8 Oct 2026 · 33m · Alison Beard, Adi Ignatius, Ethan Mollick · 17 quoted facts
Ethan Mollick argues that treating AI as a productivity tool misses the point. AI agents can now do weeks of work on their own, so leaders must rethink what their organization is for, how jobs are divided, and where humans belong. Because every company can use the same frontier models, advantage comes from how people and AI are combined in process and culture, not from the tools themselves.
The main tension is between automation and imagination. Mollick says handing tasks to AI is the easy default, but it produces the same company as everyone else and removes humans from meaningful work. His alternative is a deliberate mix of leadership, a lab and a crowd, all experimenting without a proven playbook. He admits that nobody knows how jobs or security risks will play out.
Mollick says that if KPIs and processes stay the same, AI simply yields more PowerPoint and more code. Leaders have to redefine the outputs they want, because people do what they are incentivized to do.
Leaders make decisions and take risks, the crowd gets access to the best models, and a lab of technical and organizational staff tests and builds. Mollick says this combination separates organizations that succeed with AI from those that don't.
Since everyone has the same models, Mollick argues for 'twilight factories' where AI brings people in for judgment, expertise, sign-off or interest. Fully automated 'dark factories' remove the differentiation and, in his view, are dangerous.
Mollick says vendors and consultants are making it up as they go, and past success stories often involve obsolete models. Companies must build their own in-house capacity to experiment and accept that they will fail along the way.
Mollick cites studies where less experienced people, such as BCG consultants, did worse with AI. He lists deep knowledge, wide knowledge, taste and agency as the human capabilities that matter most.
Mollick says productivity gains are real but get eaten up when the organization's processes and KPIs (key performance indicators, the measures of success) stay the same. You end up with more PowerPoint and code rather than better outcomes.
Mollick describes it as assigning work to AI the way you would brief a team, for example with a project request or standard operating procedure. The AI can now do many hours of work from one instruction, so you no longer work back and forth with it.
Mollick says the difference comes from how a company blends AI and human work through process and culture. He suggests 'twilight factories', where the AI brings people in at the right moments, instead of fully automating everything.
Mollick recommends leaders use the latest paid frontier models themselves, give staff access to them, and set up a team to experiment. He also asks four questions: what have you stopped doing because of AI, what impossible thing are you building, and do you have leadership, lab and crowd in place.
Mollick says nobody knows yet. Junior jobs may be squeezed, but companies may also hire more as they bring work in-house. He adds that the AI labs are explicitly aiming to replace human labor, which he thinks they may get wrong and which society shouldn't let happen.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Ethan Mollick says AI benchmarking company Epoch reported getting two to eighteen weeks of engineering work from a single prompt to Claude for a couple hundred dollars.
“Epoch, which is a AI, uh, benchmarking company, was reporting that they could get regularly sort of between two and eighteen weeks of engineers' work out of a single prompt to Claude for a couple $100, right?”
Ethan Mollick says the company StrongDM ships product to customers with no human involved in writing or reviewing the code.
“I know one company, Strong DM, that has been shipping product to customers, and, uh, there's no human involved in that.”
Ethan Mollick cites an Anthropic study finding that managers were the most successful at first-pass coding with AI because they specify what they want well.
“there's a great study out of Anthropic that shows that the people who are the most successful coders at the first pass of coding are actually managers, not coders”
Ethan Mollick says early studies with colleagues at Harvard and Warwick found that less experienced BCG consultants were worse at using AI.
“we found that, like, BCG consultants less experience were worse at using AI.”
Ethan Mollick says OpenAI moved 20% of its staff off research to harden its systems against security threats.
“OpenAI pulled 20% of their staff off of research to 20% of the resources to hardening their systems.”
Ethan Mollick says leaders of large financial institutions are moving more than half of their security staff onto responding to AI-driven threats.
“I have talked to leaders of large financial institutions who are pulling 50% plus of their CIS- their staff”
Ethan Mollick says the biggest thing he got wrong in the past year was expecting agents to be hard to organize, when swarms of 100,000 agents now self-organize.
“There are 100,000 agent swarms that are self-organizing. I didn't expect that to happen”
Ethan Mollick says he runs projects in which a swarm of a hundred AI agents works on a problem and reports back to him like a manager.
“I literally have projects where I have a swarm of a hundred agents solving a problem for me.”
Ethan Mollick says that in his first startup he spent about eight hours a week doing payroll taxes by hand in Excel before learning a payroll company could do it.
“I would do taxes by hand in Excel for every person at the company. It took, like, eight hours every week before I realized better way.”
Ethan Mollick defines a 'dark factory' as a factory where robots do all the work, so no humans or lights are needed.
“So dark factory is factory where all the work is done by robots.”
