A Debate about Work Design and AI Enterprise Value

[00:00:00] Welcome to The Debate. You know, imagine spending, I don't know, hundreds of millions of dollars building a state-of-the-art nuclear reactor. Okay. And then you just plug it into a single flickering 60-watt light bulb. Right. A massive mismatch. Exactly. The power generation is massive, the capital investment is totally unprecedented, but the actual illumination in the room is practically nonexistent. And that discrepancy, I mean, it perfectly captures the landscape of artificial intelligence in the enterprise world right now.

[00:00:29] Yeah, it really does. We are looking at a phenomenon that is currently just dominating boardrooms, which people are calling the AI value gap. And the numbers are, well, they're staggering. Oh, totally. According to recent Gartner statistics, while nearly every organization is utilizing AI in some capacity, only 13% of chief information officers report that their enterprise has realized, you know, significant or transformational value from it. Which is wild. It is. And the boardroom panic is palpable.

[00:00:56] To add some color to that statistic, 44% of CEOs are now openly struggling to demonstrate any AI-driven financial returns to their stakeholders. Zero returns. Yeah, exactly. Software budgets are ballooning. The agentic tools themselves are objectively incredible. But the fundamental unit economics of the businesses deploying them, they just aren't changing. So the uncomfortable question becomes, when a disruptive technology fails to deliver on its financial promise, is the problem inherent to the technology itself?

[00:01:24] Or is the failure entirely on the shoulders of the managers deploying it? And that tension is exactly the central focus of Jonathan H. Westover's 2026 article in the Human Capital Leadership Review. It's titled, The Redesigned Dividend. Why Most Organizations Still Aren't Getting Value from AI, and How Leaders Can Redesign Work for a Human Agent Workforce. It's a great piece. It is. And the premise forces us into a very distinct corner regarding how to solve this value gap.

[00:01:50] I'll be taking the stance that, without fundamentally tearing down and rebuilding the socio-technical system at work, AI is just a costly bolt-on that disrupts more than it delivers. Right. The structural redesign argument. Exactly. You have to redesign the actual structure of the roles. Well, and I'll be arguing from a perspective I find much more grounded in reality. AI capabilities are highly probabilistic and notoriously jagged.

[00:02:12] So because the technology is so inherently unpredictable, launching a massive top-down structural redesign of your company is, it's entirely premature. Premature? Yes. The real differentiator between organizations that capture value and those that just burn capital isn't structural redesign. It's rigorous, outcome-based governance and strict stopping rules. Well, I firmly believe the AI value gap is fundamentally a work design problem. It is not a technology problem.

[00:02:40] Westover's paper draws heavily on socio-technical systems theory, specifically pointing back to Tristan Bamford's famous 1951 research on British coal mining. Oh, the Longwall Method study. Exactly. If we look at that study, it wasn't just that they brought in new machines. The coal industry introduced highly efficient mechanical equipment, but to use it, they broke up the tight-knit autonomous teams of miners. Right. They isolated them. Yeah. They split them into isolated, specialized shifts on a mechanical conveyor belt.

[00:03:09] The technology was objectively better, but they totally destroyed the human communication loop. Which obviously backfired. Exactly. Morale plummeted, absenteeism spiked, and productivity cratered. And I'd argue we are doing the exact same thing today with generative AI. You think it's a direct parallel? I really do. If you just hand an employee an AI assistant, like a new mechanical conveyor belt, and leave their role exactly the same, the few minutes they save simply disappear into the cracks of their day.

[00:03:37] Unless you explicitly rebuild the human communication loops and the roles themselves for a human agent workforce, you cannot capture what Westover calls the redesign dividend. I mean, I look at this issue from an entirely different angle, specifically a constraints perspective. Okay. The primary barrier to capturing value isn't a lack of structural redesign. It's the psychological escalation of commitment, colliding with the realities of the jagged technological frontier. Okay. Explain what you mean by that.

