Bad setup kills good AI ⚙️ Ben Tasker

even if you are afraid of that replacement, it's better to learn this technology while you can because it's paying a premium and all jobs are gonna be impacted by AI. Every job's gonna be augmented by AI. So it's better to learn those AI skills in combination with human skills. And then any amount of change happens, you'll have two very distinct skill sets that you can lean on and can pivot through the change. And I think a lot of individuals just think that this is another, software implementation that they can just put their head in the sand and hope it goes away. AI is getting better about every 2 days. By the time this podcast is through, don't think it's going away. There's been billions of dollars pumped into it.

This week's Millennial Master is Ben Tasker, an AI product leader and workforce transformation expert. Ben works at the point where a lot of companies get stuck, not buying AI tools, but but actually getting people to use them properly. His background spans data science, healthcare, education, and product, and he now helps large organizations upskill and reskill their teams for the AI era. In this episode, we cover why most AI projects fall inside companies, what founders still get wrong about AI at work, and why upskilling matters more than buying another tool. Ben, welcome to Millennial Masters. Where are you joining us from today?

Hey Daniel, thanks for having me. Uh, I'm from New Hampshire.

Well, we're back across the pond, but AI is everywhere. And I wanted to ask you, Ben, before we get into frameworks and to the data science and everything in between, take me back to your first real interaction with AI and maybe like the moment you clocked it was going to change work and everything else in between.

So my first interaction with AI was actually not specifically AI, it was with data, which is the oil to AI. But I graduated with my undergraduate degree many years ago and I thought I wanted to run a hospital. So it was hospital administration that I earned that degree in. And my first task was to collect data for doctors' offices. And it was a very human problem, but we had to figure out the volumes of those doctors' offices. It could impact scheduling, it could impact the services that the hospital system provides. And I was like, ah, no problem. This, this is an easy one, what an easy first project. I was solely mistaken. Some of the data wasn't in databases. It was still written on paper. Some of the physicians estimated what they were doing. It was, it was a very interesting first project. Not, not as easy as, as I thought, but that really then made me start to think about data-driven decision-making. At the same time, master's degrees around data science, data analytics, data this, data that. I decided to do data science and data analytics because that's the background of artificial intelligence, the algorithms. I love the data. I knew a little bit of code, but I, I wouldn't say I was a computer programmer, but that program would teach me that. It would teach me the algorithms, and I knew that if organizations at that time period were just getting into data, the next logical step was data science, which is now AI. So that's really when I first interacted with AI and artificial intelligence, with the data, because I saw the disconnect between what the organization thought was happening and what was really happening.

But how about the moment when, say, ChatGPT was launched? That's when everybody got a taste of what interacting with AI and thinking how it can be implemented in a business and in day-to-day life was really that point. What did you think of it when it became so easy to use in a chat interface?

So I was a data scientist once I had my master's. That's obviously what I went to school for. I enjoyed that. And a lot of individuals have a different conception of AI and data science. Some of it's magic to some folks. It's very interesting what they think data modeling is. OpenAI launched ChatGPT. It was, I mean, I still think this, I know I'm a more innovative type. I still think we're in a change mode, but it completely changed the way individuals view data, AI, how to interact with this. I think they saw usually data science is behind the room. I think conceptually the lights turned on and everybody was in the same room. They realized the potential of it, which was great as a data scientist because now it was less, less fictional to describe some of the stuff. That a data scientist can do for an organization with the proper data. Now you can see it. And it's more than just chat that this, that that tool can do. 5 hours, I think ChatGPT gained a million users. There's no other product on the planet in current time that's gained that much traction ever. So I think that speaks volumes to what I just said.

And how did you first use it in your first use case when, when you got your hands on it? I'm assuming you were one of those first million people who use it in the first 5 minutes?

I was, it was very exciting for me. I knew the world was going to see the potential. There's some risks associated with it, obviously, but I wanted to make sure I could trust it because it was, it's a new company, well, a newer company. They were around before this, but first time a lot of folks saw a company, how are they collecting this data? How is my data going to be stored? They didn't really have the paid tier at that time. Everything was free. So my first interaction was actually doing a fitness plan, taking photos of equipment. At my house and having it plan a workout for me. I can honestly tell you I haven't planned a workout since, since we launched in 2022. I don't know. Huge benefit. Food. You can go take a picture of your refrigerator with all this, all the ingredients and everything that will make a recipe based upon a diet that you want to try. Super beneficial. But if you get that information right, like I used to do, I still do CrossFit, but which is a prescribed workout, a high intensity, but now that AI knows me a little bit better and I take those photos, it's, it's phenomenal. It's very cross— I wouldn't say it's full CrossFit, but it's CrossFit aligned.

