The Human Side of AI: Agentic AI & The Future of Workflows

Hello and welcome to another marvelous episode of 42 Causes podcast with Sir Lord Connor McCarthy live from uh Dublin, the largest city in the world because it keeps Dublin in Dublin. Ah, [laughter] >> I haven't heard that one in well over 10 years. [laughter] >> But um welcome. It's so lovely to chat with you again. Um, >> yeah. Yeah. Thank you. It's great to be here. >> For those of you who don't know, Connor is uh is one of the world's leading experts in in AI and uh helped us with our AI course and and travels all over the world telling people how they should be using AI and as well as building things himself. So yeah, it's a sort of a marvelous chap and obviously in most of our lives AI is affecting everything. So I thought today one thing that I that I think is going to be probably the biggest change this year in business is that I think we will see lots of the bigger companies switch from trying to master just you know chat GPT and Gemini and Claude um into trying to really understand and master aentic AI. Um, >> when I chat to to different people, I think most people are probably still stuck on trying to figure out how to best use chat GPT and and Claude and those kind of things just in as a normal chatbot, let alone Agentic AI. So, I thought since Connor knows lots about this and I've been writing a bit about it recently, I thought we could have a conversation about that. But, um, yeah, I'll leave most of this to you, Connor, because you probably know way more than me and you also see what people are doing. No, like with AI, the rule of thumb is if you think you know something, you probably your knowledge is going to change in the next probably 36 hours. So >> I still take great uh solace from Andre Karpathy saying that even he feels behind. So uh and he you know he was he invented vibe coding essentially. So uh right good news for the rest of us. But yeah, Gent. Yeah, it's it's God even that is weird, isn't it? It's like I feel like only a year ago it it came onto the scene and now it's kind of everywhere. But it's not, you know, it's not buzzwordy in the sense that Claudebot this week was super hype buzzword >> and is I think is slowly kind of going away. Agentic is here to stay. I think I think an understanding of the fundamentals which we can maybe cover in this call. >> It's just good to for people to get their heads around. It's um yeah because it is it is different. I actually I'll just maybe I'll just give something we can we can talk about like kind of three three levels of interaction you might say with AIS that that I see um and you know people think actually I'll just I'll I'll just kick it straight in. So, let's say there's LLMs, there's uh, you know, the usual chatbot experience, back and forth, back and forth. Nothing happens unless you go in and start asking questions, right, >> and getting the responses and then taking a direction and going on from there. That's most people's interaction with an AI tool at the moment. And that's certainly most of my interactions. Um, and that's great. Obviously, you need to be there to do any kind of uh you need your agency to use the tool effectively. [snorts] >> Then there's there's a kind of next level up which is not agentic. Not yet. It's workflows. Now, I I add this because I think a lot of people kind of go, "Oh, it's either talking to the chatbot or it's agentic." There is a middle tier of just workflows. Now, workflows are interesting because let's say tools that people might have heard of like Zapier or uh N8 or things like that. Workflows basically let you string together a number of um of actions um based on decisions that you decide to make within that workflow to get you from point A to point B. >> Right? >> So very very simple workflows could be you know um when someone enters something in a row on a spreadsheet, you know, if that the thing they enter is a uh customer name, then you know, put that send an email to that customer. >> So, it's it's it's very much if then if then uh that's a kind of standard loop. >> And you can get really fancy with that workflow. I I had a client come to me and say, "Oh, I need an agentic thing." It turns out they needed just a really robust workflow. There was nothing special about it. They just needed to make some decisions about what was happening when, >> right? >> Okay. So, very kind of log logically minded, you might say. Um, if if the LLM, if the first case, if that's the brain in the jar, um, the this next case is kind of it's starting to get limbs, starting to do things. >> Yeah. >> Then aentic, which is what we really want to talk about. Um, this falls, I think, under the the the label of decisions. And that's really what this comes down to. Okay? Because with agentic um usage, the decision maker changes. So in in cases one and two, you're still the decision maker. You're deciding all the things that happen. But in case three, agentic workflows, the model is deciding and you know nothing about that. >> Okay? Okay, >> a very a very quick rule of thumb that I saw a kind of well- reggarded engineer say online, he said a aentic workflows are essentially LLMs running in a loop