[00:04:05] Let's unpack the MIT Nanda findings. They reviewed over 300 AI initiatives and found that the vast majority yield zero measurable return. And why? Because the tools fail to adapt to the highly dynamic context of human work. But that's exactly why you need to redesign the work around the tool. Though wait, on top of that, Gartner's data highlights that generative AI runtime costs are probabilistic. Traditional software has a fixed cost, right? Right. You buy a license. Exactly. But AI scales based on token usage and compute.

[00:04:34] If your employees use it inefficiently, your margins just vanish overnight. Massive role redesign assumes that we actually understand the boundaries of what the AI can and cannot do reliably. And you're saying we don't? We absolutely don't. The technology is completely unpredictable. What organizations desperately need first are rigorous pilot constraints and production stopping rules. You have to know how to kill a failing project before you reorganize your entire company around it. Okay, let me challenge this idea that the tech is too unpredictable for structural design.

[00:05:02] Because I think it misses why these early attempts to capture value are actually failing. Go on. The primary reason AI pilots fail is what we call the headcount fallacy. Leaders use AI as a blunt instrument for immediate task substitution, literally just to fire people. Which is a terrible strategy, I agree. Right. Look at Morgan Stanley's debrief tool, which is largely considered the gold standard right now. They didn't roll out this AI to replace financial advisors. They changed the structure of the advisor's job. They automated the busy work. Exactly.

[00:05:32] The AI listens to the client meeting, synthesizes the complex notes, drafts the follow-up communications, and updates the CRM. It saves the advisor about 30 minutes per meeting. That's a massive amount of time. It is. But they didn't just let the advisor go home 30 minutes early or fire half the team. Because they redesigned the workflow, that save time is deliberately reinvested into the human element. They're building deeper client relationships, calling more prospects, doing the empathy-driven work that machines literally can't do. Right.

[00:06:02] That reinvestment mechanism is where the actual revenue comes from. Well, look, reinvesting capacity sounds magnificent in theory, but it completely ignores the very real, very heavy cognitive tax that AI introduces into a system. A cognitive tax? Yes. AI doesn't just seamlessly free up 30 minutes. It creates entirely new demands on the human worker to orchestrate the output. We have to look at the internal data from Anthropik's own economics team. Okay, what did they find?

[00:06:28] They found that when their employees shifted from doing tasks directly to managing multiple instances of their AI cloud, they didn't suddenly have all this free time to innovate. The median employee actually ended up spending 20% more time in meetings. Wait, hold on. You're saying that by giving employees an AI assistant that literally does their work for them, they end up in meetings 20% more of the time? Yes. How does that math work? Because the orchestration itself becomes a massive bottleneck. The agent completes a complex task in, what, five seconds?

[00:06:58] But now? The human has to review it, verify it for hallucinations, check the logic, and then coordinate with three other departments who are also reviewing agent outputs just to make sure everything aligns. Well, sure, oversight is necessary. But it scales poorly. Furthermore, the Anthropik data showed that the cloud sessions were idle 80% of the time, just waiting for human input and approval. This proves that managing AI agents causes immense hidden strain. Human attention becomes the scarce resource.

[00:07:26] A top-down structural redesign doesn't magically erase the fact that verifying and coordinating AI output is exhausting. But, okay, but that strain, the 20% spike in meetings, the bottlenecks, that is exactly why organizations need extreme structural changes. You think more structure solves the fatigue? I'm saying this structure is currently broken. Look, bolting generative AI onto an untouched 2010 workflow is like putting a jet engine on a horse-drawn carriage. You don't just fail to go faster, you actively break the carriage.

[00:07:55] Okay, that's a dramatic analogy. But it's true. Treating AI like a standard software license is like buying a grand piano for your office and assuming everyone is suddenly going to start producing symphonies. You didn't buy a simple tool, you bought a complex instrument, and nobody's workflow is set up to play it yet. So you're blaming the users? No, I'm blaming the design. The Anthropik data doesn't prove that redesign fails, it proves that failing to redesign causes burnout.