Ethan Mollick draws on Harvard's Mike Tushman for the idea of organizational ambidexterity, doing two things at once: exploiting what works today while exploring the future.
“to take an idea from Harvard, from Mike Tushman and company, uh, it it was the idea that you would need ambidexterity.”
Ethan Mollick says the coding profession moved from writing elegant code to systems engineering as AI took over code writing, and AI is now doing some systems engineering too.
“You're hiring coders to be the best at writing elegant lines of code. Then AI could write the code. So then it became about sort of systems engineering.”
Ethan Mollick says the AI labs explicitly aim to replace all human labor, a goal he thinks they may be wrong about and that society shouldn't let happen.
“the AI labs, to be clear, are aiming for the job apocalypse. They want AI to replace all human labor.”
Ethan Mollick says consultants' claimed AI success stories are often based on GPT-4, a model he calls obsolete.
“if they have successful cases, it's for GPT four, a model that's obsolete and no one would use for anything anymore”
Ethan Mollick says one of the first effects of AI that executives see, which also appeared in his research, is that job boundaries blur.
“jobs start to blur. Coders do design work, designers do coding, product managers do everything.”
Ethan Mollick says he has spoken to C-level leaders who used AI to query company information and diagnose problems at plants that no one had raised with them.
“who've actually diagnosed problems at plants that no one brought up to them because they could just literally query all of the information at their company's disposal”
Ethan Mollick says there is ample evidence that AI can regularly do work that would take a human two to eighteen weeks, at reasonable cost and with high success rates.
“AI can regularly do the kind of work that would take a human two to eighteen weeks in a reasonable amount of time, a reasonable amount of money with high success rates.”
Mollick explains that AI keeps improving and that autonomous agents are now a viable way to do work. This sets up the argument that strategy must change.
“the exponential curve continues, AI just keeps getting better”
Productivity gains are absorbed unless processes and KPIs change, and a default of automation leaves humans out. Leaders must decide what AI is for.
“those productivity gains get eaten up by the organization”
Mollick lays out an organizational model and says leaders must take risks and use the latest AI themselves.
“So you need three things to succeed.”
AI can now handle many hours of work per instruction, so everyone becomes a manager and management skills become AI skills.
“conjuring a team into existence when you need it”
Fully automated 'dark factories' remove differentiation, so Mollick proposes designs where AI brings humans in at the right points.
“I think we need to think about building twilight factories”
Deep knowledge, wide knowledge, taste and agency matter most, and senior leaders have an advantage because expertise improves AI results.
“I think there are four things that are very useful.”
Mollick says both sides gain: incumbents have experience and resources, while startups can now cover their weaknesses with AI.
“it's never been a better time to be a startup company”
Mollick expects a disruptive security environment in which AI agents probe exposed systems, even without malicious intent.
“every part of your system exposed to the internet is going to be a target for agents”
Mollick warns against anchoring on today's capabilities and trusting vendors, and says the labor-market outcome is unknown.
“Every consultant you talk to does not have seventeen successful cases”
I'm Allison Beard.
And I'm Adi Ignatius. This is the HBR IdeaCast.
Adi, for the last few weeks, we've been exploring the ways that AI is changing leadership and organizations today. We've looked at talent, communication, innovation, and now we're gonna put it all together and look big picture at strategy.
Yeah, so there are a couple big questions on my mind. First, if every company can access these amazing AI tools to generate ideas, to analyze information, write software, to carry out complex tasks, what does competitive advantage even look like? And second, you know, how should companies start redesigning their organizations, their workflows to take advantage of what AI and what humans can do?
And how do you do that every day when the technology is so unpredictable yet getting more powerful? So those are just a few of the questions that I put to our guest today, Ethan Mollick. He is an AI expert, an associate professor at Wharton, and author of the forthcoming book, Co-Existence: The Next Phase of AI. He has lots of thoughts about what strategy looks like in the AI era, how companies should be using it to rethink their business and organizational structure, and what human capabilities become even more important as these new tools improve.
Here's our conversation. You've been closely tracking advances in AI and constantly experimenting with all the new tools. What is different today than even a few months ago in terms of how new developments should guide the way that leaders are thinking about strategy in particular?
One is the exponential curve continues, AI just keeps getting better. There isn't really a sign of that stopping right now. And the second is that agents are real, which are AI tools who can go off and do work on their own autonomously and do real process work, real important things. And that has become a a a viable approach to work.
Yeah, it seems like a lot of companies right now are thinking of AI as really a productivity exercise. You know, you're automating tasks, you're saving time, you're reducing costs. How does, you know, a genuinely strategic AI transformation strategy differ from that? What do you think companies should be doing now?