that are able to use tools. That that's it in a nutshell. And that's very much true for for most use cases. Um I'll stop there because I've been talking. No, I mean I was I was going to say it's I mean as I'm exadvertising a lot of my friends are still in there and and I think the way they kind of when I ask for updates on them and how they're doing, how the different agencies are doing, I see these kind of you they they seem to be having lots of conversations about aentic workflows and how they're going to use them internally to help things run smoother, more efficiently, faster. I wondered what are kind of the you know what are the most interesting use cases you've seen on it and how how prevalent do you think this would be like is it and is this is this kind of that you talked about um Claudebot earlier the the thing that I can't remember what they've renamed it to >> um the the that that was kind of it's almost like a very unsafe but first iteration of essentially like Jarvis from you know the the Matrix films where you've got sort of a a true I mean I wish this was already possible but yeah I mean it kind of is I guess yeah like a a true assistant who sees what's going on in in your day-to-day life and then goes hey like I see you've got a podcast coming up in you know later tomorrow do you want me to send you some notes notes that might be helpful for you so that you've got some prep notes for it. And I've I've not had to do anything. It's just done that for me. It's been proactive. Or, hey, I saw there was an email here that needed an urgent reply. I know you're busy, so I sent them an email saying like, you're just in another meeting and you'll get back to them in 20 minutes. Like it's it's all this stuff that that that you you know just kind of almost like having a really smart, you know, real life assistant and they just do stuff and they're seeing like what's coming into your inbox all the time and sorting stuff out. Normally only billionaires would have access to something like that. Now we all come in. Does that is that is that still an agentic thing then or is it >> Yeah, because when like again your sniff test is like is this thing making decisions without me? >> Right. Okay. >> So in those so um Claudebot does do that the horror stories about Claudebot and maybe their their um their good father in this conversation because you know I have one friend who installed Claudebot and gave it access to his WhatsApp and woke up the next day and I had sent WhatsApp messages to people in his contact list about something random that he was doing that day on his calendar. So obviously he didn't want that to happen a and it was quite embarrassing but Claudebot made decisions on his behalf. right? For some obscure reason. Um and this is the this is the um the thing about agents is you are handing over your decision-m small and large to these agents. >> And that opens up a whole world of you know security, access, autonomy, privacy, all that kind of stuff. Because if Claudebot had that guy's credit card details, Clawbot might have gone and shared those or bought something or whatever. It would have started making decisions and that's that's where things get a little bit sticky. >> Yeah. Because I think a lot of people I mean I I've had this this recently a few times where Gemini or Claude will will say, "Hey, you can now link this to your email, to your Google Drive. if you want to have access to all these other tools like Slack and um I'm always rather hesitant at the moment um with my main, you know, if I had a burner laptop with with really bad information, I'd probably do it. But I I'm I've done it for some things where I know up front it's like, you know, we'll never I think with with Gemini, they send tend to be quite um overly cautious on it. So they're like, you know, we we'll have access to it, but we'll only use Drive if you tell us to go and use Drive and blah blah blah. So with that kind of instance, I'm kind of okay with it. But >> yeah, it's it's for Aentic AI to really work. You kind of need that. I guess for businesses, it's different though because you're just trying to create is it plugins or sort of link them to different tools that you exist that you have existingly and and Yeah. >> Yeah. Yeah. Yeah. So, so the tools is obviously important like I don't know maybe more maybe an example that would suit this conversation is of an agentic workflow >> would be something that you know >> if you're running ads running Facebook ads Google ads what have you analyzes >> the ad performance from the last month finds patterns in all that data AI is brilliant at doing thinks up new angles based on the best performing ads tests those ads and either launches them or pings you to say, "Hey, here are the three what what I think are the best three performing ads." >> Right? >> So there's little you could say in there like if you gave it full autonomy, it would go and just publish those ads and and you would not know anything about it. So it's making all those little decisions or sorry that big decision at the end I guess. >> Mhm. But similarly the decision which again is more of a workflow thing as in right when you get to this point send me an