Guest: John Thompson, AI Thought Leader, Author
Host: Andreas Welsch
Episode: What’s the BUZZ?, Season 4, Episode 15
Published: June 17, 2025
Episode Summary
John Thompson, Senior Vice President and Principal at The Hackett Group, joins Andreas Welsch to answer which skills leaders and teams need to work with AI agents. John says an agent can do anything a person can do and can be controlled in a more fine-grained way than an employee, and he holds that AI models infer rather than think. He expects companies to automate mechanistic roles, hire different profiles and keep growing people with AI, instead of switching hiring off. His advice for leaders is to build a safe, secure, private Gen AI environment that everyone can use, and to have technical teams learn MCP and A2A.
Key Takeaways
- John describes an AI agent as a proxy for a person: anything a person can do, a model can do, and a builder can control it more precisely than an employee.
- John says AI infers rather than thinks. The work it takes over is the mechanistic, repeatable kind, while creative synthesis and original thinking stay with people.
- John treats AI adoption as a continuum, not an on-or-off choice. Companies automate mechanistic roles, hire a different profile and keep growing and augmenting people.
- John recounts David Solomon of Goldman Sachs saying at Davos that 95% of an investment banking prospectus can be done with AI, with experts reviewing the last 5%.
- John’s three takeaways: have technical teams learn MCP and A2A, move past hesitation about AI, and build safe, secure environments where everyone can use Gen AI.
Timestamps
- 00:01:08 Guest background: Hackett Group, EY, five books
- 00:01:52 What changes when AI agents enter the workplace
- 00:04:10 What has to change, and why not to anthropomorphize models
- 00:05:50 Apple’s paper, AI infers rather than thinks, and entry-level jobs
- 00:08:20 The skills leaders need as teams grow uneasy about AI-first memos
- 00:10:33 Treat AI as a continuum, not an on-off switch for hiring
- 00:12:48 Agents as RPA 2.0 and agentic workflows
- 00:17:12 Critical thinking and AI slop
- 00:20:51 The one thing to do today: a safe, private Gen AI environment
- 00:22:25 Three takeaways: MCP, A2A and safe environments
About John Thompson
John Thompson is Senior Vice President and Principal at The Hackett Group, where he sets up a Gen AI services practice and helps build out Gen AI products. Before that he was Global Head of AI at EY and Global Head of AI and Advanced Analytics at the biopharmaceutical company CSL, and he worked at Dell and IBM.
John has written five books about data, AI and analytics and is a professor at the University of Michigan.
Full Transcript
Read the full transcript
Andreas Welsch: Today, we’ll talk about the top skills for working with AI agents, and who better to talk about it than someone who’s actively working and thinking about that. John Thompson. Hey John. Thank you so much for joining.
John Thompson: Hey, Andreas. It’s nice to see you. How you doing today?
Andreas Welsch: Doing very well. Hey, I’m so excited to have you on the show. I’ve been following your content on LinkedIn for a number of years. So I’m really thankful that you’re making time for me and making time for us.
John Thompson: As we said in the pre-show, I’ve followed your content for a long time, and when I got the invitation, I was really excited. I’m really looking forward to the conversation.
Andreas Welsch: Fantastic. For those of our guests in the audience who don’t know you yet, maybe you can share a little bit about yourself, who you are and what you do.
John Thompson: Sure. John Thompson. I’m the Senior Vice President and Principal at The Hackett Group, and my job there is to set up a Gen AI services practice and help build out the Gen AI products. Previous to that, I was the Global Head of AI at EY. Previous to that I was the Global Head of AI and Advanced Analytics at the second largest biopharmaceutical company called CSL Bearing. Before that I was at Dell and IBM and a few other companies, and at the same time I’ve written, you can see over my shoulder, I’ve written five books about data, AI and analytics, and I’m a professor at the University of Michigan as well.
Andreas Welsch: You’ve been in the industry for quite some time. You’ve seen it from many different angles. This will be a really good and meaty conversation. I’m super excited. Why don’t we jump straight to the questions that we talked about for today’s episode. There’s so much hype and so much buzz in the market, and it’s been there for the last nine months probably, around the topic of AI agents, agentic AI. We can delegate goals to software now all of a sudden, and it will work on all these magical things that we have been working on ourselves before. It will give us predictions and recommendations. It can also implement actions. And while that sounds very nice and very lofty in terms of goals, I think there’s also a very tangible aspect to it. And that is what happens when you actually introduce that kind of AI into a workplace. What changes, if anything? So I’m curious, what are you seeing in your conversations with businesses and with leaders? What changes?