You brought up a good point, trust, because it's already well known that ChatGPT and a lot of the models out there have been trained on the entirety of the internet. They scraped everything they could find. We don't know exactly what. Or how much, but we have an idea that everything that's publicly available has been scraped. And how could— can you trust a tool that you don't know what it's been trained on?

Well, I think you have a good point there too. I mean, at the time it launched, we didn't know how it was being trained, which is why I was a little bit more cautious with it. But now we do know it's an input-output model. Any chatbot or natural language process model is an input-output model. Models, anything you put into it potentially could be trained to enhance the output. That's how they get more intelligent. So, but when it's free and it's open to everybody and it's more accessible, that's a little skeptic, but I wanted to make sure that what I was putting into it was accurate. It became personalized. It wasn't super high risk. It was photos of work of equipment, right? Like that's not gonna— it could have impact, but it's a low impact, right? Like lower risk. Today still, I probably wouldn't have it do my taxes, right? I like, I wouldn't put in private information about myself. I scrub off my name. I give it a persona about myself, so I describe what I want out of it instead of just giving it all the information. So those are, those risks are probably even more relevant today than it was back then because we know that's how the models are being trained. But we really have to, I like to use a tool before I judge it. So that was when we get used to interacting So that's interesting.

All these years in, you still don't trust AI enough to give it your name or personal or financial information about yourself.

Not just the AI, it's the organizations too. Like they have masses, mass amount of data. I don't know what they're doing with the data. It's not super well regulated. Company could get that data. And even if the AI companies have really good intent behind it, some other company may not. So for me, it's use the tool to enhance your capabilities. You can use the tool to automate some processes that you don't really enjoy. It's fantastic at that. You can use it to have fun. Everything's fun. There's a lot of information on it, but you also have to be careful, right? Like, I don't want to overshare information just like any other tool, like same with my social media and stuff like that.

So for an AI expert and veteran in your case, you don't trust the tools enough to give them personal information, but a lot of businesses across the world are relying on AI to a certain extent. It's the smaller they are, possibly the more they rely on it in order to direct and assist their business. So for all the founders listening and putting all their information into one of these models, whether it's ChatGPT, whether it's Claude, is their information at risk?

I think so. AI gives non-deterministic results, which means the result, if I prompt, a new prompt, we'll get similar but different results. I think it's an excellent exercise to see because even though with the difference, it's really not that much different. Here's a great example that listeners can try. It's kind of a magic trick. I think it will still work. I tried it this morning, but generate a random number 1 through 10. Put that directly into any AI chat tool. Generate a number, a random number 1 through 10. And it should give you the number 7. So it's not selecting a random number. I've gotten that 99% of the time. The next number that it selects is a 5, followed by a 3. That's because it's the most popular information that it has stored. Uh, 7, 7 days in a week, Bible text references 7, lucky number 7, 5, I just think it's close to 7, and then 3, right? Like, it kind of makes sense.

One of the The magic numbers. Yes.

So how it gets that information is through the web scraping, mostly Reddit, Facebook, probably not the best scholarly sources. It doesn't really know what it's not really making a prediction. It's just giving you what you think, what it thinks based upon a probability is true. Organizations can rely on AI, but I think a lot of organizations are not fully understanding that. The AI doesn't truly know, it just makes a prediction. So it's intelligent, intelligent enough, but it's not intelligent in the human sense. It's just making a prediction. So having common sense and understanding, it's just— I also think a lot of organizations don't think about the implications of AI. They think of it more in the terms of automation. And AI can help you automate, but it's not necessarily used for automation. So they're spending more money to automate something that they could just get an automation tool. Really what I'm trying to say here is that 95% of AI implementations fail across organizations. There's a study proven on this one, 2025, they're going to update it. This is because organizations aren't thinking about the system. They're just thinking about a pinpoint solution. You can go search hundreds of articles about failed AI implementation. One such is an organization trying to replace a call center. They implemented AI, they laid off all the employees. A couple of weeks later, they realized that the AI was horrible and they had to hire everybody back and get rid of the AI and start anew. It's not an uncommon story.