I think it's actually pretty urgent to do a few things. One is viewing this purely as a productivity tool gets you in trouble in a lot of different ways. People are getting ROI from AI, right? Productivity gains are real. The problem is is that those productivity gains get eaten up by the organization, which is if you don't change any part of the process, AI is mostly a single player entity and you end up with more of everything.
So if you don't change your KPIs, people get produce more PowerPoint, they produce more code, that's not actually what you want. So leadership needs to decide what this thing is. And the second is AI is getting very capable, and I worry that if our default is automation, we end up in a situation where we just keep handing tasks to AI and don't think hard enough about where humans fall into this process. And that leads to a lot of danger as well.
Is your argument that leaders need to move from asking sort of where can we use AI in this existing business alongside humans to a question that's more like, you know, how does AI change our business going forward?
At this point, we have ample evidence that AI can regularly do the kind of work that would take a human two to eighteen weeks in a reasonable amount of time, a reasonable amount of money with high success rates. That implies a lot of different things. It implies that what your organization does has to change. How you think about the boundaries between jobs is changing. I talk to C-level leaders all the time, and one of the first effects they see of AI, and we saw this also in our research, is that jobs start to blur.
Coders do design work, designers do coding, product managers do everything. How do you start thinking about what the boundaries of your organization are? What's the product you're trying to produce? Is it an interim product of a PowerPoint? What was the point of the PowerPoint in the first place? You start to run into really fundamental questions about what the goal of the organization is, how you accomplish that, what are your values for getting there, and they're urgent.
This isn't sort of theoretical, you know, let's sit down in an academic way. It is, your organization is already changing from the inside. What do you do about that?
So what are some best practices that you recommend to leaders who are grappling with these issues right now that need to start having those really urgent conversations and not just have them, but sort of make them progress enough so that they can actually start making some real decisions?
There's some individual feedback and there's some organizational feedback. On the organization side, the model I've seen works we- best is sort of like a meta model, which is leadership lab and crowd. So you need three things to succeed. You need the leadership team to actually lead in this, right? That's not just say, use AI. They have to make decisions about how AI is used, what the organization does with AI, how do we incentivize people. The crowd means giving people access to these tools.
We're at a point now where giving organizations inferior AI access because of either just because you haven't bothered to improve the most recent models leads you into trouble. Like, you'd actually become very deluded about what AI can and can't do. And then you need a lab. You need people doing twenty four seven work who are not primarily technical, by the way. There'll be technical people there, but also organizational people who are thinking about how do we restructure the organization around AI.
They're benchmarking what systems could do, building systems of the future, taking ideas from the crowd and ideas from leadership and helping build them. So you need to be doing this as a whole organization response. At the individual level, the factor that I'm seeing that differentiates leadership that succeeds at AI use and organizations that succeed and those that don't, is leaders who are willing to eat risk, who are willing to say, look, nobody knows what's gonna happen here, but unless we do active experimentation, we're not gonna know anything.
And if you just let everybody in the organization have vetoes over AI use because it might be risky, it is, but not doing anything is risky too. That also means actively using it yourself. So leaders who don't use it the latest AI systems, who wait for reports from other people, who are waiting for a training some week, fall behind those who just are experimenting.
I think the last time that you and I talked about this, you sort of had this model for human-AI collaboration that was either a centaur, you're half using your human skills and then half using AI, or a cyborg where you're going back and forth between the two. Talk about how that is evolving to one that you now describe as conjurer work, that you think, you know, leaders should not only be doing themselves as individuals, but also enabling in their organizations as part of their strategy.
The biggest change, if we really look at the sweep of the last year, was that in every independent study we have, basically, AI went from regularly being able to do a half hour of work, forty five minutes of work, to you can get sixteen plus hours of work out of a single AI instruction with good chance of success. And that means you don't walk work next to the AI anymore. You don't work back and forth with it in the same way you did before.
You are actually assigning the AI to tasks almost like you're a team you are assigning, and it does the work. So a conjurer is really about conjuring a team into existence when you need it. It's almost like writing an RFP for a project you need done, or SOP, or any of the other standard TLA three-letter acronyms that we use in business to assign work. You send that out and the AI just does that work. That's a very different situation than I prompt the AI, look at the results, I prompt it again.
This is a very different, you know, action-oriented field.
How does that change how leaders approach the structure of the organization, what it even means to be a manager?
So everyone's a manager to some degree right now. Everyone has AI to manage, and AI skills look like management skills. In fact, there's a great study out of Anthropic that shows that the people who are the most successful coders at the first pass of coding are actually managers, not coders, because they do a good job specifying what they want, and the AI executes on that. So that is the big problem, right? Your organization's suddenly capacity increases.