[00:08:21] If your people are spending more time in meetings, it means your approval processes, your managerial hierarchies, and your intake cues haven't been updated to handle the sheer velocity of the AI output. So what's the solution then? Well, this is exactly why Moderna took the radical step of completely merging their HR and IT departments into a single unit, led by a single executive. Which is unheard of. Totally unheard of. But they realize you can't have IT handing out the grand pianos while HR manages the musicians.

[00:08:47] They had to merge to, quote, architect the flow of work across both people and agents simultaneously. Well, merging HR and IT is certainly a bold structural move. But I'd argue it still relies on a fundamentally flawed assumption. Which is? That the AI will perform predictably once that new structure is in place. And that brings us to the actual quality of the output, which is where things get truly perilous. We have to talk about the jagged technological frontier. Ah, you're referring to the BCG field experiment by Delacqua and colleagues. I am.

[00:09:17] That BCG study of over 700 consultants mapped out exactly what happens when humans interact with AI capabilities. And when AI is applied to tasks securely within its capabilities, say, drafting a persuasive executive summary, it's phenomenal. Right. Huge productivity boosts. Exactly. Consultants work 25% faster and produce higher quality work. But here's the catch. The frontier of what AI can do is jagged. It is not a straight predictable line. Meaning it has blind spots. Massive blind spots.

[00:09:45] It can write a brilliant sonnet but completely hallucinate the basic logic of a supply chain math problem. And the boundary between those two tasks is completely invisible to the user. Yeah, that's a danger zone. Right. When the BCG consultants applied AI to tasks just outside its capabilities, the edge of that jagged frontier, performance drastically plummeted. They were markedly less likely to reach the correct answer than consultants working completely without AI. You cannot design a perfect workflow when the frontier of what the technology can do is invisible and constantly shifting.

[00:10:14] I mean, I don't disagree that the frontier is jagged, but how does that negate the need for structural redesign? Because structural redesign commits you to a path you can't guarantee. Look at Klarna. They are the ultimate cautionary tale for this. The customer service bot. Yes. They deployed an AI customer service assistant, publicly bragged that it was doing the work of 700 full-time human agents, paused all their hiring, and celebrated the massive cost cuts on Wall Street. And then it all fell apart. Exactly. A year later, what happened?

[00:10:44] The AI couldn't handle the edge cases. Quality dropped so significantly and customers were trapped in such frustrating loops that the CEO had to publicly reverse course. They had to hire again. Rapidly hire remote human workers, again, just to fix the mess. Klarna didn't need a better workflow. They needed a predefined stopping rule. They needed a metric that said if customer resolution drops by 4%, we immediately kill the pilot. That outcome scorecard is infinitely more critical than trying to guess how to redesign the company.

[00:11:13] See, I look at that data completely differently. I don't look at Klarna as a failure of AI, and I certainly don't look at it as proof that we just need better kill switches. How could you see it as anything else? Because Klarna is a textbook failure of role redesign. They engaged naked task substitution. They just swapped 700 humans for a machine to cut costs without preserving the human judgment required for complex issues. But the machine was supposed to handle the judgment. And that's exactly the problem.

[00:11:39] By doing that, they triggered exactly what Lizanne Bainbridge outlined in her famous 1983 paper on the ironies of automation. Remind me what you mean by that in this context. So the core irony of automation is that if you automate away all the routine, easy tasks, like the 90% of the daily work, you leave the human operator with only the most complex, catastrophic exceptions. Right, the edge cases.

[00:12:02] But because the human is no longer doing the routine work, they lose the daily practice and the tacit knowledge required to actually handle those complex exceptions when they inevitably pop up. The system collapses under its own weight. Which is what happened to Klarna. Yeah. Precisely. Klarna automated the easy stuff, got rid of the humans, and when the jagged edge of the AI's capability failed, there was no human system in place to catch it. Contrast Klarna with Ikea. Oh, with her bot Billy. Yeah.

[00:12:27] Ikea had an AI bot Billy that successfully deflected 47% of routine customer queries. Where's my package? What are your hours? Stuff like that. Right. But instead of firing the humans used to answer those questions and celebrating a cost cut, Ikea retained 8,500 call center workers and structurally redesigned their roles. They trained them to use 3D spatial planning software to become remote interior design advisors. That is a massive pivot. It is.

