.png)
Real Work: AI and the Human Team
Join an AI expert and a team coach as they dive into the real-world dynamics of humans and AI collaborating in the workplace. Through candid conversations, they explore the challenges and opportunities of this partnership, tackling topics like trust, fear, diversity, and choosing the right tools for the job. No hype, no alarmism—just practical insights for making AI a valuable teammate.
Real Work: AI and the Human Team
What stage are you?
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Episode four opens with Dom's five stages of AI adoption and a blunt read on where Sean sits: solidly stage two.
So what gets you to stage three?
Go back through your own chat history, find the prompt you keep retyping, and turn it into a skill.
Dom walks through the email triage skill he runs across four accounts, on demand or on a schedule.
Dom reports back from Anthropic's developer conference in Tokyo and lands on the thing that separates good AI output from bad: give it a way to verify its own work. There's a warning here for anyone about to point an agent at their SharePoint. If seventeen versions of the revenue target are floating around, the agent will confidently hand you the wrong one.
Plus the gym waitlist agent that found its way into a booking system, and why that story ends better than the headlines suggested.
Welcome to Real Work, AI and the Human Team, where human potential meets artificial intelligence. We explore the critical space where people and AI collide, collaborate, and create something better together. From trust and fear to teamwork and transformation, we tackle the real challenges of making this partnership work. Here are your hosts, Dominic and Sean. Hey Tom, good morning. How are you today?
SPEAKER_01Good morning, Sean. I'm doing very well. How about you? Not too bad at all, not too bad. So we'd plan to talk about your trip. You went to Japan and you you went and saw the Claude community in Japan and a few other very exciting things. But before we get into that, I'd really love to hear a little bit more about your post this morning. It intrigued me. As you know, as you probably know, but I'll just go through it just in case. I use it AI and I use Claude quite extensively to uh draft up proposals, to research clients, to record records, to actually, you know, keep me in line, so to speak, and and and keep a lot of projects and a lot of balls balanced. I spent a lot of time with Claude and we get a lot more done than we would otherwise. I would really like to understand. You posted about these five stages, and I'm not single business, but I'd really love to know where you see me sitting and what would be next for me along that.
SPEAKER_02Yeah. So at the stages, maybe for those that haven't seen the post, the stages are kind of stage one, more ad hoc usage. So in an organization, people are sometimes secretly using AI, and the a technical term that is used for that is often shadow AI, because people use it even if the company doesn't support it or doesn't give access to it directly. And then stage two is more guided. There is a practice amongst the people in an organization how to use things. And then stage three, really, that's where things become systematic, right? Where there's maybe some people that are kind of the knowledge bearers that are the early adopters that help others in the organization to use AI. And that's where we start to share workflows, where we share certain context information, we document patterns. And then for I'll go through stage four and five quickly. The stage four would be we start measuring things, right? Because if the boss comes and says, Hey, we need to use AI now, and then everyone's like, Okay, like how much? What should I do? And to really make sure that the company gets its money's worth, we need to start measuring and see where where are we using it, where might we need to do some more training. And then stage five really is kind of where what I call supercharged people, where the humans and the AI starts to work hand in hand, and where people use things across the team, and where essentially AI becomes part of the team and where the team can achieve things that might not have been possible the year before. Now you're a little in maybe a different setup than many because you're like me as well, a single entrepreneur. You're working in your business, but you don't necessarily always have a big team. You might be working with others from time to time. But the way I see when we talk, you're definitely at stage two. It's guided that you have an agreed practice, even though you're mostly working with Claude, and Claude is your teammate, but you know what you're doing, and that works for you. So you're solidly on stage two.
SPEAKER_01So, what value would stage three be to me?
SPEAKER_02Stage three really comes where you review essentially what you have been doing over the last few months. So there, if you go through the history of your chats, or even if you go and ask Claude to go and look at things, you will see that there is certain patterns, certain things you do over and over, right? Maybe with new customers, maybe with research that you're doing about customers, but it's repeating things. And I'm pretty sure you're gonna find chats where you put essentially the exact same prompt in over and over again, maybe in different chats, but that's where you can say, hey, let's look at turning this into a skill, for example, in Claude terms, and also where you can start using Claude to document things like that. And so Claude can go and take reference of like, oh, last time we built a proposal for a customer, this is how we did it, this is the the key information that we need, and so it improves the improves the results.