Automation and artificial intelligence, there's a clear distinction that you're making there. And if you want to automate things in your business, it doesn't mean that you have to use AI for it. So at this stage then, what is the best use case for AI in a small business, in a medium business, and in a large business?

I think for a small business, can complete tasks and potentially, like, not just automation tasks, but there's a lot of paperwork that you have to fill out that's not always exact that you can train AI on. Recently, a friend of mine wanted to profile customers and he has all the call transcripts either from his iPhone, just from recording calls or from Teams or other virtual meeting tools. And that has the transcripts. You can take off all the personal information, come up with customer profiles. You can make what's called a, a mini customer service management tool. That way when you get a new client, you can interact better. You can train yourself on how to talk to people better. It's not probably as good as human coaching or someone that's been doing it for a while, but he's really improved his skills and he's been able to land more clients. So I think AI could help in that sense. Like if you're using it to enable you, just not automate, like you're like you're actually interacting with it, like mimimimim, you're reviewing what it does, it's probably very good at those tasks. At a medium or large size organization, it needs to be more methodical. I think there's different layers to AI. So the larger the organization, there's really a leadership tier, a specialist tier, all your knowledge workers, not just AI, and kind of everyone else. What levels of information around AI and data do these individuals need. So one way to look at that, instead of just thinking about point solutions, you can start to think about the system. So one way I like to establish this is by trying to figure out pain points at that foundational tier, because you can link together all of those pain points to create that system. You can then go to the, the top of the triangle, the leaders, and explain some of those pain points and, and the value that you could get by creating that system and having human, validate it and ensuring that they're getting the proper training. And then the specialist here to help implement that for you. That way, everybody's now involved in the AI process. They're technically learning about it because the whole organization's involved and it's, it's more methodical. So instead of just putting in a system to do one thing, it's impacting multiple different departments and making them a much better impact and it's moving people. Automation. It's actually moving into AI. An example of this could be project management, for example. It could help break down tasks, find problems that potentially human isn't looking at, put those tasks forward. The team can decide whether they're valid or not, disperse the tasks to the data science team. If you have a data science issue, if there's some sort of ethical concern, it could for compliance, and, and that's just one use case example, but the ROI on that is getting products completed faster. There's less risk on the team. You're including multiple parts of the organization, so it's breaking down silos, uh, and, and there's many more examples of that.

So for those people then who are using AI as an advisor in their business, whether it's directly for the management a big company or a small company and trying to guide them. If you are telling me that this is just a deterministic answer and it's not really putting things in context, then does it mean that it's a false prophet and that what it's advising you is not really the real deal?

I don't think it's as real as, uh, as interacting with a human being, but if if you're trying to upskill and reskill your organization and it, you know, if you take the one-person example and now apply it to like an entire customer service team, you're having them train with the AI while they're not taking calls for an outage example or a billing issue. You could roleplay with that and at least create some muscle memory there and some common cues. I'm not saying it's a full replacement, but it's a great way to get individuals that might not trust AI to interact with it. It doesn't involve much— someone else could set this up and folks could interact with AI. So it doesn't involve prompt engineering. And that might also inspire the individuals interacting with it to take upon their own learning journey and learn more about the tool, which even from an organizational level is beneficial. But from an individual perspective, it's incredibly important to have AI skills. And human skills.

How do you get AI skills into your team? I know a lot of founders, including myself, who the first port of call is to open the chat window and then look at it and play with it and experiment and see what it can do and where it can go. But as you said, a lot of people might have staff in their companies who are not as keen on the idea of dealing with AI and interacting with it. So where do you start with that?

I always try to solve a problem. So for, for my first use case, I was trying to— I got sick of planning my workouts. It took a lot of time. There's a lot of cognitive load to that. I was spending more time doing that than actually working out. Sounds silly, but I was methodical about it. So that, that was a very good use case. Same thing with anyone else. Find something that bothers you. That you really don't like doing and then try doing it with AI. Even if you know the solution and you felt like it, chat engineer, for example, it's gotten a lot better with it. I don't use it for everything, but for specific types of appointments, arranging appointments, I have it send me an update and then I read the update and I approve it.

What tool do you use for that?