John Thompson: There is a lot of buzz and a lot of hype around Gen AI and large language models and large reasoning models and AI agents, and they’re all intertwined. Some people say, I want to talk about agents only. I don’t want to talk about Gen AI. And I say you really can’t do that. They’re intertwined, they’re the same. And we’ve had agents for a long time. We’ve had agents for maybe 40 years or something like that. Those are symbolic, they’re not neural network, but there are symbolic-driven, rules-based agents. So agents are not new. We’ve had them for quite some time. But what changes with AI agents is that AI agents can do anything a person can do. If you think about a Gen AI-driven agent, you can have them do pretty much anything. They can spend money, they can book airfares, they can book hotels, they can fire people, they can hire people. We’ve all seen those missteps. I think the Canadian Airlines one is probably the most famous, where they were putting out tickets for 79 cents or something like that. And the airline, to their credit, honored all of those. So a lot can change. What has to change? Not much really, because you can build an agent and you can control that agent in a much more fine-grained way than you can control an employee. I always tell people, don’t anthropomorphize models. Don’t use pronouns like he and she and talk to them as if they’re John or Andreas. They’re just models. They don’t think, they don’t have emotions. But it’s helpful to think of a model as a proxy for a person. Anything a person can do, a model can do. So the breadth of change is either nothing or everything.
Andreas Welsch: I love that, and I must say I’m of the same opinion, that we shouldn’t anthropomorphize systems and call them Ava or Eva or something else, or John or Andreas, because it doesn’t mirror what they can do. They’re still based on data. They’re still working with probabilities and making predictions. But I find a lot of times now, when we talk about agents and these systems that can take over more complex tasks that we’ve been doing so far, that these analogies to how you would manage a team of people actually work quite well. What are the parts of the tasks that you are doing that we could now delegate, or that you would now delegate to an agent? How far can we actually compare humans and the way we work in an organization with how AI is now working?
John Thompson: The one thing that runs on a parallel path, or very close to what we’re talking about here, is all the buzz in the last week. Apple put out their paper, does AI really think. I’ve been very vocal about this for the last three years: AI does not think, AI infers. It infers on steroids. It can infer billions and billions of times a second, but it’s not thinking. It’s clearly not doing that. There’s also a great deal of conversation about the crisis of AI. Is it going to take all the entry-level jobs? You mentioned it in the intro, and Dario Amodei of Anthropic stirred the pot a couple of weeks ago by going in a big forum and saying nobody’s going to hire any kids out of college anymore, which is not true. That’s not the case. I think what we’re getting confused about is that many people are looking at what AI does now and saying, oh, it’s taking over the entry-level jobs. That’s not true. What AI does is a good job, as you said moments ago, of predicting the next letter, the next word, the next frame, the next concept, the next paragraph, whatever it is. So if you are a person or an employee who just does rote tasks, A, B, C, that kind of thing, parts of your job are going to be automated away. But if you’re a higher-level thinker, someone who’s doing creative work, synthesizing different parts of information, coming up with original thoughts and adding value to the organization, you have no worries. AI doesn’t do that. AI is nowhere near that. So what AI will take over is the very low-level, mechanistic, repeatable work. That’s gone. If half of your role is mechanistic and half of your role is creative, stop doing the mechanistic part and fill out your whole job with the creative part.
Andreas Welsch: This is a pretty big change to some extent, right? On one hand, we’ve always had this technological change where things we’ve done before have been automated. That’s the whole premise of software to begin with. But it feels like there’s a sense of urgency, and an order of magnitude that seems to be a lot bigger than what we’ve seen from previous waves of innovation. Not all leaders are comfortable with this change. First of all, their employees, their team members, are not comfortable. If you read the news and the headlines, whether it’s the entry-level roles or the ripple effects that go up the hierarchy, what’s the role of a professional? What’s the role of an expert? What’s the role of a leader? You see the CEO memos from Shopify, from Duolingo. We want to be AI first. We’re not hiring new people unless and until we can do that with AI. There’s a lot of angst and confusion and fear there. So how do you see leaders becoming more comfortable with this situation and also helping their team members go through that transformation? What are the skills that leaders need?