SPEAKER_01So I actually do do some of that already. So I've I've actually flowed into with co-work actually creating projects. Now I think my projects are a little messy, and some of them overlap a little bit. So I'm getting to do one of those what are your markup documents to move from one to the other. I find myself stuffing up sometimes. I'm talking to Claude on the wrong project, and I go, What the hell are you talking about? Yeah, you agree. But that journey being quite exciting. One of the things, one of the barriers for me is Claude keeps telling me it's dangerous to do this.
SPEAKER_02That's that's interesting. You mentioned that before, and I'm I'm still not entirely sure what what exactly is the dangerous part. Right. So, do you do you have any any specific example?
SPEAKER_01So basically, what Claude will refuse to do is is hold any confidential information in any project that is allowed on the internet. Yeah. No, one note will do the backup, etc. So it won't disappear altogether. But Claude's extremely worried about that.
SPEAKER_02Yeah. I mean, that's to some extent okay, right? And and Claude and Antropic needs to kind of make sure that nothing nothing goes on that that shouldn't, or no data, no personal information goes out. I think that's where, for example, you were talking about cowork, where cowork can come in, right? Because in cowork you can actually connect it to folders that are locally. That are in your folders directly. And then there is other options, right? To say, hey, let's put certain information into Google Drive or into OneNote or something, depending on the environment that you've connected to. Because you can connect Cloud to Google Drive, and essentially it can use the drive as a storage medium for information.
SPEAKER_01The other end is that if you have an internet that's connected to the files in your OneDrive, Cloud's very worried about a document coming through that will just eat your drive.
SPEAKER_02I mean, it's always what you need to pay attention to what you put into Cloud, right? And essentially you don't want to just accept things from the internet or from external people and just throw that into Cloud. And Dropic has done a lot of work on this recently, where they improved security measures. And if you especially using skills, so that's that's more kind of the extension. If you don't write your own skills, you can also find some on the internet. But what you essentially have to remember is that if you if you go and get something from the internet from a source that you don't trust, then that means you're essentially installing and running software on your computer that can access anything. And that's where things get tricky, especially around those skills. People tend to forget and it's like, oh, this is just a new skill for my AI assistant. I just want to use this now. And the tip that I'm giving to people, these skills are essentially just text files. So that you can read them and use your eyeballs to look at the skills and see if there is anything nefarious going on in there. And Claude, you can also like say, Hey Claude, look at this without executing it. So there's they these features exist.
SPEAKER_01So you so I I don't have any skills online at all, and and I do notice Claude going and finding skills and using them sometimes, for instance, writing a Word document, that sort of thing, or moving from Word to PDF, or whichever. So these skills you can actually write them yourself and just like little subroutines almost for us people who used to use PLCs back in 1984.
SPEAKER_02Yeah, yeah, exactly. So it's it's a skill. The easiest is probably to just tell Claude, like, hey, look over the conversation that we just had and turn this into a skill. So one example is that you could say, Hey, I want to do a market research on a specific customer, and for this market research, go to their website, go to the business register, go to go to news sources and find all of these and then give me a report in a specific format. It's in one of my projects as part of a doc. So I have skills. So the next step, if you go and say, Hey, do this on customer ABC, and then tomorrow you come back and you say this on customer XYZ, then and you're always typing essentially the same thing, or you go over copy-paste, and then you miss something. Rather than doing that, you can go into the conversation that you just had with Claude, where you got a good result and say, Claude, turn this into a skill. This is a hidden phrase for Claude, and Claude knows how to create the skill, and it will then turn this into a skill, it will show you the skill, and then you can install that. So one skill that I created, for example, was email triage. I have several different email accounts. Besides my private one, I have a company one, and then some other organization one, and cloud community one. And sometimes things go missing, or not missing, but might slip through the cracks. And so I have a skill that knows or tells Claude to go to my private one, my company one, and the other ones, and then find out what is high priority, what I have not yet responded to, and then give me an update. So now I can run this skill and I get a little update. It's like, oh hey, Sean actually messaged you a week ago about setting up a new date for a podcast conversation. Let's get this done. And with Claude, I can actually run this manually, or I can say, hey, you can go and do that into a scheduled task. So every morning, maybe Claude goes, looks at what emails came in over the last 24 hours and builds a little dashboard essentially and gives me a feedback, and then I can just go and look at that.
SPEAKER_01That's nice, but it's sort of uh I've noticed that in my projects, and it might be my discipline being a little bit looser than yours, though. I don't know if anybody can notice the different accents, and sometimes the the national disciplines are seen as different, but I have a B D system whereby sometimes Claude gets very hung up on some crap that I don't think is important.