email. >> You know >> I mean that that's a lovely example when when you're saying that like I mean how >> right now like how how easy or hard is that to build? >> It's it's pretty straightforward. It's one of those like uh logically it's very straightforward like you'd have to kind of get in and there's a lot of testing of these things to make sure that you have the right guard rails up etc. And I should highlight like you know what's the most important part in that. Okay, finding last month's ad ads that's easy. It's a database call. Um finding patterns in those ads that's relatively straightforward and again AIS are really good at doing that. Coming up with ideas for new ads based on those patterns like that's also like fairly straightforward for an LLM at this point. It's the kind of um it's the decision in there of like Well, should this ad go out? Like, has has it made the right call on these three ads? The human judgment factor, like maybe you want it to kind of ping you and say, "Hey, these are the three ads." And you as the as the account exec or what have you, you might be able to kind of go, "Oh, no. We shouldn't publish this ad because this thing just happened in the news and this would reflect badly, you know, on us because, you know, it's not very sensitive to what just happened in the news." So that's that's an example of a guardrail and you could probably build against that. You could probably kind of go you know what just check the news in the last seven days since sense check these three ads to see you know will anything you know cause a problem here. >> Um but sorry your original question that yeah I pretty sure um those marketing automations there's a zillion of them now. I'm I'm assuming that exists somewhere right now. >> Yeah. I mean, when you're chatting with clients and trying to get them to think about, what things they might want to put an agent an agentic workflow together for? What kind of things do you suggest that people think about um so that they find the you know the the golden stuff rather than just you know I've got this idea let's try it because I I I have heard also there's tons of data showing that these big companies have been wasting trillions of dollars on on AI experiments most of them have been an abject failure um maybe they tried too hard too early >> um there seems to be this course correction on yes let's keep experimenting but let's fine tune tune those experiments so that we only do the stuff that we really need to do. What kind of recommendations do you suggest when you're chatting with people and how to think about, you know, what are the key things that are going to really make a difference? Um, start like start small with something that will make like a tangible difference. So, like one thing that I've done a few times for people because I know it makes such a difference is like an inbox triage. like just something as simple as that where you can hand off your your um your email inbox to a simple agentic workflow. And even if it does something like there's various levels. Let's say you just wanted it to label your emails in a certain way. >> But emails can be tricky because you know yourself there's everything from spam through to super important emails. >> Yeah. >> And you can't you can't really if then those emails, >> right? There has to be some kind of decision on behalf of the LLM to kind of go, you know what, this the voice in this email sounds urgent and it seems to be, you know, from a client, so this should be, you know, top priority, you know, ping ping Connor, ping Chris, say you really need to check this out now. So like that that sounds really kind of boring or whatever. But I think when I've done that for people and they see like oh my god suddenly their email is manageable again or suddenly their life is better because they don't have to kind of >> worry about things they're things that are sipping through the net or all that and they're saving hours every week. Then they kind of go, "Oh, okay. Right. So what else can we do here?" And then then you get into uh maybe some SAS platforms that they would use that we can kind of hook together, hook into something agentic um and kind of start to figure out better ways to um just help save time or just save headache in a lot of cases. >> I guess I had like two questions based on what you said because I'd imag I'd imagine you're absolutely right. I mean, email just takes up so much of of most people's time, especially now with >> AI sending emails. I mean, I I I don't know what happened, but there was last I think it was last week or the week before I got in one day probably 10, you know, very personalized emails, but obviously AI written >> saying like, oh, you know, don't you want to have this person on your podcast or like >> and they were all podcast ones and I was like, how who's got my email and why are you sending me 5,000 versions of basically the same thing from different random people I've never heard of in my life. >> Um, so I guess I guess the the but yeah, sorting out your email is a good thing. I guess the two questions I had was one is like is there an easy sort of go-to tool that that that's, you know, not too complicated