John Thompson: I see it all the time. We’ve seen it with Klarna and Shopify and all these different companies. We’re AI first. I’ve talked to over 500 companies over the last three years. Managers have said, when I go to hire a new person, I have to justify why I’m going to hire this person and why I can’t do it with AI. Fair. I get it. But the VP, SVP, EVP, C-level executives need to understand that you can’t really just turn these things off. You can migrate to it. You still need to hire people. You still need to do your jobs. You still need to build the technology. We’re not there yet. And the whole debate over the last week about large language models, large reasoning models, talking about, hey, these things are going to take over thinking positions, that’s not the case. Again, I’ll say it over and over again. Large reasoning models do not think, they do not reason. They infer at a high rate. So you just can’t say, we’re not going to hire any more people, we’re going to do it all with AI. What you can do is say, we’re going to look at some of the more mechanistic roles, we’re going to automate those, and we’re going to hire different people. Maybe you don’t hire the same profile you did before, but you’re still going to hire people. You’re still going to bring people into the business, you’re still going to grow them, and you’re going to augment them with AI. So to answer your question directly, I think leaders at the top of the pyramid need to have a more nuanced view of this. It’s not an on-or-off question. It’s a continuum. We want to take AI this far, and we’re still going to hire people over here. Where we were hiring people with no knowledge of our skills or our industry, we’re going to need people with a more nuanced and subtle understanding of what we do. So I don’t think it’s going to turn things on or off. I think it’s going to change them materially.
Andreas Welsch: That deeply resonates with how I’m thinking about this too. I’m visualizing it as a sliding window, from where you’re further to the left, more basic research, gathering information, gathering data. If we talk about entry-level roles, that shifts further to the right. You have software that does the legwork for you, but you are still in charge of ensuring it’s complete and accurate. Are there any aspects that we have not explored, that AI can maybe help us explore, or that we should be exploring and synthesizing, and then working more on the decision proposal as opposed to gathering as much data as you can and putting it together?
John Thompson: A really good example is David Solomon at Goldman at the last World Economic Forum, the last Davos meeting. He got up in front of all the world leaders and said, we generate all these different documents for investment banking, a prospectus or something like that. He said 95% of that can be done with AI, and that is true. A lot of it is gathering information from the SEC and getting information from the company and the market and competitors and Bloomberg and different places, and putting it all together in a predetermined form that everybody understands very well. And then for the last 5% of the process, you bring in experts who review it. That’s exactly how all these processes are going to work. If it’s all about gathering and collating and stacking up information, AI can do that all day long.
Andreas Welsch: That brings up another question for me. We’ve talked so much about automation for the last 10, 15 years. Obviously we’ve had rule-based automation for a very long time. Download a vendor invoice, save it locally, extract information, put it into your ERP and then kick off the process. So things like robotic process automation. Are we now talking about agents as RPA 2.0 or 3.0? Or is it a big fundamental shift in how we work?
John Thompson: Andreas, you’re the first person that’s put it that way, and I love it, because I’ve been positioning it that way for about a year and a half now. When people ask me about automation, the first thing I say to them is, how are you thinking about automation? Are you thinking about it the old-line RPA way, which a lot of people do? You could do it that way, but that’s A to B, C to D, X to Y, it never changes. It’s always the same. So I say, you really should think about this as an agentic workflow, which is very dynamic. The conditions are such that this really shouldn’t go to the governance department, this should go to the legal department. Or the velocity of change is such that you really need to talk to someone who’s managing the plants, because we’re seeing the demand signals ramp up significantly, but our production plan is not changing. So automation from an AI perspective can be very dynamic and very exciting, sensing what’s going on in the world. As you said, RPA is A to B, C to D, it just runs. It doesn’t change. Automation from a Gen AI perspective can be really dynamic. That automation conversation really needs to be grounded in a context of what you are expecting automation to do for you.