And I think organizations that don't have imagination are going to retreat to automating away the work. And automation's a well-worn playbook. I think it's a recipe for disaster. I mean, it's a disaster for all of us if all of our work becomes automated or we just become verifiers of AI content. That's, like, not a good situation to be in. But also, it removes all the differentiation. Leaders need to be building companies that blend AI and human work together with process and with culture in ways that accomplish things that neither could do alone.
If you just move to automation, you're gonna have the same cloud-based company everyone else does.
Yeah. Well, that's an interesting point. So, you know, when we're talking about strategy, we're also talking about finding competitive advantage. So if every company has access to the same frontier models, where exactly can leaders differentiate their organizations from competitors?
That's where the key kind of comes in. You are probably not using AI enough inside your organization, not trusting enough, but you then want to bring people back in. I I mean, we're already seeing as the the most extreme version of this, right, in terms of organizational change are things like software dark factory. So dark factory is factory where all the work is done by robots. You don't need humans in it, you don't need any lights.
I know one company, Strong DM, that has been shipping product to customers, and, uh, there's no human involved in that. Their rules are no human can look at code, no human can write code. It is 100% AI automated. I think that's a dangerous approach because you're just letting the code do the work. I think we need to think about building twilight factories, situations where the AI proactively reaches out to people and brings them in because they have variance in views, because their process or expertise matters, because there's sign off and approvals needed, or because they find the work interesting.
And I think we need to start restructuring work around the idea that these systems really can accomplish many things, and we'll be able to accomplish more in the near future. So it's integrating with humans at the right places at the right time and process that's going to give you a differentiator.
Where are the specific ways in which you think that, as fast as AI is evolving, humans still have a role, and particularly leaders still have a role?
Uh, one thing is you can't really be in the loop in the same way you were, right? It was easy to be in the loop when you were when it was a chatbot. It's not easy when, I I mean, I literally have projects where I have a swarm of a hundred agents solving a problem for me. I have the report out to me like I'm a manager, so I I am in the loop to this the extent that, you know, they give me a PowerPoint about their progress and they show me demos of what they're building.
But that's very different than, you know, I'm approving every action. Leaders, actually, one of the outputs of AI is leaders are super empowered. I keep seeing organizations where leadership basically takes back control over their company because they can now directly ask the AI for a, you know, for the information they need. I have talked to, um, people who are like C-level leaders that who've actually diagnosed problems at plants that no one brought up to them because they could just literally query all of the information at their company's disposal and ask questions and have the AI go off and do tasks for them.
So it's a really interesting change in the span of control if you set this up right, which lets leadership actually think more about things rather than hearing, uh, things through three or four different layers of leadership.
So would you say then that AI agents are making strategic decisions on behalf of leaders and organizations with the right context, you know, given to them?
I mean, they already are. Epoch, which is a AI, uh, benchmarking company, was reporting that they could get regularly sort of between two and eighteen weeks of engineers' work out of a single prompt to Claude for a couple $100, right? That involves tons of strategic decisions and judgment calls and creativity calls. They may not be your judgment or creativity or or, but, like, we can't say anymore the AI isn't making strategic decisions. It is. The question for you is where do you want to make decisions and where not?
And just like any team, you can give it those instructions. You can say, listen, I want you to come with the, like, the five best options, your defenses for each of them, surface those, pressure test them for me, and then deliver f- so I can pick between them, and this is the format I want them in. Or you could say, just go ahead and execute, build the prototype for it, and test it with a few people.
Those are now options you have available to you. So some of this is reconstruction we've already done with humans. What's the process by using humans and AI together that I get a working idea that's much better than I had as before? How does that get done much faster and and better?
Where exactly, though, does a leader know where they add the most value and where they don't? Like, what are the the points at which they should be involved absolutely every time?
I think part of this is you have to make decisions now about what your how to best accomplish what your company wants to accomplish. And you obviously don't want to give this all up because the AI is not as smart as you at what you're best at. Because the AI doesn't have enough variation because it doesn't actually absorb your culture in a deep way. You know, it can give some of that as context, but that's part of process design.
I mean, we do this all the time. What are the limits of your product engineering group versus your marketing group versus your legal team? We have to think about the same kind of process design here, which is where are those limits and limitations? I think the danger is is that by skipping doing that, you either move to an all automation world where all these choices are made by AI, or you underestimate what these systems can actually do.
So it is actually disciplined decision making that gets you there.
And what are the things, the aspects of strategic leadership that you don't think can ever be delegated or outsourced?
So I'm always careful about bright lines, where we say nothing is something can be never done, right? Never ever, But these systems are jagged. They're good at a lot of stuff in bad at ways, stuff you wouldn't expect. And that means, by the way, there's different things the jobs do. Think about coding, right? You're hiring coders to be the best at writing elegant lines of code. Then AI could write the code. So then it became about sort of systems engineering.
