mastered it because they're no longer using Google, they're using ChatGPT.

12:55And they're using it a lot. They say, oh, yeah, I really get it. I'm no longer using Google. It's like, okay, that's another level. Then you have those that start to make the difference between, okay, well, Google gives you one thing, ChatGPT or AI assistant gives you something entirely different. So you get to the point where you understand that almost by definition, if it was a good Google query, then you should not be using it. in an AI assistant. And if you're using it in an AI assistant, it probably shouldn't be something you send to Google. Once you start making that difference, you start to look at a whole bunch of other use cases.

13:33This often comes with people looking at reasoning tasks. So that's when you're no longer looking at it as an answer engine, but you're looking at it as a kind of a starting point of reasoning through specific cue points. But you still go into, I guess, conversation mode. That is probably the next step. And then the step after that is when you start using it as a proper thinking partner. And the biggest signal for that is when you start to use AI to ask you questions instead of asking questions to the AI and expecting answers. And that's where it now starts to stimulate your own thinking. step after that is when you start to go outside of ai assistance and you start to explore other apps like perplexity notebook lm gamma recently powerpoint copilot starting to do some stuff that's interesting but basically going a little bit outside of the the general ai assistance and started to look at tasks that that touch multiple apps and then the step after that is when you start touching on building stuff which is the what happens when you almost similar to someone who explores and for the first time understands the power of excel and for a very long time like basically excel runs entire companies in some cases well that's where we're getting to now with app building like you don't have to be too technical to be able to put together a general app that just manages all your employee holidays or something like that like that's very very simple to go into and then you've got another three levels behind that xa that are more engineering led no but when you get to building stuff i mean so there's a bunch of stuff here we could break down um and i know that you know building tools are not even the peak but they're another kind of area talk for a second i've heard you use this phrase before josh which i really like which is a utility threshold can you talk about how you define that and how a user can know whether they've crossed the utility threshold so i think that's one of the first kind of sound barriers a user has to break so to speak absolutely and this is this ties to some of my kind of pet themes where i hate people talking about how these models are not getting or how we're getting less out of their progress.

16:02But the idea, you can think about the capability of AI and this AI threshold at the point at which an AI systems becomes useful. It is directly related to its level of intelligence, but it's basically the point where the time it wins you or the quality increase that you get is worth the time you put in. Now, that means that if the model is 98 % of the way there, which means you're putting 100 % of effort, you get 98 % back, it is useless. It can be extremely interesting and it can be mind-blowing from a technology perspective, but it's useless because you've done the thing yourself. So if it creates a great poem, but it takes you, I don't know, everyone has their own utility threshold here.

16:54Me, it would take me weeks to write a great poem. Somebody else would probably be able to do that instantly. The AI, if it takes them more time than it, or if it takes me plus the AI more time than it would take me on my own, not you. But even like an email, right? I mean, an email is a good example for most people. Most lay users aren't, you know, poets, but you go, how long would it have taken me to craft an email I'm proud of? Versus if what you're saying for the utility threshold is if collaborating with AI, doesn't result in a, call it meaningfully improved or meaningfully faster outcome than you on your own, meaning you get something from it, then you got to refine it and work it, whatever.

17:32It has yet to cross the utility threshold and it remains a toy. Yeah. Now, and what is interesting is once you cross the utility threshold, a tiny increase in either your ability to wield the model or the model itself yields dramatic increase in utility. because now imagine that you're getting initially get 102 % out and it takes you 100, right? So you take 10 minutes and the AI does it in whatever, nine minutes, 58, whatever. So you get the two seconds. Now the model gets a little bit better and it does it in nine minutes, right? So actually what has happened is you just won that two seconds, but times 50 in this case, right?

18:15So So you're really, you're getting a dramatic uplift with a small increase in the model. And this is the bit that many people are currently, especially technologists, are just not getting in the conversation of how this technology relates to actual productivity, which is, everyone talks about these benchmarks and how the AI is getting only slightly better on a particular benchmark. A 2 % uplift on a particular benchmark might mean that there are a thousand different use cases that just went from being useless to useful. And so it's unlocking a whole bunch of other use cases, which then in turn gets those people to the other side of the utility threshold.

18:53Then they start to use it more. And when they start to use it more, their own proficiency becomes higher, which also has this effect on the utility threshold, right? Because the easier it is for me to use the model, the less effort it means I need to put, the more utility I get out. You know, there's a phrase that someone used on the podcast. We had Bryce Shalemel, who's the head of AI at Moderna. And one thing he said, which has stuck with me, and I think resonated with a lot of our audiences, I can't imagine doing any part of my job without layers of AI baked into it. I think he said something like that.