Gemini can, because I use Gmail, is a great, you don't have to do much of a setup for that. It's a great way to see if you can trust the AI, interact with it. Now, I did, I did try to do that a little bit much and it screwed up some of my emails. So there, there is some risk to it. Even an AI expert sometimes doesn't set it up correctly. So start, start small. And I guess that would be my second piece of advice is find a pain point, put the pain points together to create that system and then start small. You don't have to replace entire customer service department or make these giant things when you're not even through the foundations yet. And then my third is, once you understand all those parts, a lot of organizations skip this. It's the change management. And with change management comes upskilling and reskilling responsibility. So the organization has to train their employees on this new system, meaning that most employees need to know how to prompt engineer, and prompt engineering can be different for different roles, but understanding the basic prompt, the risks behind it, the information we don't put into the AI, what happens if there's an AI catastrophe at your organization, what's the responsible AI team doing, do you have a responsible AI team, does it just sit in compliance and security, should it be its own? So these questions, an AI— so to answer your question, an AI problem is really a human problem., and you have to figure out how that's going to work with your new organization that involves change management, coming up to the line, you can't build a team, whatever tools you're using, you have some training, technical training, skills training, it's a little bit different, and you have to realign the organization to that new system.

What would be the quickest risk check before an AI project goes live? And we're not talking here necessarily about large organization with an AI team and data and compliance, but somebody in a small or medium business, say the founder with, you know, 10, 25 people on their team who are starting to implement AI in one way or another. What's the quickest risk check?

Accuracy, whatever task that you're trying to have it do, can't do it correctly. So a lot of small businesses, um, medium-sized businesses are using AI agents to help find leads, customer leads. So they'll post like a YouTube video or some sort of digital media ad. It will collect that data and then the AI agent will follow up for those leads like, like human would. And then if there's some response and interaction, the probability that they become a customer probably increases and then it bumps it to a human. You could do a small Try only value 100 people to sit to see if that works for your product and your organization, the people behind. If it doesn't work, if it miscalculates the leads, it doesn't get the right people, if it doesn't follow up well, if you're not involved in that process too, like maybe you should approve the follow-up messages, it still saves you a lot of the steps, right? Like Thinking about this more slowly also helps, but if you can't do that, maybe that's not the first best task to do. Maybe taking the notes and summarizing for your team is important. The return on investment at that, and that is usually pretty accurate and limitless, right? So I guess it's really thinking about the system, what you want to do for your organization, but what your product is, but you also get that value returned by doing it longer publicly and understanding that by just going right for the most advanced AI might not necessarily trade off even the smaller tasks that it can do, still might have a return.

So you made a point there, Ben, about reviewing AI work before it's live. And everybody understands how to do that, or they have a certain standard or threshold to do so. But my question for you would be, How do you review that all AI work without burning the whole team's time? Because if you implemented it at different levels for different members of different teams, they can spend all their time reviewing AI work versus doing the work itself.

Well, there's different— correct. But you could also hire a team that, or maybe hire a person or two to implement this system for for you and they would check it and then they would have what I would call an expert reviewer of the department, right? So if you're doing digital marketing, you have a digital marketing person, the digital marketer would review this to proof it. So it takes, yes, that will take some time, but it's less time than building the AI, right? Like, so that involves all the upskilling and everything. The digital marketer will then still learn about AI because of the output for the validation. So that's one way we could do it. The other way you could check the AI piece is you could do cross-collaboration. So if you have 5 individuals working for you, you could have them switch the AI outputs just to do a quick check. Yes, initially that might take more time, but think about the lost time of doing something. It depends on the impact of it, but there's always that risk, right? Like of doing it so wrong that it might not be worth doing. And then over time. You'll, you'll build that confidence and for specific tasks, you know that AI can handle it. So you'll then just collect the log of what it does and you can even input that log because it's data into AI and have AI validate AI.

One of the things that I've, I've actually written a newsletter item went out today is to get good outcomes from AI, you can't let it guess. About your business. And I made the analogy of you have to train it and develop it like a new staff member. It needs to know what's good, what's bad, how do you do things, why do you do things, what are the desired outcomes. And it seems that a lot of people are skipping past this part and then they're disappointed with the output.