Could you see the b- bigger picture? Now AI is increasingly doing some of the systems engineering work. In some ways, it's very upsetting if you're a coder because the nature of the job is changing in ways that are good or bad all the time, right? And what you learned as craft no longer is. I think we're gonna see the same thing about leadership. Like, right now, your your maybe your judgment calls are about, uh, you know, about the details of pricing strategy.
The AI might be better at pricing than you. You won't know till your lab tests it for you. I don't I can't give you the answer in your industry or not. You'll only know by experimentation. And you don't have to do that all alone, right? That's where your lab and your crowd comes in. Your crowd is surfacing use cases and values of AI. You don't have to do this all yourself. But I think a big danger of companies is they think someone solved this problem for them.
Someone's figured this out. We're all faking it. Nobody knows anything. Like, the AI labs, you know, don't know anything. The w- the vendors are all making these up. This is a three-year-old technology. So you have to sort of seize control to some extent of this yourself to figure out what you want to use it for.
Yeah. Are there capabilities that you suggest your students, your MBA students, your kids, um, aspiring leaders that you suggest they develop because you believe they'll be increasingly important in an AI-dominated world?
So I think there are four things that are very useful. They're not the only things. Uh, one of those is deep knowledge, which is the ability to actually understand a field really well, right? So we're on this podcast, and you know from the beginning of my question or not, before I even say, like, a few sentences, whether this is gonna be an interesting answer or not. I hope this is one of them. Uh, and y- and whether you need to redirect things or not, that comes from deep expertise.
Similarly, in your industry area, you can look at a spreadsheet and know whether something is wrong. You don't have to read those 400 pages of AI content, because you know right away because you're an expert. It's one of the reasons why junior people are often struggling now. They don't have that expertise. Second thing you need is wide knowledge. As people start to cross field boundaries, knowing a little bit about many things helps you give the AI better feedback and ask for the things you want.
The third thing you want is taste. You want to be able to, like, pick among things, and that could be company taste too. What strategy feels like the kind of strategy our company would do? Out of these many designs the AI came up with, which is the right one? Selecting among many options. And I think we can develop taste. And then the fourth one is agency, the willingness to take action under uncertainty. And I think that's a little bit harder to teach.
We try and do that. I'm an entrepreneurship professor. Um, but I think giving people the leeway to to engage in experimentation is important.
Yeah. How are you encouraging the sort of old guard of leaders who aren't accustomed to such uncertainty to become more comfortable with it?
I think there's a danger of people overstating the generational aspect of this technology. Like, it certainly was true that digital natives were a thing. Like, you didn't grow up with the internet. There's a lot of weird stuff about the internet you don't understand as an internet user. It doesn't work the same way with AI because AI doesn't have the same sort of innate background. It's sort of a single player setup at this point, and because it rewards expertise.
In our early studies with, uh, some of my colleagues at Harvard and my team leaders at Warwick, we found that, like, BCG consultants less experience were worse at using AI. And there's other studies that show less expertise results in worse AI outcomes, including for Anthropic itself. And so I think one of the dangers you have is you're like, Oh, I need these, like, you know, it's there's some group of people who gets this. Nobody gets this, right?
So as a senior leader, you actually have the best tools available because you have that expertise and judgment to make decisions about things. And you probably have a thermometer or two too. You know, if you've been around for twenty years or, you know, you've been through financial crisis, you've been through wars. Like, this is this is going to be a time of a lot of change. That doesn't mean that you're inflexible because you're a senior leader.
In fact, often the best senior leaders who got where they are are good at deciding how to make changes.
Have you seen companies or leaders that, you know, really successfully use AI in strategy development itself?
I've talked to C-level leaders at very large, uh, funds who've done this. I've spoken to senior leaders in large financial services organizations, uh, certainly in software. Yes, there are people doing this all the time. And in fact, I would argue you're probably making a mistake if you're not getting a second opinion from the AI. Like, the models are very smart right now. It's just like if you're if you're getting medical advice, absolutely see your doctor. But every piece of evidence is you should absolutely be feeding that into the AI system too, right?
I would do the same thing with legal advice. To listen to your lawyer first, but why not get a second opinion? Second opinions are cheap right now, and the best models, and again, I can't emphasize enough, you need to use one of the frontier models, which, as of the time of the recording, that means, you know, Claude Fable of 5.1 or whatever it'll be in the next couple days, and, uh, and GPT six. And, you know, maybe Google comes up with a a really leading model again, but you need to use one of these frontier models that you have to pay for.