19:25I can't imagine. And then he went on to say things like, it'd be so lazy and stupid and reckless, which I love the kind of impact of that. But I think speaking of the utility threshold, you have to get to the point for a particular, set aside all of your work? Is there any part of your work for which you as a knowledge worker could say, I can't imagine doing it without AI? And I think for a lot of people, probably the truth is, oh yeah, I can kind of go back to the normal way. But there are, I would say for myself, there are foundational work products that I go, it would set me back a long time if I no longer had the ability to collaborate with AI.

20:07And I think that to me is, I don't know if that's evidence of utility threshold, but it may be like a mile marker, so to speak on the journey. How many of those things can you point to that you can't imagine doing without AI, right? So honestly, at this point, from a almost, if not all of my work, no, I do live delivery, but my live delivery is showcasing AI. So I could argue every single aspect of my job is knowledge work, in which case I don't think there is a single part of my job that doesn't get dramatically enhanced with AI. To the point now, so the last few years have been interesting where we went, I don't know, when was it, eight years ago or something like that, through the whole wave of getting Wi-Fi on planes.

20:56And for a bit, extremely good. It worked very well, started to be able to do some emails, and you were kind of in between because you were hoping, okay, you get some internet, there's an additional productivity boost you go through. Now, when I don't have Wi-Fi on a flight, even the thinking work feels like it's wasted work because I know I'm going to have to go through the exact same steps again because I don't want to miss out of the input that the AI would have given in the thinking work. The boost. Yeah. It's like, do I want a 10x think through this or do I want a 1x think through it? Do you have anything, because I think that's very true on a lot of different dimensions.

21:36The thing that I've been increasingly thoughtful on is what is it that where I have to force myself to kind of like lean into my humanity? Like, what would you say is the thing that, you know, even if you can create a bot that could do that, you would really want to do that yourself? from a work perspective i don't see it it would be on the at a personal level um it would be about the relationships i craft with the people i care about around me i guess that would be the only and even there i would still think there was room for ai assistance in like making sure i don't forget a birthday making sure i don't forget mother's day but i would want for myself for that relationship to build for there to be a real connection and that would that can only come through me one might call it relationship capital no kidding it's only because it's a core theme of a book that i have coming out so like nice set up there but it is fascinating right like the people like you who's such a super user increasingly obviously would be completely deprived it was gone right so therefore like that's obviously you know interesting but also increasingly have a little bit of a tough time seeing what is it that we as humans are completely unique to to kind of do except our ability to emotionally connect with other humans yes and now i think we're seeing with the rise of mates and chat bots you know set aside i mean even romantic partners.

23:17But I think for a lot of people, there's a question of whether the, whether that's a uniquely human ability. Uh, that's not my personal feeling, but I would say, as I observed the world around, there's an increasing number of people who go, you know, I saw in the paper the other day, I, yes, I read the paper. It was like sitting on a hotel desk. It was like, somebody's marrying a chat bot. It was like, and so you start to see there's a lot of, I mean, think about someone who's always available, right? Like your human partner has to sleep, for example. But a jab bot could be available to have deep emotional conversation, right?

23:52Right. I think you touch on something interesting here because for me, there are two different components. So for me, the component that you feel, your need, my need might be met in different ways. My need to feel heard in a way or to feel like I understand myself or to have an outlet where I can out my frustrations, whatever that is, I can imagine that need being met in some cases, maybe not fully, but in some cases to a degree by an AI. But that is not the same as the need that I have to express myself to somebody else. We talked a little bit about it in an earlier part where, you know, you can get your GPT to count your calories, but for some reason, it feels better to send it to your trainer because there's a need for connectivity.

24:45If I can just, you know, speaking of need, like you obviously are so good of kind of doing all these tools. And I think for a lot of us who kind of like are in this world, like that's a little bit like CatNet first. You talk, I saw on your LinkedIn about making all people like 40 % more productive. And I think a lot of people who listen to the podcast, they sit with a role where it's their job either because they're self-selected or because their company have asked them to to kind of like help the organization go through that. What is some of your current advice when somebody goes like, hey, I have an organization, people are kind of curious, but like it's being used, but only a little bit.