SPEAKER_02Yeah. Yeah, so this is about feedback, right? This is important for me. Um if you're within a project in Claude, these instructions can go into the project instructions, but also depends a bit on the configuration. If you have enabled memory and project memory, you can actually tell Claude, hey, remember this in this project. And this is like an additional information, additional instructions for Claude that is like, hey, if I ask you to go and do the customer research, don't like dig down into 20 years of history, like the last five months are okay. Something like that.
SPEAKER_01Yeah, yeah, I've got some of those that are going on at the minute, but I do find that he's not as good a coach as me. I hear people using AI as a coach, and I've found that Claude's jump to assumptions and jump to to this is a fact is is quite dangerous for coaching. But anyway, we were here to talk about your Japan trip.
SPEAKER_02Yeah, well, the the Japan trip has been a while ago, actually. That was going to Japan for the Claude Developer Conference. That was in June. Actually, uh two days of conference, 600 people per day in Tokyo. It was the first time for me to be in Tokyo and was quite mind-blowing. Also, the first time in a long time where I arrived in a city, and like as soon as you step outside the airport terminal, it's like, okay, there is still a lot of English text, but most of the signs are starting to be only in Japanese. You get on the train from the airport, and at least the announcements are in two languages, so in Japanese and English. But to be fair, I have used Claude a lot as a travel guide where it's like, hey Claude, I'm here, I need to go there. How do I get there? And what should I look out for? Like, how do I buy a transit card and stuff like that? Like, hey, there's like three different options. Which one is right for my three, four days in Tokyo? And then coming to the conference itself, this was more kind of focused around the development work done with Claude. Many, many people from Anthropic have been there. It was really cool to see kind of a few people that I only know from videos, like Brian Cherney, the person behind Claude Code, and I got a little photo up with him, all very humble, not kind of even though they're essentially celebrities, they're not like don't seem to have the the the issues that sometimes come with that. And was also some security there, right? I mean, speaking of celebrity status, yes, there is security at the gates, there is security if Brian walks around the the area. But it was really interesting to see what different people or different companies across the the world are doing with cloud, especially around the the development side, a lot of new features which might not necessarily come into play for for someone like you, with but in the background, these are getting added to the to the cloud platform, and sooner or later we get we're gonna have like integrations with more agentic work, right? So like some of that now has come into cloud co-work, for example, where cowork used to be only available locally on your computer, you can now do certain things in cowork as well on the internet. So when you're on the go, you have access to it via your smartphone, or you can if you have a computer and a laptop, if you're out with a customer, you can access the same co-work sessions as well. So it was really good to see the development and where things are going.
SPEAKER_01Here's a question that has not been primed, so it's going to be a hard one for Don. You said that you saw what a lot of companies over the world were doing with Claude. What would be the star attraction that you would love to sell in to help Australian companies do better?
SPEAKER_02So this is probably more from an engineering or like from a software engineering point of view, because that's really what the conference was about, and many of the talks were about that. But more and more kind of not outsourcing all the work to Claude, but really enabling things that have not been possible. So, especially for people working in the software industry, we all have been working on applications on code bases that are maybe five, maybe 10, maybe 20 years old, and there is work to be done on them that like, oh, this is too big. Like, oh, if we would do this, that would mean our whole team would be busy for three months and we would not get anything else done. And so these things are never getting done. Like updating to latest versions, bringing in better practices. And this is something that really is something that we can now do because with Claude, you can send off Claude and get work done that is very viable and that can be tested, and then Claude, that's where Claude really excels, right? And it's a little bit like with humans. If you give someone a task and they might not really quite know yet how to do the task or what how to verify the task, if you just send someone off that's like, oh, go and do this, but don't give them any hints on how to verify things, then the quality will be lower. Same is true with with these AI models and claude environments. If you give a way to actually verify the work, then Claude can go and do really good work because it can do iterations, it automatically goes, does loops. And I guess this is also applicable to many businesses that are not working with the coding environment, even if you're looking at using co-work. If you give Claude a task, the context information and the way to verify what the result is, then and you know how to how to ask for for it, you can get much higher results or much higher quality results. And as a side note, what I've also seen is rather than Claude being doing the maybe the engineering work or the thinking work, companies using Claude to verify, right? Not to not to check if the human is doing it. Oh but hey, like you're you're building this software, this is the input, this is or this is what the the feature, the plan was to be building. Like, is are all the security issues fixed? Are all the are is everything nailed down when it comes to security access? Is it documented? Is the documentation anything missing? And doing all those checks and kind of giving a feedback to the human of like, hey, there is something missing, and maybe here is how you can solve it, that makes a big difference. And I see that when I talk to people at my cloud events, a lot of people go, like, yeah, I don't want to outsource all my work, but I would like someone cross-checking what I've been doing before I'm like sending it off to a client.