And you should be asking it questions. Like, the only way you're gonna find out what this is good or bad at is by using it. So, by all means, use it for everything during the day, take its opinions, decide where it's dumb. That's a useful set of information for you.
When it comes to strategy, does AI do you believe it ultimately favors large companies with more resources and data and those experienced leaders that you were talking about? Or, you know, does it allow much smaller companies to compete, you know, with capabilities they couldn't access before?
Both of those things. One of the most depressing things I see when I talk to boards of very large companies and to C-level leaders, is a lot of them have come to believe they just can't change, right? So I show them what AI can do, and they're like, Well, I guess we'll ride this out. And I'm like, That's an insane answer. You have 100,000 people working for you. The smartest people in the world are working for your company.
You have customers and resources. The idea that you can't be flexible and compete seems to be overly baked in right now. And like I'm a business school professor, I understand we talk a lot about, you know, how difficult it is to survive disruption, but it's also a very weird decision because you have all the resources. Because who's going to figure out how to use AI in your organization is going to be people in your company. Experience has returns, right?
The the startup doesn't have any of the same barriers you do, but they don't have the experience. And that experience matters. It's no longer a twenty-something knows everything about this. The more experience you have, the more you know about the in AI in your field. And everyone's using the same AI models. It's not like the startup you're talking to is going to have a better model than, you know, than Claude, right? So I worry a little bit that people are sort of giving up.
On the startup side, it's never been a better time to be a startup company. Startups are all about that you're the top point 1% at something and everything else you're terrible at and you hope that doesn't destroy your company. Now we have a tool that makes you good at everything you weren't good at. You know, I I'm somebody who, when I was in my first startup company, we had to do payroll, and I didn't realize you could pay a payroll company a couple cents, uh, per payroll period to do taxes for you.
So I would do taxes by hand in Excel for every person at the company. It took, like, eight hours every week before I realized better way. Like, AI will just tell me I didn't have to do that. It could take care of it for me. There's so much that it can help solve. So it's accelerating everybody, right? And we haven't talked enough about the individuals in a company. They're also accelerating. The question is, what do you do with that?
I have to ask about sort of all the, um, fearmongering now that's going on about sort of unleashing agents in your organization and having them create chaos or start doing whatever they want to without listening to what leaders want. So tell me your reaction to sort of all the recent news with OpenAI and Anthropic calling for maybe a slowdown in AI development.
So I think we need to differentiate between future model capabilities, the unguarded version of models that are being tested now, and current capabilities. So I think there is enough smart people worried about future model development, and we're still on this exponential curve, that if they're worried about a slowdown, government action or, you know, other action to slow to pace the frontier and to figure out what the right thing to do is and give people time, like, so that we are in the loop as deciding what happens seems reasonable, right?
What form that takes is up for grabs, how risky it is really is kind of up for grabs, but, like, there's no reason not to have some, you know, to have policy to prevent existential risk. I think we're in for a cybersecurity environment that is apocalyptic in the near term. It might be solvable, but I think we are in a for a world where when we get unguarded models, cybersecurity is going to be a real issue.
Within organizations using guarded models, uh, there just isn't the same kind of risk. So I think we need to differentiate between those things.
But how should organizational leaders be thinking about that environment, which you said might become apocalyptic for, you know, rogue agents from outside thwarting their competitive strategy because they're stealing information or causing disruption?
I I think that you should think about it not so much as your as your competitors unleashing things on you to cause trouble. I think it's going to be that that you're going to have to deal with, uh, with security environments where agents, for many, many reasons, are going to be attempting to penetrate your security. I think that when I talk to CISOs, um, I mean, OpenAI pulled 20% of their staff off of research to 20% of the resources to hardening their systems.
I have talked to leaders of large financial institutions who are pulling 50% plus of their CIS- their staff and try everyone's trying to figure out what to do. At the same time, defenses increased, the amount of patching you need to do is increased, which also creates problems. We don't know how this goes through, but I think we're going to have a very disruptive period where the cybersecurity environment is going to get kind of, uh, intense.
And that will look very, you know, like, crazy from the outside. I think that thinking about how you're going to deal with that now is really important. I think you're probably underestimating how difficult that situation's gonna be based on the experts I talked to. I'm not a cybersecurity expert. So I don't think it's so much attackers bothering you as much as every part of your system exposed to the internet is going to be a target for agents.
Maybe accomplishing completely random, benign tasks, right? If I ask an agent's form to price products, they might it might decide to really get good pricing for, you know, for the Christmas because you wanted to buy a, you know, a sweater. I really need to get inside of this, uh, inside of this wool mill's systems because that's the only way I can find out the pricing I need. And relentlessly attack that wool mill. That wasn't an malicious attack.