to use for that? Um, and the second question I've totally forgotten, but I'll remember in a minute. [laughter] That's okay. Um, one and the the the I'm finding with AI the the recency bias is very powerful because there's so many tools and ideas floating around that I'm like >> I'm sure I've seen a hundred of these before, but the one that springs to mind is lindy.ai. >> I'm not a reseller. There's no there's no referral fee, but Lindy. I heard someone on a podcast say um that they started using it and it, you know, changed their life essentially. Um you can it's basically uh easy to use automations. Um everything from customer support, email management, lead genen, um LinkedIn outreach, it it kind of does everything. Now, you know, you got to get in there and use these tools. And you got to be really careful again with agentic stuff like how much do you want certain tasks taken off your plate entirely safe ones >> versus how how much do you want to step in and you really need to kind of get into the nitty-gritty. um like a lot of the a lot of the talk about agentic workflows and I think one of the reasons that they're so popular is that in the in the engineering world in the software development world they've been and the they're doing like unbelievable amounts of work good work for engineers engineering like it's it's usually the canary in the gold mine for a lot of AI stuff for obvious reasons but it's um engineering is kind of a kind environment as well like it's code it either it works or it doesn't. You know, you can generate tons of it, but you can also test like does this work? Is this any good? >> Right? >> So, agentic workflows in engineering have really taken off and now if they're to be believed like you read stories of um engineers just literally overnight um you know a customer request would come in for some feature to be added, be picked up by an agent. it would, you know, create a a pull request in GitHub. D would, you know, write the code, write a series of tests to test the code. If you're really brave, it would post the code and and you're done. Um, that kind of stuff is happening. Like people are literally opening up their email in the morning kind of going, "Okay, it took that problem and solved it and there's actually nothing for me to do here." >> Yeah. >> Like that's that's happening now with with certain types of software. I I remember reading an article yesterday. There was a I think it was actually even a tweet last week from the CTO of of um of of of Claude um saying that they they don't they actually he's not written a line of code. >> Oh yeah. >> The whole year. So so far this year he's not written any code. He's just been using claw code to do all of it and then since checkeing it >> 100%. And and just the the caveat with this and with all things like someone in the comments I saw that tweet and someone was like like you wrote everything using flow code and he's like no just the code the architecture the security the tech like he named five or six things that he was like obviously they were done by humans or assisted >> you know AI AI assisted with humans and you know code is is just words on a page there's so much more that goes into developing goods software. >> Yeah. And I remember my second question, the you because we're talking about the Genti and you've got, you know, it has to then talk to so many different services. Um, it made me think of of a of a of a similar sort of thing where in in in smart homes, if anyone is a is a gadget nerd a bit like myself, in smart homes, one of the things that was always infuriating is everyone used a different standard. So, >> you know, if you had Philips light bulbs, they would only talk to other Philips products, but maybe your Google or Apple Home or Amazon wouldn't Alexa wouldn't talk to certain things. So, you ended up having to always have a million apps. And finally, I think it took them 10 years, they figured out, okay, well, we use matter as the kind of uh the the the common ground uh base that every everyone everyone interacts with. Is there a similar one for Aentic AI that that that people are using? I heard I think there's one that's been done by Anthropic and or champion mainly by Anthropic and and Google. I couldn't um yeah uh I'm also a smart home nerd. The um there is well Anthropic released they they created and and generously kind of released for free this this idea of MCP servers. >> So yeah, model context protocol. Such a clunky name. Oh my god. Model context protocol and you're called everything. But yeah, that that's um that's a way of giving AI tool or giving LLMs access to um platforms and services outside of itself. Again, it's got it's giving you more arms. >> Yeah. >> Um so it's similar. It's like superpowered APIs. I don't know if you know what APIs are that >> Yeah. >> I mean, you can you can you can share a belief explanation. anyone who doesn't doesn't >> that's a programming interface. It was essentially a way for systems to talk to each other. Um so uh so you're able to kind of ping a silo of let's say all your information is in Salesforce. It would have an API and I could write a piece of software that I could ask the API hey can you get me