[00:12:53] They used the time freed up by the AI to have humans walk customers through complex kitchen renovations. They turned a cost center into a massive revenue generator that brought in 1.3 billion euros in 2022. That is the redesigned dividend. I mean, that's an impressive case study, for sure. So I have to push back on your reliance on stopping rules. Aren't pilot stopping rules, which Gartner knows only 31% of CIOs even have, just a band-aid for fundamentally bad organizational design?

[00:13:21] If you design the rule correctly from the start like Ikea did, you don't need to kill the pilot. Calling stopping rules a band-aid fundamentally misunderstands corporate psychology, if I'm being honest. Really? How so? Stopping rules are the bedrock of psychological safety and financial discipline in an unpredictable environment. The escalation of commitment is a very real psychological trap. Sure. The sunk cost fallacy. Exactly. When a senior leader publicly champions a million-dollar AI pilot, their reputation is on the line.

[00:13:50] When it starts failing, when it hits that jagged frontier, they won't stop it. They will keep throwing money at it, rationalizing that it just needs, you know, a slightly better model or more clean data or more employee training. Yeah, I've seen that happen. We all have. Without a strict predefined stopping rule agreed upon before the pilot even launches, you end up scaling a system that actively bleeds money.

[00:14:13] And speaking of bleeding value, we have to look at the metrics organizations are using to try and force this redesign because they are creating dangerous activity traps. Activity traps? Activity traps? How are metrics creating activity traps? Consider Shopify. They recently mandated that employees must explicitly demonstrate why they cannot accomplish a piece of work with AI before they are allowed to request more hit count. Oh, wow. Yeah. They even added AI usage targets to their performance reviews. Now, on the surface, you would probably argue that looks like a great mandate for redesign, right?

[00:14:42] I mean, yeah. It forces the employee to rethink how they accomplish their daily tasks. But let's look at what actually happens. If you're listening to this and your company just mandated AI usage targets for your quarterly bonus, you already know what happens next. You game the system. You game the system. You're going to log in, prompt the AI to summarize an email chain you've already read and log out just to hit your KPI. If you tell employees their bonus depends on AI usage, usage metrics will naturally skyrocket. The dashboard looks green. But it's meaningless. Entirely.

[00:15:11] Does that actually move the needle on the business's unit economics? No. It just means people are prioritizing logins over actual business outcomes. It is an activity trap. And the long-term consequences of this forced automation are terrifying for the future talent pipeline. Okay, I assume you're looking at the impact on entry-level knowledge work. Exactly. Let's go back to that anthropic internal data. As they integrated AI heavily into their own coding workflows, the share of junior technical hires they brought on dropped from 27% all the way down to 9%.

[00:15:41] That's a huge drop. The talent pipeline practically collapsed. If we blindly redesign workflows to eliminate all entry-level work because an AI agent can do it cheaper and faster, where do the senior experts come from in 10 years? That's the million-dollar question. Right. How do you become a senior developer if you never struggle through the junior developer tasks? You can't structurally redesign a company if you are simultaneously suffocating the very mechanism that produces tomorrow's experience to professionals. Look, that is a very real concern.

[00:16:09] And I absolutely don't want to minimize the risk of the talent pipeline. But I'm not convinced by that line of reasoning because it assumes that workflow redesign inherently means eliminating learning. How does it not? If the machine does the basic work, the human doesn't learn it. Let's look at the empirical evidence of how AI actually impacts learning when it's deployed correctly. Eric Brynjolfsson and his colleagues conducted a massive study of over 5,000 customer support agents. Okay.

[00:16:35] They found that providing access to a generative AI assistant didn't hollow out the junior staff. It actually accelerated the learning curve for lower-skilled and novice agents by 15%. Wait, but how? If the AI is doing the work, what is the junior agent actually learning? They are learning the tacit knowledge of your best employees. The mechanism is absolutely fascinating. The AI ingested the chat logs of the top 10% of performers.