Andreas Welsch: I like that part. First of all, that we’re throwing the ball back and forth on this. But the part specifically where you said, making a decision, should this go to route A, to route B, to route C, and using AI to make that decision. We’ve used machine learning for a while to make decisions, for example, is this an expense that can be approved easily because it’s below a certain threshold, or is this actually something where a person needs to review it? Now we’re taking this a step further by saying, based on the information we have and all these different input factors, which route is the optimal one. So great to hear that from you as well. Now, when we do that and we outsource a lot of our thinking to AI, there have been some recent studies about this. What do we actually retain? What’s the role of us as team members, as team leaders, if we just say, hey, AI, agentic AI, whatever, agentic workflow, go figure this out for me, I don’t want to think about this? Is that taking it too far?
John Thompson: No, and I don’t think, at this point with the technology, as we’ve talked about, large reasoning models and the inability of LLMs to think, I don’t think you can take it too far. I think it’s really mechanistic, the way the AI that we have today works. It’s great, it’s fantastic. It does many things that we can’t do. If you brought to me, let’s say, 25, 30 page legal documents and you said compare them all and let us know where the privacy shortcomings are, it would take me weeks and I would be very bad at it. AI could probably do that in, I don’t know, 10 minutes, five minutes, six seconds. It depends on the model and the compute environment. So no, I don’t think we’re taking it too far. I think what we’re doing is we’re still feeling around in the dark trying to understand what is the domain of computing, which we’ve always had to do, and what is the domain of humans. In general we understand it, but it’s not a widespread understanding. Once we get that more clear, these comparative exercises against large corpuses of information are definitely in the domain of computing. The idea of creativity, keeping things under control, and making decisions with the vast context we have in our minds, that’s definitely human. We just need to bring those together.
Andreas Welsch: To what extent do you see people more or less blindly, or in good faith, using AI and saying, I let the thing do the groundwork for me, I don’t need to think about this, and then I just send this off, and we end up with AI slop?
John Thompson: That’s a great question and a great point. I’ve been feeling around in the dark myself trying to figure out what my next book is, my sixth book, and it’s actually going to be about how people need to be developing and refining critical thinking in the age of AI, which is just what you described. It’s one of those things where, when you ask people to think critically, you can almost see them go blank. They’re like, I don’t even know what you’re talking about. I think that’s an indictment of our education system. We’re not teaching people to think critically and we need to do a better job on that.
Andreas Welsch: It’s difficult, right? Even before Gen AI, you could make the argument that the value of critical thinking has been diminishing through social media and compartmentalizing, or living in your own social bubble. Now you add AI in when we’re already not as critical thinkers as we used to be. That gets more difficult. One of the other questions that I keep pondering with experts in the industry is, is agentic AI so much different from how we’ve usually and traditionally used software? Are there even new skills that we need to learn? Or is this just another conversation with a more capable chatbot, a more capable assistant, with almost an iceberg underneath the surface that we don’t see as users? I just ask it a question, it does the legwork for me, then comes back with something I would have had to do before. Where do you see this? Is it just another evolution on a continuum? Or is it a revolutionary thing that we now need to train everybody on and change how they think about it and how they articulate their goals?
John Thompson: Another great question. I think, Andreas, we could be here for about five hours doing this, but we’re going to run out of time. I always get painted with the brush of being a pessimist, and that’s not true. I’m an AI optimist. I think we have over the last three years seen a step function. It’s not a smooth evolution. It is definitely a step function in the capabilities of AI. But the way we’re going to use it and spread it more widely across the world, professionally, personally, educationally, academics, universities, research, is that we’re going to understand that AI in this respect, what I’m talking about now, is no different than any other AI system we have built. It’s all about the data. It’s all about managing it correctly. It’s all about integrating it. It’s all about bringing an ensemble of data together in an intelligent way that allows AI to understand, analyze, and project from it. What we’ve done over the last two or three years is go from using 10% of the world’s information to a hundred-plus percent of the world’s information, and people have not got their heads wrapped around that. But once everybody in information management understands that, in systems and development and AI, we’re going to have just a golden age of AI, in my opinion. So is it different? Yes. Is it really different? No. I’m on both sides of that equation.
Andreas Welsch: I like that. That’s wonderful. So maybe then to bring it home and make it tangible for business leaders, maybe those that are not quite sure, is this real? How big of a thing is this actually? Where should I even start? Are my people using this and do I even want them to use this? What’s the one thing that they can implement or encourage the team to do today?