The change management's often skipped and the upskilling and the reskilling's often Daniel Ionescu. Back to thinking of AI as automation, I really think like there are a lot of individuals who will just pay for the $200 to $500 license, put in, put in the API, don't personalize anything, don't do much of the setup, and it will be able to do very complex things for you. And it's further from the truth. To get good AI, you have to be able to set it up. You have to understand how to prompt it. You have to give it a personality and templates. You have to set guardrails on what it can and can't do initially. If you don't have those guardrails, then I think it's hard for the AI to learn your business and it's hard for you to understand where it's going in your business. The second thing that Skip did, the change management, and I can't emphasize this enough, is creating the learning plan for everybody that this impacts. How the prompts, and it's more than just a lot of folks still think that AI is Greek and it's not. Like, you can't just write a sentence and sometimes you have to write very extensive, sometimes multiple paragraph system prompts in order to get what you want out of it. It also depends on the use case too. Like, if you're making images, movies, that's more, and that needs to be more accurate. Those take very, very, like, I haven't made that great of an AI movie yet, and I'm pretty good on prompt engineering. So that takes very advanced prompting. Try it. It's very hard. So that would be my recommendation in that case.

Even for a simple illustration, if you want, you can have a really long prompt, even if it's just something really simple.

Exactly. And it, and it might be simple to us, but it's— think about what the— right? Like the AI doesn't actually know what you're trying to do. It's making a prediction. So by prompting it that way, you're trying to pinpoint it in the right direction to give you the correct prediction.

And where would you draw the line between something being good enough to ship versus this is risky and we need to stop and reevaluate the output?

That's a fantastic question. I think that's a product decision, what you're trying to sell. Everybody's a little bit different. I think for lower-stake items, messing up an email might not be that big of a deal out of 100 emails, depending on what the emails contain and the sensitivity behind them. But if you're using AI to draft plans for something you're going to 3D print, like if you're an architecture company and you're using it to print house molds or whatever your 3D printing company does, and you spend thousands of dollars on the equipment and the supplies, and then it's all wrong, that will take human double-check. I don't think the AI is advanced enough to do the task. So I, the way I would do it. Is I would think, is this task more advanced or not? And if it's a less advanced task, AI can probably handle that. But if it's a very advanced task, AI can assist with it, but you're still going to need human oversight. So it's not going to be a full AI-fledged project.

So have you seen any examples in your field of work where the output was fine when it was in a draft or in the testing stages, but once it went live and in production, things went sideways?

It happens with coding all the time. In a small environment with small data or smaller data or synthetic data, the code may run fine, the code stack may be great, the model may perform well, it's giving good accuracy results. But then when you take that demo code to full-stack code and it interacts with different systems and maybe it's more live data and the rest of the data ecosystem impacting those model results, the, the code may break and the error might not be where you thought it may be. So then you have to take more time assessing where that code broke versus just writing it from scratch. But the way that you can use AI to help in that case is that you could, instead of doing a full product deployment, you could deploy in pieces and then you would, um, exactly where the AI might not be that great yet. And maybe we should have Department X work on this code.

So are there any early warning signs that you can look out for ahead of that happening? Because as we were just chatting, you can, things can look good in, in the test and in a small production. And once you move into, into live, it can easily break and you, or you push an update. If we're talking about code, and then everything can break. And you're probably going to tell me about human review.

There is, yes, human review. There's not just human review though. And let's use a different example instead of the code. Let's pretend we're doing the customer service, you know, the organization that thought they had a great customer service model and then replaced their customer service agents and then the model was horrible. Clearly something happened. In the live ecosystem that, that broke that model. So you really have to pressure test the AI as well. You have to give a different scenario. You have to have your best people interact with the AI. By pressure testing it and with that human view, it will get better. But you also have to understand the system. And I think what you're trying to articulate there is that just shipping a part of a product is a pinpoint solution. But what is the rest of the system dependent upon for this one task that we're having the AI do? And you can test that by not using the AI and what happens if this code breaks? Where does this fail in a demo system, right? Because most organizations have backups that you can demo on. Sometimes that's a change management piece.

You mentioned at the beginning of our discussion, Ben, that a lot of organizations or businesses have data in different silos, in different formats. And we both know, and anybody who played with AI a little bit knows that AI can confidently lie to you that it understands things, that it can read a webpage when it can't, and it can only see the headline maybe and the subheading. And I've been to a couple of conferences on AI and most of the people inside those organizations who were trying to implement it were telling me that data hygiene, as the term that they were using, is very important in order for AI to be able to read it and interpret it. And even a comma or a colon in the wrong place can throw off an entire result. How have you seen businesses approaching this data hygiene and arranging their older data, their customer discussions, old formats, and so on in order to allow AI to do a better job in their organizations?