[00:16:59] It figured out their phrasing, their emotional intelligence cues, their specific problem-solving pathways. Okay, so it learned the patterns of the experts. Right. Then it fed those prompts to the junior workers in real time while they were on calls. It acted as an instant, personalized mentor. But here is the critical point. That only works if you deliberately structure the work to allow for it. There's that word again. Structure. Because it matters. Westover explicitly calls out the need to protect learning tasks.

[00:17:28] If you just leave AI deployment up to middle managers who are trying to hit quarterly cost targets, yes, they will use AI to cut junior headcount. Which is what almost everyone is doing. But they shouldn't. If you have a top-down human agent workforce plan, you can deliberately preserve specific tasks purely to build expertise for early career employees. You design the roles so that AI becomes a tool for rapid upskilling rather than a cheap replacement for your junior staff. The technology itself doesn't hollow out your company.

[00:17:56] Lazy organizational design does. Look, I hear your passion for this. But calling it lazy organizational design entirely dismisses the staggering complexity of what business leaders are actually facing today. I don't think I'm dismissing it. You are asking them to perfectly architect a socio-technical system around a technology that fundamentally reinvents its own capabilities every 6 to 12 months. Yes, you have to adapt. But when the capabilities are a moving target, over-designing the workflow is a fool's errand. You'll spend a year mapping a workflow for a human-agent interaction,

[00:18:24] only for the next iteration of the model to render that entire workflow completely obsolete. The safest, most effective route to value isn't a theoretical top-down redesign. It is robust, defensive governance. You track unit economics relentlessly, you measure customer outcomes rather than internal usage metrics, and you implement strict stopping rules. The kill switches. Exactly. If the AI doesn't hit the predefined business metric, you kill the pilot without hesitation. It's that financial discipline, not structural redesign, that prevents the AI value gap from widening.

[00:18:52] Well, I think we have reached a very clear distillation of our two perspectives on this incredibly complex issue. From where I sit, the historical precedent and the current data are just overwhelming. In favor of redesign? Absolutely. Overcoming the J-curve of productivity, you know, that initial period where new technology actually hurts output before it helps, requires brave, deliberate investments in intangible assets. AI does not produce value simply by being present in an organization. You agree with that part, at least.

[00:19:22] Right, but you cannot capture the redesign dividend without doing the hard, messy work of mapping tasks, rebuilding human handoffs, and fully redesigning workflows end-to-end for a human agent workforce. And I maintain that in a landscape defined by highly probabilistic runtime costs and a constantly shifting, jagged technological frontier, assuming you can perfectly architect a workflow is pure hubris. Hubris. It is. The organizations that will actually survive this transition

[00:19:51] and capture real value are the ones that prioritize rigorous governance. They implement strict stopping rules to counteract the psychological escalation of commitment, and they relentlessly measure actual business outcomes rather than just cheerleading AI usage. Well, despite our fundamentally different approaches to solving the value gap, I do think there is a profound point of convergence here. Which is? We both entirely agree that the bolt-on deployment strategy, the act of just buying expensive software licenses,

[00:20:21] handing them out to employees, helping for the best, and subtly threatening headcount reductions, is a guaranteed path to organizational failure. Oh, unquestionably. Treating artificial intelligence merely as a fast-track to payroll reduction without addressing the underlying mechanics of the work itself is toxic. It destroys company culture, and it ultimately degrades productivity. It really is an incredibly fascinating challenge for modern leadership. The sheer complexity of these socio-technical systems cannot be overstated,

[00:20:50] and I highly suggest our listeners explore Jonathan Westover's material to think more deeply about the deliberate design of their own daily work. It's definitely worth a read. Because at the end of the day, if you plug a billion-dollar nuclear reactor into a 60-watt light bulb, you aren't just wasting power. You might just blunt the entire circuit. Leaving everyone in the dark. Something to think about next time you log in and ask an agent to summarize your inbox. You might just be able to see your inbox.