The Bad Math of AI Job Displacement

This special edition is coauthored with Alec Levenson of USC Marshall Center for Effective Organizations

Since the release of ChatGPT, there has been a constant drumbeat of loud voices forecasting massive job displacement. The viral chart from Anthropic last month is a clear indication that AI anxiety is wide-spread.  AI tools are developing incredibly fast, and we are very early in the game – too early for anyone to forecast accurately what the final tally of displaced jobs will be.

Source: Anthropic report, March 2026

Yet it is not too early to call out the misguided assumptions used by the doomsayers. Their assumptions suffer from some pretty major flaws. Which means their forecasts of job displacement are likely overly excessive, by potentially large degrees. 

The problems are threefold:

1. Indivisibility: you can’t carve up most people’s jobs just because AI can automate specific tasks
2. Additivity: you can’t just insert AI into the workflow without taking account of how the humans do their work
3. Transitivity: substituting AI for humans can degrade the overall customer experience, leading to negative effects on revenue, customer satisfaction, etc.

 The basic problem of the doomsayers is they focus too much on the ability of AI to automate specific tasks, while ignoring the challenges of shifting work within business processes from humans to AI. The excess focus at the task level discounts the essential role that humans play ensuring everything stiches together properly within the business processes. That critical role will not disappear even in the face of largescale AI disruption.

Indivisibility: You can’t arbitrarily use AI to eliminate the humans in the loop

We have already seen significant job disruptions in occupations such as software programmer and language translator. Other knowledge-centric occupations in law, finance, accounting, and many other disciplines are starting to see largescale introduction of AI. These clear disruptions are the reason for the doomsday forecasts of massive job destruction.

The problem with the forecasts is you cannot project job losses by taking the percentage of tasks automated within a job times the number of people working in the role, across the economy. Which, for the most part, is what the doomsday forecasters are doing. The reason this is bad math is twofold:

• The divisibility assumption: It assumes that jobs are easily divisible and reconfigurable
• The large organization illusion: It assumes what applies for large organizations applies for medium and small sized organizations, which is a corollary of the divisibility assumption

The divisibility assumption: The divisibility assumption is the most egregious. Just because AI can automate, say, 80% of software code writing does not mean 80% of software developer jobs will disappear. As we have seen, the shifting of work from human to AI redefines much of what the human does, away from creating something new, towards reviewing what the AI creates.

Rather than get rid of 80% of the human hours, the actual figure will be much lower. Will it be 60%? 40%? 20%? We don’t know yet. And the lower cost work done by AI means companies can produce more and/or better software for the same price, which increases the demand for all inputs, including the human ones. This will also reduce the net displacement of humans in the system.

Take AI in recruiting for example, while AI can screen thousands of résumés in seconds and rank candidates, it cannot read the room in a final-round interview or navigate the politics of a difficult hiring committee. The human in that loop is doing more than just adding a "review" step, they are providing judgment, context, and accountability that AI cannot replicate in hiring. It is way too early to know precisely how much additional human labor is needed to do the reviewing. But it’s certainly a large portion of the 80% which appears as if it should be made obsolete by AI.

The large organization illusion: Most people do not work in the biggest companies. Smaller to medium sized organizations have few to no people doing the exact same job. Which means they lack the scale of operations that could allow reconfiguring work among multitudes of people doing the same job – a core requirement for AI enabling headcount reductions in large numbers.

Even in the largest organizations where some roles have large numbers of people doing approximately the same thing, most work is done by people working in different roles from each other. So only a minority of jobs, even in the biggest organizations, in theory could be offer major headcount reductions as a result of AI productivity gains. And that’s not counting for the divisibility problem (described above).

Additivity: You can’t ignore the negative spillover effects from adding AI

The decision to add AI cannot be made solely based on AI performing some tasks more cheaply. Why? Because of the negative spillover effects from the AI onto the humans’ work.

What matters is not the reduced cost of AI doing certain tasks. Instead, most important is the total cost of production, as any introductory economics course teaches. The total cost of production, in turn, depends on what happens to overall productivity across all inputs and tasks. Which means we have to account for any spillover effects from AI onto the tasks that the humans can uniquely do – especially negative spillovers.

The most commonly cited these days is the increased time humans now have to spend reviewing and correcting the AI’s work. This is not some minor, insignificant cost that will magically disappear with more experience working with AI. It’s a fundamental challenge that could make or break the economic case for introducing AI in the first place.

Here is a data point that should be on every leader's radar: knowledge workers using AI completed tasks significantly faster and produced measurably better results, but only for tasks that fell within AI's "jagged frontier" of capability (2026 Organization Science article). Step outside that frontier, and performance actually dropped, even with AI assistance. In practice, integrating AI isn't going to yield a smooth and uniform productivity gain across your organization. Employees need to learn where the edges are and that learning takes time, energy, and human judgment to navigate. This real cost is overlooked in the simple task-automation math.