Data is definitely the oil to the AI engine's input-output base. So if you have not that great data, you're not going to have that great AI. One way that you can look at data, so data is usually a, in most organizations, is a service. So there's maybe a couple groups in the organization or one person that's really good with the data and they help deliver that data in the form of reports or model updates maybe for the rest of the organization. And I think that helps streamline the reporting and potentially the data science. But what it doesn't help is, is the data hygiene problem. So now a lot of organizations are flipping to sort of data as a service, data as a product. So they're reviewing what data they actually need, maybe what data they don't need. Do they have to put a lot of cleaning behind the data when they have a similar variable or feature that will provide the same result? Could you make synthetic data that would resemble what, what the organization has so you don't have to invest in cleaning the data and eventually making one, one large, very large dataset that individuals can then make their own reports off of, make their own models off of. And then depending on the level of security and access that you need, they'll just get more slices of that dataset. So it becomes more of a product than a service. So you're enabling everybody and encourage— encourages data-driven decision-making. It's more agile. Allows everybody the same access to the data, so it's transparent, accessible. One thing though that I think that model doesn't really address is how you do the data science piece, but maybe a lot of organizations are thinking that in the future AI can help with basic analysis, but you have to understand the level of analysis that you want to do. So that's where I think the upscaling and rescaling comes. For that type of product where you have to, you learn how to prompt the AI on what you want. Do you want a table? Do you want a bar chart? Do you want a line chart? When you use those, are you trying to measure time? Are you trying to measure category? Does this report already exist? Is there a common repo that I can search so we don't have to put time into it? I can just use a template somewhere else. That's data as a product, and I agree with you, oil data is the oil to the AI engine. So without it, you're not going to have, you're not going to have much Yeah.

And one of the words that kept coming up in our conversation as well, Ben, is replacement. I think you said it 4 times now. And in your experience, did you see people in organizations who were being trained to use AI fearing that they will be replaced by AI?

That's definitely a fear. I do see it. Where I see it is where organizations have loose guardrails and they're not really expressing why individuals should be learning the AI, where, and that's the change management piece that folks are skipping. And even if, so what I want to express is even if you are afraid of that replacement, it's better to learn this technology while you can because it's paying a premium and all jobs are going to be impacted by AI. Every job is going to be augmented by AI. So it's better to learn those AI skills in combination with human skills, and then any amount of change happens, you'll have two very distinct skill sets that you can lean on and you can, you can pivot through the change. And I think a lot of individuals just think that this is another software implementation that they can just put their head in the sand and hope it goes away. AI is getting better about every two days. By the time this podcast is through, through. There's going to be some sort of either AI news article, a new app's going to hit the market. It updates every 30 minutes with something new going on around it. It updates in intelligence every 2 days. So don't think it's going away. There's been billions of dollars pumped into it. I do think it might slow down because of some AI whoopsies, which then companies or even governments might be too loose with the guardrails. But that doesn't mean that you should be afraid of it. It means that you should put some time to yourself and learn how to prompt engineer, understand risks. You could become an expert in AI for your domain with just point-and-click solutions without needing to know the AI theory. But if you're in any of the departments that you think AI is going to impact, you can learn how to set up those point-and-click agents. You're going to be the most valuable person in that organization, and they're still going to need humans to validate the outputs, to build new models once they come into existence, to learn the tools. So it's not all doom and gloom. I think there's actually going to be more jobs created because of AI.

So for a founder or an employee or a manager in a company, what would be the first AI skills that they should start learning or perfecting?

The first is, so don't think about it in the sense of tools, think about it in the sense of skills. And, and there's many different skills frameworks. I'm not gonna bore everyone with all the differences, but the World Economic Forum has a great one. They have what's called AI skills and human skills. They're two swim lanes. The reason why they're calling AI skills AI skills, because right now there's a premium for those skills. Humans can perform those skills, but over time, AI's gonna be really good at those skills. Human skills on the, on the flip side, humans right now don't get paid a premium for the human skills. Unfortunately, but over time, since AI can replicate them, but not fully do them, they should have a premium. So some examples of the AI skills are prompt engineering, basic coding, systems thinking, analytical thinking, data-driven decision making, being able to form a complex dataset for the data as a product. On the flip side, you'll have leadership, compassion, communication, some elements of project management. Those human skills, even though society might underlook them right now, think about like teaching, being a firefighter, police officer. There's probably a lot of job gap working in healthcare. There's probably a lot of job gaps in those industries because of the human element. I think there's going to be a lot of demand in those areas because it's going to be harder for AI to fill those roles than the knowledge-based roles in the AI.