That was a I just need to find the price per wool that you have in the futures contract market. And I become obsessed with this. So it's gonna get very complicated. We're dealing with kinda alien technology to some extent.
What's the biggest strategic mistake that you see companies making right now as they're trying to figure out what AI means for their business?
So either, uh, there's two giant ones. One of them is fixing on the ability of AI right now. I have this list I have of, like, what the hot thing is in AI, and every three months it changes and everyone goes all in. All of these things are because we're anchoring on what the capabilities of AR- AI are today. So, like, the hot thing that's has been recently is everyone's spending too much on tokens. Well, guess what?
Um, once you start thinking about multiplayer AI, where many people in your organization are using it, token costs become more reasonable. All the AI companies are now reducing the costs of models and having models use dumber agents to do work that are cheaper. Like, everyone who went all in on cost was the main issue. That may or may not be the concern a few months from now. I mean, I can't tell you. It depends on your organization and the choices you're making.
So I think there's a danger in fixating on what you have now and satisfying based on the s to systems and capabilities you have today and not realizing that we're still on the exponential. That is by far the number one problem. The second problem is trusting that other people have got this nailed down. Like, every vendor you talk to is also making this up as we go along. Every consultant you talk to does not have seventeen successful cases, and if they have successful cases, it's for GPT four, a model that's obsolete and no one would use for anything anymore, and your phone's model is more powerful than that.
So you have to build the capacity to make your own solutions in-house. That doesn't mean write all your code or anything else yourself, but it does mean building that capacity internally.
Let's talk about that first problem or mistake that you see. How do you teach yourself to embrace the idea of exponential advancement and be ready sort of for where the puck is going?
So to take an idea from Harvard, from Mike Tushman and company, uh, it it was the idea that you would need ambidexterity. You need to be able to do two things at once. You need to be able to exploit what's happening today, but realize you're still exploring. That's why the lab is so important. Without your own internal flexible team that's building impossible things for the future, you get stuck. So there's a few questions I ask organizations whenever I talk to them.
The first is what have you stopped doing as a result of AI? AI has made something in your organization trivial. What is the trivial thing that used to be really important, but is no longer important? You have to stop doing it. AI just does it. Second thing is what are you building that's impossible? You should be building at least one thing that is unbelievable and that would change the industry. If you're not building that, you're not being ambitious enough.
Even more so, you should be building something that doesn't work yet, because you're betting that exponential curve will keep growing, and at some point it will work. And then the fourth question is, do you have leadership lab and crowd in place? So I think if you have those four questions answered, you're in much better shape for being ambidextrous.
How do you staff your entire organization with the talent that's able to execute on everything you're talking about?
It's also a really hard question. Everything from psychological safety to how do we build the right structure. All of that now is you're cashing all that in. If you have a good structure and a good organization and a good culture and the right reward systems, people in your organization are discovering use cases and they're telling you about it. If you're not, then everyone's hiding all their uses, and nothing you could do can change that because they're scared.
If they tell you, they'll get fired, or they won't get respected, they won't get rewarded. And why would they ever tell you about how they're using these systems? So you need expertise in your organization finding this and people willing to talk about that. And that comes from having a good culture, structure, reward system, you know, openness, good management. Those are more urgent than ever.
You talked about sort of the first questions that you ask leaders when you're beginning to advise them on on how they should be managing this. As you sort of get down the line and they they sort of face challenges, what are sort of the the most common hurdles you see to prevent them from sort of fully moving forward into this new world of sort of AI-led strategy?
I mean, I think that that your danger is in a lot of organizations making this a normal technology. Like, there's a desperate desire to make this like any other technology that we can swallow and we'll turn into a project that has X dollars associated with it and has a and, you know, we have a six month RP process and then a eight month discovery process that, like, that is not gonna work. You're going to experiment. You're going to have to fail.
Like, nobody has answers to your question here. So if you try and do this as a complete de-risked thing, you end up in a position where you end up replicating other large technology projects, and that is going to be obsolete the moment you launch it. And you haven't built in any capability to change what you change. Making sure that you stay hungry for the future is the real problem here.
So it's not sort of an AI adoption or even transformation plan. It's really revolutionizing the way you think about what your business is designed to do.
The one thing I've gotten most wrong about AI in the last year is I thought agents would be harder to pull off than they are. I thought you would to organize teams of agents, I thought we'd need to use all of our organizational design knowledge, and it turns out we don't. There are 100,000 agent swarms that are self-organizing. I didn't expect that to happen, but it makes sense when you think about the reason we have organizations and they're complicated is people have limited knowledge.
People have limited, uh, have different interests in mind. They have conflicts. They don't they don't have the same goal. AI agents share all those things. Coordination turns out to be an easier problem than we thought. And as a result of all of that, suddenly you can get things done with AI in a way you didn't expect. So the problem becomes now on the leadership side, what do I want the organization to do? And if your KPIs and other material are all baked around the idea of human outputs, you don't want a hundred times more PowerPoint.