John Thompson: You know what? I think this is a great idea. I did it at EY and it worked out exceptionally well. Every organization, every company, no matter what size you are, you should build a safe, secure, private Gen AI environment, and you should let everybody use it. For unfettered reasons. Let them write their weekend holiday agendas. Let them write sonnets to their puppies. Let them use it for anything, because what happens is you raise the level of capability across your whole organization. They’re going to use it for work reasons too, of course. They’re not going to waste a whole lot of your time and money, but you really should have every part of your organization using Gen AI in a safe, secure way. Don’t wait. This is not a hallucination. This is a real transformation.
Andreas Welsch: Great, and again, very actionable. I’ve seen many large enterprises do that, in fact, make different models from different vendors available to their teams. Then it’s up to the transformation team, the transformation office, to get people enabled and on board to use our safe and secure environment rather than some public version of whatever assistant where we don’t know what’s happening with our corporate data. Great point. Now, we’re getting close to the end of the show, John, and I was wondering if you can summarize the key three takeaways for our audience today. How can you build the top skills for working with AI and agents?
John Thompson: First and foremost, you should have some part of your technical team spending their time learning Anthropic’s Model Context Protocol and Google’s A2A protocol, because that’s how agents are going to talk to each other. MCP is going to be at the low level, A2A is going to be across multi-vendor, multi-agent. So you should have your teams learning that now. They’re going to need it. It’s like a 101 skill, they have to have it. So get them learning that. Secondly, if you’re concerned about AI and you’re reticent about it and you’re worried about moving forward, you’ve got to get past that. You need to talk to Andreas or me, or whoever on your team, or whatever consultant you want to bring in to make you feel comfortable. But now’s the time. Every day that you wait, the further behind you fall. And then lastly, I implied this, but I’ll be explicit. You really need to build environments that are safe and secure, and it’s not hard to do. It’s quite easy. The one thing that non-technical people need to understand is that Gen AI and models and all these different kinds of agents and chatbots are easily knitted together. You and I, Andreas, have been doing this for decades. You remember when it was really difficult to get two systems to talk to each other, almost impossible. Now you put a model repository together, you drop in Claude one day, you drop in GPT another day, you drop in a Cohere model, anybody can point at any model and use it. That’s just wild, it’s great. The tech has really enabled us to build stuff that your people can use, and let them go in an unfettered way, but make sure it’s safe.
Andreas Welsch: Wonderful. I love that. That really brings us home. So John, thank you so much for joining us and for sharing your experience with us. I really appreciated the conversation.
John Thompson: Thanks, I’ve really enjoyed it. Hopefully we’ll do it again soon.
Andreas Welsch: That’d be great. So folks, for those of you in the audience, thank you for joining us as well, and see you next time for another episode of What’s the BUZZ? Bye-bye.
John Thompson: Bye.
Listen and Subscribe
- Listen to this episode on Buzzsprout
- Watch this episode on YouTube
- All episodes of What’s the BUZZ? on Buzzsprout
Related Reading
- Agentic AI for Enterprise Leaders
- AI Leadership for Enterprise Leaders
- AI Leadership Handbook by Andreas Welsch
- AI Agents in the Workplace: What a Benchmark Shows
- Don’t Outsource Your Thinking to AI: A Leader’s Guide
- RPA and Agentic AI: What Leaders Need to Know
Frequently Asked Questions
Can AI agents replace entry-level jobs?
John Thompson says no. In his view AI predicts the next word, frame or concept, so it takes over low-level mechanistic and repeatable work, not creative synthesis. He expects companies to automate mechanistic roles and hire different profiles, not to stop hiring people.
Are AI agents the next generation of robotic process automation?
John says leaders should think about AI automation as an agentic workflow, which is dynamic, instead of the old RPA pattern of A to B, C to D that never changes. An agentic workflow can route work differently as conditions change, such as sending a case to legal instead of the governance department.
What is the first step for an organization that wants to prepare for AI agents?
John recommends building a safe, secure, private Gen AI environment and letting everybody use it, including for personal tasks, because it raises capability across the whole organization. He did this at EY. He also advises technical teams to start learning MCP and A2A.
Do large reasoning models think?
John says they do not. He argues that large reasoning models infer at a high rate rather than think or reason, and he refers to the Apple paper that asked whether AI really thinks. His conclusion is that leaders cannot treat models as a replacement for thinking positions.