SPEAKER_01Kind of like that chair, you know, whenever you see the pictures of the chairs, I can't remember who made it, where they bounce something on it repeatedly, yeah, and they do 20 years' work in 10 minutes. That sort of robust testing where they try different options, different ways. Yeah, that sounds like a really good use for Claude.
SPEAKER_02Yeah, yeah.
SPEAKER_01As you're talking, I can see, you know, and bringing that out of code into sort of my word. I've it's been a while since I've actually been employed by a corporate in consulting, but I still see the same problem for for my clients. We have an intranet that has grown, and we as humans have a respect for the authors of the previous document, and we don't get rid of the bloody thing, yeah. So it just grows and grows. And I've seen disciplinary procedures that were out of date. And there was the two new versions of that disciplinary procedure posted, but nobody knew about them because they sat in the wrong place and they were getting a link. There was a link somewhere in something that went straight to this one. A search of the internet pulled the wrong one. Whereas Claude can do that quite easily, I'm sure.
SPEAKER_02This is kind of two-sided, right? One side Is yes, Claude can do these things, but this is also where a lot of companies are struggling if they come into this world of AI, of like, oh, we're gonna set up an agent that does all these processes, and like an agent that goes and finds the documents. But if your SharePoint or your drive is chaotic and doesn't follow the right naming standards, doesn't follow the right filing, right? And as you said, it's like, oh, one version of the document is here, and then the other version is somewhere over here. And yes, Maggie from accounting knows, but not necessarily a stand-in when she's on holiday, and then suddenly the wrong thing gets updated or the wrong report is sent to the board. And so the first step when looking at automating things and introducing agents is actually making sure that the data context is clean and like that. We go through maybe all of SharePoint or Drive or the bits that we want the agent to touch and actually do an analysis. And we can use AI to do that, to do kind of the indexing and to go through and like, hey, find duplications, maybe not find out which is the latest version, but find all the versions and then make a recommendation which is probably the latest one. Find obsolete information and archive them, or give at least a report to the human, or like, hey, this really should be archived because it is five years out of date.
SPEAKER_01Yeah, two documents exactly the same with a different number and a different date. Get rid of one of them, don't you?
SPEAKER_02And it's not only the bound to update the wrong one, it is also about if you then start implementing AI assistance that can ask that people can ask for answers, right? Of like, hey, what is our revenue target for next month? If there is two documents that are about the revenue target for next month, and one is a higher version, but it is in a different directory because somebody filed it wrong, then there is a chance that the AI agent will answer with the wrong answer, give you the wrong answer. And that can have downstream consequences. Because if this agent, if it's just a human asking the agent for the revenue target, then that's one thing. But if it's another task that comes in and has like, hey, ask the agent for the information for the revenue target, and then that gets passed on the wrong information, and so you have a perpetuation of essentially not made up, but the wrong data, and that's where the adoption of this really often fails.
SPEAKER_01So AI performance is a function of the environment they're working in, as well as humans' performance.
SPEAKER_02Exactly. Right. And this is nothing new, right? I've been working in software engineering for 20 years, and one thing I learned very early on is never trust input data. Right. Like if you have an external system that sends you information, you always have to verify that the data is in the format that you expect, the length that you expect, has no special characters in there, that at least not those that you don't expect. And the same thing holds true in the age of AI, because if you give overwhelm the AI with 17 versions of the revenue target, then the AI has to choose. And what might be a simple task for a human to identify which one is the latest, this might actually be much harder for AI. And so you have to clean up the data.
SPEAKER_01The human will have heard the CEO talk about the revenue target.
SPEAKER_00Yeah, exactly.
SPEAKER_01No, that one's wrong and this one's right. Yeah, yeah. I do find that with AI it will latch on to the first truth it finds. And sometimes it's a bit of an argument to get it to go, no, no, not really.
SPEAKER_02There could be two things that are true here. Yeah, yeah. This often depends on how you set things up in the first place, kind of what is called the system prompt in the background, what drives the AI agents or clawed or what have you.
SPEAKER_01Well, I think mine might be don't tell me any lies. I told it. Therefore, whatever it believes not to be a lie, I will hold on to.
SPEAKER_02But yeah, I mean, there's tactics and ways to essentially say, hey, go and review your assumptions, don't stick to the first one that you have. Or with at least with things like cowork, you can say, hey, use multiple subagents, and like this is kind of the text you would use, use multiple subagents to look at a problem from different angles, and then look at the problem from different angles, and then do a refinement and say, hey, which one is maybe correct, or what are our options, right? Rather than like, oh, look for options. Oh, this one sounds good. Let's do this.