You don't want a hundred times more code. You want different outputs that are better. So you need to sit back and think about what those outputs and what those goals are and be aiming to maximize those goals. Otherwise, you have internal things and you end up just producing a hundred times more code, a hundred times more PowerPoint. I mean, if there's one lesson I learned from my MBA program before I went and got my PhD, is that people do what they're incented to do.
And you need to think really hard about the incentives in your organization. So the incentives of your organization are minimize the impact of AI, use AI as a productivity tool to do more of what my KPI tells me I should do. It's not transformation.
Let's zoom out and talk about a few bigger picture issues. First, you know, obviously, there's a fear that this will lead to massive unemployment. So what do you see as sort of the future labor market in this new AI-led world?
That is the biggest question and debate among economists about this topic, and we don't have easy answers, right? So there you can easily theorize that, like, the junior market gets squeezed because junior employees don't have the expertise. Why use them when you can use AI? And there's some indications of that, but there's also some indications there may be more hiring because, I mean, you know, now you can use coders to do things that organizations didn't have coders, now can write their own code.
So you start insourcing, and now everybody's doing their own marketing firm. So maybe you lose jobs at, like, large centralized firms, but inside the organization, jobs grow because there's more demand. It's really hard to know, right? Um, you can imagine and the AI labs, to be clear, are aiming for the job apocalypse. They want AI to replace all human labor. Like, that is their explicit goal. I think there's reasons w- that, A, they might be wrong about this, and B, we don't want them to win, so we shouldn't let that be the future that we get.
But, like, that is their goal. So I don't know what the outcome is in the long term in terms of employment, you know, per se, like, what the what the impact on jobs is going to be, and nobody does. So we're watching that closely, and everything is very confused right now in terms of the indicators. We don't really have an answer.
As a technology advances so quickly, and we don't even know what's gonna happen with it, as you said, you know, sort of you have to be developing your strategy dynamically. You have to be prepared for change. You have to experiment and fail. You have to be ready for exponential advancements. How do we get better at predicting not just the opportunities, but also the really, um, terrifying risks?
I I mean, things are accelerating quite quickly. I think it is hard to bounce those things out. I think that this is a general purpose technology. It's many different things at once. There are existential risks associated with it potentially. There's certainly near-term risks. I worry a lot about, um, again, automation of jobs instead of, um, augmentation of jobs. I worry about the the people being taken out of all the interesting parts of their job and their only job is verifying the results of AI content, which would just end be fired if it's wrong, seems like a terrible way to live.
I worry about, um, you know, impacts on our information environment. I mean, at the same time, I'm incredibly powe- and I'm worried about, like, as a tea as an instructor, that people are cheating with AI and then they they don't learn as much. At the same time, we know AI is also an incredibly good tutor. So it can transform teaching for the positive, just like it can for the negative. I think that it's a great freeing aspect for people who are stuck, uh, that they can get more stuff done before.
We certainly know it gives good medical advice, legal advice, things like that. Um, still consult your doctor and lawyer. I'm not trying to tell you not to do that. Um, but I you know, I I think that there is, um, it's a mixed bag. So how do we react to that? We react with policy to encourage good uses and mitigate bad ones. We encourage it by being meaningful about how we're using these systems and not doing them blindly.
We do that by, as leaders, making choices about how these systems are used and working in concert with our team and our organizations to make sure that we're not disempowering people en masse. And I think good leadership will start to show through in these kind of circumstances.
Well, Ethan, thank you so much for wrestling with all these really tricky issues with us. I've learned a lot about AI from you over the years, and especially a lot from your new book. Thanks so much for being with me today.
Thank you for having me.
That was Ethan Mollick, an associate professor at Wharton and author of the forthcoming book, Co-Existence: The Next Phase of AI. Be sure to check out all four of our episodes on how AI is changing leadership at hbr.org/podcasts. If you found this episode helpful, please share it with a colleague and be sure to subscribe and rate IdeaCast in Apple Podcasts, Spotify, or wherever you listen. If you want to help leaders move the world forward, consider subscribing to Harvard Business Review.
You'll get access to the HBR mobile app, the weekly exclusive Insider newsletter, and unlimited access to HBR online. Just head to hbr.org/subscribe. Thanks to our team, senior producer Mary Du, senior production editor Kristen Murphy Romano, and our broadcast team, Dave DeJulio, Ellie Honan, and Simona Sprain. And thanks to you for listening to the HBR IdeaCast. We'll be back with a new regular episode on Tuesday. I'm Allison Beard.