Well, somebody was saying on a different podcast I was tuning into the other day that plumbers will earn more money than a lawyer in the future if we continue to evolve at this pace.

I mean, potentially it's— Anthropic actually released an AI job exposure index a few days ago. And what I found, and some of this, you meaning, but what I found interesting that most jobs on that index were computer science jobs. Most, not all of them, but computer programming is the highest to be replaced by AI at 74.5%. It doesn't really go into the level of computer science. I think that's more entry level potentially, but it's, it's top, top of the list. The second is customer service, which, which we've seen. The third that Anthropic's Ranking is data entry, then medical records specialist, which I think is still data entry potentially. Market research analysts, sales representatives, wholesale manufacturing, financial investment analysts, software quality assurance, information security analysis, and then computer user support specialist. So, so essentially basic coding and data and then basic customer service would be the two large domains, but the largest percentage for computer user support was a 46.5% probability. So it's almost all the jobs that I just listed have a 50/50 shot of being fully automated with AI. With that said though, if you learn the AI skills, you can help implement the AI around those solutions. The AI is probably going to make some mistakes. So these percentages are probably a little bit more optimistic.

So would humans, they would also do mistakes.

I mean, that is fair, but I think companies are going to rely too much on the AI and it's going to make a larger impact.

You made a point which I've kept on hearing across multiple of my guests in the last year or so. It's that AI is replacing the junior roles because it's good enough, if not better than somebody in a junior role. And it costs less to implement than a junior role. So in your view then, what happens if we lose this progression ladder in a, in a business, in a profession? What are we talking about? Web design, coding, writing, journalism, copywriting, marketing, all, all of these areas where AI is getting pretty good now.

Interesting. This, this Job Index report that Anthropic has also has data that, that you're mentioning. I don't remember where entry-level jobs fit into it. I think it was just for computer science, but essentially you're not wrong. Since OpenAI has launched, it's considerably kind of, I think it's about 22 or something percent like that. Uh, what I think education should be doing is, so it takes at least in the US, It takes 4 to 6 years to get a degree. Then usually what a degree was, was one and done, unless you want to go get a master's or, or something like that. And, and they taught very strict to a curriculum, but with this rapid change, I think it's very hard to, to stick to a regular coding curriculum or project management curriculum. It takes different, it has to be a little bit more pop is what I'm trying to say. So I would encourage all the students that are going through this, it really doesn't matter if you're an English major or if you're a computer science major to learn how to implement these AI tools, learn the backend to them, learn some data science. There's, there were more AI jobs open in 2025 than AI replaced. There was about 55,000 jobs. Difference between, between what it, what it replaced and the openings just by generative AI. And there's different forms of AI. So just by understanding generative AI, which is probably one of the easier, unless you're developing a chatbot, but most of it is easier than doing physical AI, which is robots and cars driving themselves. Just by doing that, you're going to set yourself up for, for higher success to get up, uh, to not be impacted by that gap.

And something else you mentioned, Ben, was augmenting your staff or augmenting yourself with AI. And I think that comes into play both for learning or teaching in that case, but also in a business sense. Again, one of my previous guests, Damon, was making the case of if you skill or upskill your best people to be using AI more, you would have more productivity than trying to replace them by AI, which ultimately would be inferior, at least at this stage.

So my thought on that is organizations don't really know where this AI thing is going. And by augmenting and upskilling your individuals, wherever this AI thing lands, you can shift the individuals around without as much impact on the change. And those skills are still going to be relevant if you're teaching coding skills instead of tools. If you're teaching the skills, they're still going to be relevant no matter how advanced the AI gets. And I know that there's this thought that there needs to be cost reduction or return on investment, but think about organizations at the current level you're at. You could probably maintain that same level, make more product, make less mistakes, have more satisfied customers. Increase your revenue and maintain the employees that you have. So it doesn't have to be one or the other. There can be an ecosystem where the augmentation supports both the AI and the individual.

In the final bit here, then, let's just do a bit of crystal gazing. You said that we're in a transition period. Even as I see it, we're in the early days, you know, ChatGPT came out just under 4 years ago. It's, it's still early days. Where do you see the future of work and AI going in, by say 2030? And then if we're going to be able to move it even further, that, that would be interesting to know. And what should people be prepared for?