25:22What's kind of like the way to kind of kick it in the butt? So there's two things. One is the simple realization that no technology has ever adopted itself. So it's very weird that we're expecting this one to somehow be that. And that is exactly why it's so treacherous because it appears we get back to these different levels of knowledge and you always think you're at the top of the ladder. And so it's so easy to get going, but it's also very easy to fool yourself to think that you now know how to use it. And then the second bit is show, don't tell. I am genuinely very sick of the amount of people that talk about all of this stuff And that honestly, when you ask them right after, OK, so how did you use AI in the last whatever, two, three days?

26:14And they can't come up with a single example. Right. Right. Or they share their screen and you look at their history and it's like there's either there's three conversations in total there. Or, you know, it says last 30 days, like there's nothing today. There's nothing last seven days. Right. Yeah. that's exactly it but um so so to get back to that is when you get the chance and it takes maybe we've gotten better at this over time so it takes maybe 30 minutes i'd say where i'm i can take any skeptic in the world probably sit down and get them to the other side to realize okay shit this is real and i really need to do something about it now but it requires a live demo i don't i don't know of any other tool to be able to get someone past that gap.

27:00I think that, and I think also like there's increasingly like a gap, knowledge gap, right between people who know a lot about it and the people who kind of like haven't really gone into it. I think what we find useful is to just kind of not personify it, but just think about it as agents. I think people are so used to thinking about colleagues that if you start to go, hey, there's like three types of organizational basically agents that we can make we can do one for the organization so like your hr kind of bot so that if people want to you're talking about like travel days you know it could be you know what's our rule for sick leave um there's like i would say very department focused one so uh you have a supply chain team they get a bunch of invoices in they'd like to get it in like a structured format and so they don't sit and then there's like the hyper individualized one where you have for example i'd like to read a lot of newsletters but i often don't get to read them all and then i feel frustrated and so now i have like a script that kind of like summarize them for me um is that a form that you use or do you have another way for people to kind of like get an easy mental model for how to think about it i think that is indeed um it's that way of thinking about it is actually very traditional the way that i hear that is every example you just mentioned is basically we could meet that need through building traditional technology.

28:26As a software engineer before generative AI, I would have been able to build a solution that automates that. What doesn't get talked enough about, and I am also not doing a good enough job at it, is the ability to dramatically expand our own thinking power using these tools. So the use cases that I think are much more powerful are helping me brainstorm solutions to a particular problem, helping me understand a particular problem from different facets, helping me understand how different stakeholders might respond to a particular situation ahead of time, helping me develop the strategy for Mindstone for 2025 and have AI as an actual thinking partner, critiquing what I'm doing, suggesting where I might want to go and improving various aspects here and there.

29:24And it also paints this picture of two different worlds. And I actually think this is probably one of the biggest risks we have, which is that AI has always been about automation up until this point. We very rarely, I know it sometimes gets soft about, but we rarely talk about augmentation. And the problem is that often when people don't understand what is happening at the moment, the responsibility of AI often lands with the CTO or the CIO because it's technology. And then a CTO or a CIO is often very, they're great in operating in a deterministic world. That is literally why they're great at their job, because that is what technology required you to be able to do, to extremely rationally be able to take a very complex system and derive a set of rules that will get you to an outcome that automates that process.

30:22That is the skill set you need to be a great technologist. But if you then take this technology, whose power is specifically at non-deterministic use cases, but you put that in the seat of the person who's great at all the deterministic scenarios, they end up using it for all the same scenarios, which is that they end up using it for automation. And so the result of this means that the organization ends up automating more rather than if, and this is one of the biggest battles that I'm trying to get to, this should be in the hands of a progressive HR team who are like, hell yeah, we are going to equip everyone in our company with the ability to do twice as much by the end of this year.

31:08And then you end up with an augmented workforce, which then means we are all reaping more of the benefits. And I really think that it might well be that difference. And if we don't get there, we just end up with a whole bunch of unemployed people. And if we do, we end up with an enriched society that actually is able to do that. I think that's such an astute point. But what I find complicated is in launch organization, and maybe we can have you talk a little bit about how you built big kind of teams before, and I assume now building them in a new way. So let's kind of like end there. But I think you're right.

31:46And obviously we see some of the new, so the Ethan Mollick kind of like research which come out of basically how if you allow people to spill into other functions and their core kind of like function, which I think is a little bit of what you're talking to. A lot of teams, of course, have a tough time doing that because while AI will allow an engineer to suddenly think as a marketeer or a marketeer to kind of code some things that they can make it actionable, you have structures that makes that complicated. Do you have a view on how this HR manager, which I think you're totally right on, how can they navigate the structures that they kind of work within?