SPEAKER_01That's uh I really like the concept, and um you know, um as you said it, I'm thinking about skills and subagents and agentic and all this sort of thing. And I got this sort of being a little bit older, whatnot all, I got this fear. Where am I gonna do all this? And then I remember having that exact same fear to get to stage two, yeah. Yeah, yeah. And it's beyond us.
SPEAKER_02Yeah, yeah. And like it is not that different to being a manager in an organization, right? Because you have to start handing off tasks, you have to trust that your employees are doing their work, that they're doing the work correctly, and you also have to learn that sometimes people take another path than you would do personally, and that it might take them, at least the first time, it might take them longer than it would have taken you. I had to go through this learning experience myself when I became a manager, and then it's like, hey, person, go and fix this bug report. Here is the ticket, and then four hours later they're still working on it, and I know exactly which line of code this is, like, why does this take so long? And okay, maybe I should give them the information if I know which or think I know which line of code it is, then maybe I should actually tell them. That would probably help. But also like just accepting that it will take them longer than it might have taken me, and that I don't have all the control. And but if I give them the right guardrails, the right environment to thrive, then they will. And this again, the same is true with AI. And I think the big takeaway is just go and do it. Like take a little bit of time out of your day, out of a task, and say, hey, let's see if I can find out how AI can help me with this or how I can pass on a task that is very tedious to AI to at least get me there 80%.
SPEAKER_01And uh when you were talking earlier, the other concept that sprang to mind was the way I described the various different directions. You said use the agents to look at a problem from different directions. I have asked Claude, hold the tension between different truths. I'm a bit nervous of that because Claude has so many truths that it could be a I could blow up on thropic.
SPEAKER_02Yeah, yeah. I mean, it will probably just drain your digital allowance. Yeah. Um I don't think that will necessarily blow up entropic, but like we have seen that in the last few months that the agents that we have today, like Claude Code or JetGPT, they can be, if you give them a task, they can be very task-focused and overly so, and go to sometimes extreme lengths to achieve their goal. A few weeks ago there was someone that had an AI agent running and he wanted to go to the gym, but his favorite gym class in the morning was always booked out, and he couldn't get onto the wait list. And then he he asked his agent, like, hey, can you get me onto the wait list? And it came to the point where the AI agent then found the website and then found out, oh, the back-end system doesn't do verification of the future date, and so I can just add future dates and I can put you on the wait list for classes that are not even listed on the website yet. And then it went a step further of like, oh, there is a way to remove you from the wait list, but it's not secured, so I can remove anyone from the wait list, and like this in the media, this is like, oh, this is a hack. And like the AI broke into the gym's system. And at its base, yes, this is true, but any human that has a bit of understanding of how these systems work could have done that in an afternoon, but it would have taken the human a lot longer, and they will probably have gotten bored because it's about the gym and it's not about national secrets, so they wouldn't probably have done it. But the AI system had a plan, it needed to put the his user on the waitlist, and it found a way. And in return, now the gym software system, which is used by many gyms across the world, is now safer.
SPEAKER_01Yes, true, true. There is also another newspaper article that might be a good one to have a discussion around, and that is uh there was a guy who took on a computer scientist who had some unfair practices at work. I believe the courts have decided they were unfair practices, not my place to decide, but he purely used AI, and uh he won against a university's high-powered lawyers and law school, etc. University got law school, they got access to plenty of lawyers, all that sort of stuff, and he won using purely AI. But there was a comment from Fair Work as well, which was that Fair Work is swamped with uh small cases raised by free AI and they're all quite just time consuming to take it with, and they're not well founded. AI, the cheaper end of the market, you know, uh that's what's them all, they all have their cheaper entry end, people are using them for too much, and it's creating hallucinations, as we call them, etc. And all those different things that are being presented to Fairwork, and Fairworks going, there is no case law, you've made that case law up to satisfy your client, which which is somebody who hasn't done that robust thing of building their own profile on it. So that's a topic for conversation at some stage.
SPEAKER_02Yeah. Well, how about we pencil it in for one of our next episodes? We're already getting to like 35 minutes, so plenty isn't good for today.
SPEAKER_01Yeah, yeah. We'll let people carry on.
SPEAKER_02Yeah. All right. Well, thank you so much for your time, Sean.
SPEAKER_01And uh, thank you, No, it's been very interesting.
SPEAKER_02All right.
