Inside the Invisible Cage: What Dr. Hatim Rahman Taught Us About Why AI Implementations Fail

On August 14, 2026, Revival Group hosted Dr. Hatim Rahman — Pepsi Co-Chair in International Management and Associate Professor of Management and Organizations (and Sociology) at Northwestern University's Kellogg School of Management — for a team consult on the real-world challenges of AI adoption. What followed was one of the most practically useful conversations we've had about why AI implementations go wrong, and what consultants and business leaders can do differently.

This post draws directly from that discussion and from Dr. Rahman's research, including his book Inside the Invisible Cage: How Algorithms Control Workers, published by the University of California Press in August 2024.

Dr. Rahman developed the "invisible cage" metaphor while conducting longitudinal research on high-skilled workers in the gig economy — on platforms like Upwork and Fiverr. The metaphor captures something most enterprise AI rollouts quietly replicate. In his words: "These algorithms feel like an invisible cage because they're opaque. Workers don't know what's going into it, how it's being processed, the outputs. And it's a cage because it increasingly controls opportunities without their say."

The cage is invisible because organizations can alter the rules for success — at scale and speed — with no explanation. It is a cage because the system dictates behavior, opportunities, and livelihood without the consent or agency of those inside it.

Rahman was careful to note this is not a condemnation of AI. These tools can be used to empower workers, to surface more equitable outcomes, to create mutually beneficial optimization. The problem is not the technology. It is the organizational choices made around it.

Why top-down rollouts backfire.

In his research — and in several high-profile corporate cases — Rahman has documented what he and his co-authors call cascading inequalities: design-phase decisions that ripple downstream to harm workers and organizations in ways that weren't anticipated.

The mechanism is straightforward. When AI systems are deployed without involving the people who actually do the work, the assumptions baked into those systems are wrong in ways no one catches until damage is done. The result is not just worker dissatisfaction — it is organizational inefficiency.

Dr. Rahman cited a recent and widely reported example: Meta's use of opaque algorithms to determine layoffs, which reportedly resulted in the company needing to rehire many of the same individuals. The algorithm optimized for something, but not for the right thing — because the people who understood the work were not in the room when the system was designed.

This pattern is consistent with independent research. A 2025 enterprise survey synthesized by MIT found that up to 95 percent of AI initiatives fail to produce measurable ROI or sustained business impact, despite tens of billions in investment. Separate Prosci research surveying over 1,100 professionals found that 63 percent of AI implementation blockers stem from human factors — change management, employee buy-in, and culture — rather than technical deficiencies.

The innovation fallacy.

We also discussed a concept Rahman calls the innovation fallacy: the mistaken belief that what a technology can do automatically determines what it will do to workers.

The reasoning goes like this. AI can read a radiology image, therefore radiologists are obsolete. AI can write code, therefore developers are obsolete. This logic is seductive, but decades of organizational research say it is wrong. As Rahman put it: "It is the organizational, the situational factors that determine a lot."

He pointed to the radiology case as an illustration. Computer scientists were predicting the end of radiology as a profession by 2020 or 2025. The reality: researchers at Northwestern published, roughly a year or two ago, the first study showing how AI in radiology can be implemented sustainably to improve both physician and patient outcomes. It took approximately a decade to get there.

The conditions were ideal — highly trained physicians with strong motivation, genuine enthusiasm for AI assistance (they were "inundated with work"), and good job security. And still it took a decade, because no one could shortcut the work of figuring out how to redesign workflows to take advantage of both human and machine capability.

When organizations skipped that work, radiologists were left asking questions like: what does an AI's 90 percent confidence score mean? Does it mean I should investigate that 10 percent more? Does it mean my diagnosis is wrong? That ambiguity didn't reduce work. It created more of it.

What this means for AI consultants.

For those of us helping small, medium, and enterprise businesses adopt AI agents, Dr. Rahman's advice was direct. Four things stood out.

First, start with first principles, not the technology. "As exciting as these tools seem," he said, "start with first principles — start with what are you trying to accomplish, what are your goals." Too much AI consulting begins with the solution and works backward. The better discipline is to define the problem first. Are you trying to improve efficiency? Reduce worker turnover? Improve customer response time? AI is not a hammer looking for a nail, and your clients deserve a consultant who won't treat it like one.

Second, involve domain experts and workers early. The people who understand the work must be in the room when workflows are redesigned — not as an afterthought, but as co-designers. This is especially true for the implementation phase. Workers with deep expertise are the ones who can identify where an AI output is ambiguous, where a confidence score is misleading, or where an automated step breaks a workflow that looked clean on paper. Rahman's recommendation is not just ethical, it is practical. Organizations that involve workers early avoid the cat-and-mouse dynamic where employees resist or work around AI systems because they feel imposed upon rather than supported.

Third, transparency reduces anxiety and improves outcomes. One of the most underrated insights from Rahman's research is that being open about AI usage reduces organizational friction, even when the news is uncertain. "I think it's okay for organizations and leaders to say: we don't know exactly what this is going to look like. We hope this leads to better outcomes. But we need your help. We need your input." When organizations don't tell workers what they're doing or why, workers fill the gap with anxiety and guesswork, often assuming the worst.

Fourth, don't confuse AI activity with AI value. Using AI to generate more emails, more slides, and more documentation is not the same as creating organizational value. Rahman was pointed on this: businesses should carefully ask whether these outputs actually serve their goals, or whether they're just producing more noise faster. For SMBs especially, the path to ROI is not adoption volume. It is disciplined experimentation, led by people capable of learning what works and adjusting when it doesn't.

The contrarian take: AI's impact will be gradual.

One of the most useful contributions Dr. Rahman makes to this conversation is his willingness to offer a contrarian perspective on the pace of AI's macro impact.

Where the dominant narrative insists transformation is imminent and total, Rahman argues the full macro impact of AI will likely be gradual — similar in arc to the introduction of the internet or search engines, both of which took years to fully restructure industries. He contrasted AI with technologies like air conditioning, which delivered benefits without requiring extensive user training or behavior change. AI is the opposite. It requires constant skill development and adaptation, which fundamentally slows the timeline for realized benefits.

This does not make AI unimportant. It makes the pace and method of implementation the most strategically important variable, which is precisely where good consulting adds value.

Learn before you automate.

Rahman's advice for students — that they should master fundamentals before relying on automation, just as pilots must learn to fly manually before engaging autopilot — applies directly to organizational AI adoption.

The businesses that will get the most from AI over the next decade are not necessarily the ones who adopt first. They are the ones who invest in understanding their workflows deeply enough to know where automation adds value and where it creates risk. That understanding cannot be outsourced to the AI itself.

The invisible cage is not an inevitability. It is the outcome of a specific set of organizational choices: deploying AI without worker involvement, without transparency, without starting from defined goals, and without the discipline to redesign workflows rather than just layering automation on top of broken ones.

For AI consultants, the lesson is clear. The technical implementation is rarely where implementations fail. The harder work, and the more valuable work, is the human and organizational layer. Getting that right is what separates AI projects that produce real outcomes from the majority that don't.

Dr. Hatim Rahman is Associate Professor of Management and Organizations at the Kellogg School of Management, Northwestern University. His book, Inside the Invisible Cage: How Algorithms Control Workers, is published by the University of California Press. Quotations are drawn from his August 14, 2026 team consult with Revival Group.