all the customer data and it will return that customer data and there's a whole authentication thing going on. So that was the simple back and forth let's say remember earlier the back and forth LLM thing APIs were v1 >> right >> then MCPs were released and suddenly it was more of a fully featured conversation where through um through claude let's say literally in claude um I could I could add the Salesforce MCP server and now I can have an English language conversation with all of my Salesforce data >> and and similarly a clawed agent could do the name and it brought this just really rich really powerful interaction where suddenly you could chain together um you know uh something like a a Salesforce MCP with your Google Drive with like looker with all these different services and and it's it's a proper conversation between these tools you know >> it's um there's another kind of underlying layer that all the big companies have kind of come up with to try and system it to make sure there isn't the same uh Alexa theory problem. Uh I think Google Google called it a agent to agent and that's a more of a kind of underlying platform, but I won't get into that now. [laughter] >> Yeah, it's uh I mean I know I know you've got to dash off in a minute, but um >> I mean what what's your is the end of January flown by in a flash. What's um at the moment what's the the tool that you're using the most? that's your your AI dayto-day like go-to go-to things. Is there anything that's that that's been standing out to you recently that you think people should have a go with or >> um I'm using Clawude Code a lot and I mentioned this last time. It's and I think you were playing with it too. Um >> yeah. >> Yeah, >> it's I really it's a funny one to recommend because it means using the terminal. I've started saying to people, you know what, it's just another chat box and we're using chat boxes anyway and let let's just go for it. It's extremely extremely powerful because um you know now you've got Claude's brain on your computer and it can do all the things on the internet that Claude can do you know through a browser but now it also has access to local files and local services. It's really really powerful when you get into it. It also has a lot of tools at its disposal. skills and as you said MCP servers and everything. I don't know. I I I would just implore anyone who's who's looking for like you know what's something really powerful that I'm missing. I think it's clawed code at the moment anyway. Um it's it's it's very safe. It's very easy to use. Um yeah, I've just got a million situations where it's like, h this isn't working with chat. Hey Claude, can you blah blah blah blah blah, you know, connect to my thing and do this and make sure this is working and it it's done. It's fixed and it it blows my mind all the time. >> So you're not using it just to to make websites. You're using it for sort of >> Oh, like all these AI tools once you realize, oh, it can do that. Oh, maybe I can use it for this over here. So, it's everything from, oh god, building pieces of software for myself, um, to fixing problems I was having in notion with really complicated formulas, uh, doing pieces of research. Um, I'm creating more and more skills to help me just kind of do marketing for my consultancy, that kind of stuff. Um, yeah, my my ultimate goal is maybe to create uh that the the kind of what Claudebot should have been like the the personal assistant, right? But using Claude so at every step I can kind of manage what's going on. >> Yeah, that sounds brilliant. Um, look, I know I know uh I know we ran out of time already. Time replies. Um, >> sorry. I I just I just blabbed the entire time, but thank you for >> Amazing. It's super super helpful. So, thank you so so much. Um, and also good timing because I can hear uh >> I can hear our little >> um Yeah, I hope this was useful. God, yeah, it's um it's it's interesting times, but we should we could do another session on Agentic if if you're still kind of going, but what about this? >> Yeah, it'd be interesting to to maybe try and actually have a have a little mini project and and show people what it's like to to build something. Um >> probably more uh YouTube friendly than podcast friendly, but uh we'll make make a plan. Um but yeah, we're um yeah, we'll uh love you and leave you and thank you so much for for your time, Connor. I really really appreciate it. You're amazing. >> Thank you everyone for listening. Yeah. >> Yeah, we've just updated the agentics section in our in our uh in our AI for marketers course as well as well as adding new lessons on uh on GEO, which is AI search and uh and all sorts of other bits and bobs. So, uh, yeah, if you if you want to have a little, uh, easy to understand read up on this, go have a look at the course. Um, yeah, >> loads of love and can't wait to chat soon. And thank you so much for the time. You're amazing. >> Thanks, Chris. >> Talk to you soon. >> Thanks. >> We hope you enjoyed listening to this 42 Courses podcast. If you did, please like and share. Uh, any comments are also very, very welcome. And of course, if you want to learn more about us, visit our website at 42courses.com. Thanks again.