I think by 2030 that we're going to definitely see AI in the organization chart. What that means is that humans will be managing AI agents similar to employees. Employees that they would manage. AI agents will have specific tasks, but those teams are going to need to understand how to interact with an AI agent, the risks, what it can and can't do. So those skills are going to become more in demand. How to manage the AI agents in the organization is also going to be probably a full-time job and how they're developed. So that's going to open up potentially more opportunities for domain experts in specific areas. I think this year we're actually also going to see physical AI, which is cars driving themselves.

Well, Elon has been promising that for years, so I'm still holding my breath for it, but sure.

Well, the World Economic Forum also believes that physical AI, that this 2026 is the year in physical AI. It's making major advances. I don't think it's going to be a full adoption of the physical AI. I do think though that understanding what that technology can do will then play into 2030, 2040. 35. To answer your question, AI in the org chart, AI agents, and then after that, physical AI to do other repetitive tasks.

So driving would be a very interesting one and it would unlock a lot of mobility, whether we're talking about transporting goods, whether we're talking about transporting people or whatever would be the lowest risk, but I would say transporting goods to start with. Although using the same infrastructure as the humans would also pose different dangers. The physical robots then, then that would take on roles and jobs. You don't see that coming until 2030 onwards, if we still don't have a self-driving car that's reliable.

Good is the keyword there. I think we're going to see it this year, but good, probably beyond 2030.

So this year would be more like a demo. It's not going to be something that you'll be able to go and buy and implemented into your business, whether it's somebody stacking the shelves in a supermarket or in, as you do now, say you would be able to replace, say, a junior copywriter with an AI bot.

I think that it will be able to do basic tasks, but it's not going to be super advanced and the investment may not be worth the upfront cost initially. There needs to be more training behind it. And I don't know if society is fully ready to see full functioning humanoid robots in a grocery store or school. So there's going to be, I think there's going to be more apprehension towards that technology than just ChatGPT or Gemini or Claude because it's web-based. But the technology is definitely there. Cars are going to be able to drive themselves. You can drive short distances. Mack, for example, already has driverless trucks. They have 100 in their fleet. They can stop at truck stops, can get loaded. They know where they need to get loaded, where they can get gas. There's no drivers in them. So, uh, this isn't far-fetched. It's already kind of happening.

Have you been in any of the Waymos or autonomous vehicles that are available at the moment in the States?

I drive an autonomous vehicle, so my car can drive itself. Yes.

But with your hands on the steering wheel, right?

I mean, well, yeah, I'm in the car, correct? That's the human oversight, but it's gotten a lot better. Okay. Okay.

So you're seeing the progress and you're trusting it.

I trust it more. I mean, I don't use it all the time, but if it's a longer drive and I'm tired, I can use it. It's a great feature. Let's put it that way.

And to wrap things up, Ben, I'm hoping you can give me a top 3 things that a business leader or a business founder from something small to something medium can look at implementing in their business in the next month or by the end of the year, since we have 3 more quarters to go, and to prepare themselves for what's coming in the AI for at-work era?

So the first thing that I would look at implementing is trying to think of a way to upskill and reskill your staff or yourself. It doesn't have to be academic. You don't have to go take a PhD-level degree program. You can make this fun. For example, Walmart just gave AI tools to all their employees and asked them to go use it. They can use it at their house. They can use it at work. It's open for everybody. They can give licenses away to other people that they know. So it's a very interesting approach to it, but that could, that could count. You could use it to create fitness plans. The next concept, if you haven't already done this, is think of prompt frameworks or prompts technology that will work for your organization. That way, those prompts then impact the next thing I'm going to suggest, what you actually want to use AI for. If you can flesh those three things out this year, you'll be set up for the next five.

So you heard it here on Millennial Masters. You got your plan. If you haven't started using AI in your business yet, this is what you need to do next. In order to keep up, not necessarily stay ahead. Thank you, Ben, for joining us today and sharing some of your ideas and your learning so far. And I definitely want a robot to maybe take on way more parts of my job.

Hold on to that. You want— humans want to be in control a little bit longer. It's not there yet, but thanks for having me, Daniel Ionescu. This was a lot of fun.

Thanks for tuning in all the way to the end of episode. If you enjoyed it, please share it with someone you think would get value from it. And don't forget to subscribe on YouTube, Spotify, or Apple Podcasts so you never miss an episode. For the best way to stay in the loop, head to millennialmasters.net. That's where you'll get all the latest features and new episodes straight to your inbox every week. Catch you next week.