Additionally, AI will not uplift everyone equally. The same study finds that workers below the average performance threshold improved the most by using AI. If AI is going to compress skill gaps rather than widen them, the workforce story becomes a lot more nuanced and complicated. The simple math of AI job displacement doesn’t apply in this instance.

Don’t get us wrong: we know that, ultimately, AI will enhance productivity and become an integral part of how much work is done in organizations. But the steep learning curve and potentially large adaptations needed by the humans in the system will substantially slow down adoption, and potentially limit the upper bounds of AI-driven efficiencies and increased profits.

Transitivity: You can’t substitute AI for humans without risking the customer experience

The products and services at the heart of the vast majority of industries and business models are not simple to produce. Nor do customers buy them only because of their cost.

This means that customer value is not created solely by lower cost delivery. The implication: focusing only on the cost benefits of AI runs the risk that other product attributes will be degraded by haphazard AI introduction, worsening the customer experience and driving them into the arms of your competitors.

This is especially true when AI takes the place of human interactions for the customer, or degrades how your people interact with your customers. History has numerous examples of new technologies that offered to replace humans at lower cost … only to have customers refuse to buy, leading to the technology being sidelined. Examples include:

• Player pianos: Over a century ago, the introduction of player pianos threatened an entire industry of musicians. Today, live piano players dominate, even in lower-cost venues like a local bar. Player pianos today are a curious historical relic.
• Self ordering at restaurants: The technology to have customers place orders themselves has existed for years. Yet despite some restaurants introducing self-ordering, the prevalence is quite rare – and only a few human roles even in those restaurants have disappeared, a very mild disruption compared to the worst case scenario.
• Automated answering systems for customer calls. The telephone switchboard operator is no longer. And we all struggle with automated answering systems that force us to choose among a number of options before reaching a human. Yet ultimately reaching a human is what happens in the overwhelming majority of cases. Full customer self-service only exists in theory on the internet, not for dial-in issues. And even then, most websites offer an option to chat live with a human or to be connected to talk to a human over the phone.

We're watching a modern version of this play out right now with AI customer service. As companies race to replace human agents with AI chatbots in the name of cost reduction, many are seeing customer satisfaction scores take the hit. The Qualtrics 2026 Global Consumer Experience Trends Report shows 13% who have used AI for customer service saw no benefit from the experience.

None of this means AI disruption isn't real. It absolutely is, and both individuals and organizations that ignore it will fall behind. We are pointing out the big difference between the actual disruption taking place, versus the doomsday scenario the loudest voices keep predicting. The organizations that will thrive are already working on skills-based workforce planning and investment in their people's ability to adapt. That strategy will enable them to capture AI's upside without sacrificing the human judgment and customer experience their business models and bottom lines require.

References:

• Fabrizio Dell’Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, Karim R. Lakhani (2026) “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality.” Organization Science.
• Maxim Massenkoff and Peter McCrory (2026). “Labor market impacts of AI: A new measure and early evidence,” Anthropic, (https://cdn.sanity.io/files/4zrzovbb/website/2b5bbaf2c1eb81dbf6e6fb813c1a24e35a64d376.pdf)
• “Qualtrics 2026 global consumer experience trends revealed” https://www.qualtrics.com/articles/customer-experience/global-consumer-experience-trends/

ICYMI:

• I got to celebrate my book's first birthday in several cities and stopped by Taipei last week - the city where it all began!
• I had the pleasure of giving a new talk "Using Al to Turn Well-Being Data into Predictive Safety Intelligence" in D.C. at Informs Analytics+ and enjoyed nerding out with practitioners.

Coming up next:

I can't wait to speak at this year's SIOP in New Orleans with several fantastic I/O friends and learning from everyone else. It's my first SIOP (IKR) so if you are attending and have suggestions for this SIOP newbie, please reach out!

Dr. Serena H. Huang, Founder of Data With Serena, Keynote Speaker & Author

After 10+ years in corporate America and on keynote stages, I have learned one thing above all: data brings credibility, but stories move people. My work lives at the intersection of both.

I help F500 leaders design and deploy AI strategies that drive measurable business performance without destroying what makes their organizations work. In my keynotes, I translate complex AI and business topics into narratives that inspire audiences by combining analytical rigor and human storytelling.

My experience leading People Analytics and AI at PayPal, GE, and Kraft Heinz spans early pilots to scaled deployment across tens of thousands of employees, with over $100 million in measurable impact. I specialize in the hard middle of transformation: turning promising tools into systems people actually use, trust, and stand behind.

I am the author of "The Inclusion Equation: Leveraging Data & AI for Organizational Diversity and Well-being" (Wiley, 2025) and my work has been featured in Fast Company, Barron’s, and CNET.

My mission: Make AI work for people, not the other way around.