The Problem With AI Job Predictions Isn't AI

Something interesting has happened over the last few weeks: the people who were most confident that AI would rapidly eliminate large numbers of jobs have become noticeably less certain. Sam Altman has softened some of his earlier warnings. Dario Amodei has become more nuanced in how he discusses labor market disruption. Meanwhile, despite extraordinary advances in the technology itself, the evidence for widespread white-collar displacement remains surprisingly thin. The gap between what many expected to happen and what has actually happened is becoming difficult to ignore. Recent reporting from both Fortune and MIT Technology Review points to the same phenomenon: AI capabilities continue to improve at an extraordinary pace, yet the labor market has not responded in the way many of the loudest forecasts anticipated.

There are several possible explanations. Technological adoption often takes longer than enthusiasts expect. Organizations tend to change more slowly than technology itself. Economic effects are rarely distributed evenly. Yet this may be less a story about AI than a story about how we think about work.

One explanation appeared recently in the Financial Times, and I think it gets closer to the heart of the issue than most of the debate has. The article argues that we tend to think about jobs as collections of tasks rather than collections of responsibilities. It sounds like a small distinction, but it isn’t: tasks are generally visible; responsibilities are not.

If you ask someone what a recruiter does, they might say reviews resumes. If you ask what a lawyer does, they might say drafts contracts. Ask what a software engineer does, and most people will say writes code. These answers are not wrong, but they are incomplete. They describe observable activities rather than the reason the role exists in the first place.

A recruiter is ultimately responsible for helping build effective teams. A lawyer is responsible for reducing risk and creating clarity. A software engineer is responsible for solving problems within a complex web of technical, organizational, and business constraints. Writing code, reviewing resumes, and drafting contracts are some of the ways those responsibilities are fulfilled, but they are not the responsibilities themselves.

The Doorman's Fallacy & Other Psychological Wrinkles

This distinction reminds me of something economists sometimes refer to as the Doorman's Fallacy. A doorman appears to spend the day opening doors. If opening doors is the job, then automation seems straightforward: just install an automatic door and move on. Except opening doors is rarely why buildings employ doormen. A good doorman provides security, recognizes residents, manages unusual situations, coordinates deliveries, answers questions, and often serves as a source of continuity for an entire building. Opening the door is simply the most visible aspect of a much larger role. The mistake is assuming that because an activity is visible, it must also be central.

The same mistake appears throughout the AI conversation. Consider software engineering, the profession that has become the poster child for AI disruption. Much of the discussion assumes that writing code is the job. But writing code is one activity software engineers perform. The actual job includes understanding tradeoffs, making architectural decisions, translating business needs into technical requirements, managing complexity, anticipating future problems, and navigating constraints that often have little to do with technology itself.

This helps explain why organizations can simultaneously report that AI is generating more code than ever while struggling to point to equivalent gains in product velocity or business outcomes. More output and more value are not necessarily the same thing.

There is also a different psychological wrinkle here. Most people understand the hidden complexity embedded within their own work while seeing only the visible activities embedded within everyone else's. Ask an engineer about engineering, and they will describe tradeoffs, ambiguity, stakeholder management, and judgment. Ask that same engineer about recruiting, and the work may suddenly appear far more procedural. Recruiters often make the same mistake about finance. Finance professionals make the same mistake about marketing.

Everyone sees the iceberg beneath their own profession; everyone else sees the tip. This asymmetry creates a predictable bias. We tend to believe AI is more capable of replacing work we do not understand than work we do. It is one reason conversations about automation often feel so detached from the lived experience of the people actually performing the jobs being discussed. This may also help explain why so many organizations are finding AI simultaneously impressive and frustrating.

The technology does not always work as advertised (Starbucks recently decommissioned an inventory tool that hallucinated), but on the whole, reports are generated faster, research can be completed more quickly, code can be produced in greater volumes, and documentation becomes easier to create. Yet many companies find themselves confronting the same organizational constraints they faced before adopting these tools: decisions remain slow, priorities remain unclear, coordination is still difficult, and strategic alignment remains elusive.

In those situations, the problem is rarely technological because even if the technology succeeds, the surrounding system remains unchanged.

This distinction becomes particularly important when discussing entry-level work. Much of the anxiety surrounding AI has focused on the possibility that junior roles could disappear as repetitive tasks become increasingly automated. The concern is understandable, but it is also incomplete. Entry-level roles have always served purposes beyond immediate productivity. They are part of the developmental infrastructure of an organization. They provide context, they train judgment, expertise, and process.

When we reduce entry-level roles to the tasks they perform, we risk overlooking the developmental function they serve. None of this means AI's impact on work has been overstated. In some respects, the opposite may be true. In fact, the long-term effects could prove more significant than many current forecasts suggest. What may be changing, however, is the nature of the transformation itself.

The most important shifts may not come from eliminating jobs outright, but rather from forcing organizations to rethink how work is structured, how judgment is developed, how accountability is assigned, and how value is created. So, what should founders and executives actually do?‍

Audit Responsibilities, Not Tasks

Many AI initiatives begin by asking which activities can be automated. It is an understandable starting point, but often the wrong one. A more useful exercise is identifying the responsibilities that each role ultimately owns and then examining how AI changes the way those responsibilities can be fulfilled. Leaders who focus exclusively on tasks often end up automating activity while leaving the underlying constraints untouched. Leaders who focus on responsibilities are more likely to redesign work itself.

Reevaluate What You Are Measuring

One of the most revealing questions an executive can ask is whether increased activity is producing better outcomes. If engineers are generating more code, marketers are producing more content, analysts are generating more reports, and customer success teams are responding faster, but the business itself is not improving proportionally, it is worth asking whether output was ever the primary constraint.

Many organizations are discovering that their bottlenecks are not as production-related as they expected. They are prioritization bottlenecks, decision-making bottlenecks, coordination bottlenecks, or managerial bottlenecks. AI may not solve those problems, but it has a remarkable ability to expose them.

Protect the Developmental Pipeline

Organizations should think carefully before eliminating junior roles without creating alternative pathways for learning and development. The efficiencies generated by automation are real. So is the long-term value of developing judgment, context, and expertise. Companies that aggressively optimize away entry-level opportunities may eventually discover they have weakened the very pipeline that produces future leaders, managers, and technical experts. Five years from now, the challenge may not be finding enough AI tools, but rather finding enough people who understand how increasingly complex organizations actually function.

In conclusion

The debate around AI and employment has largely focused on what the technology can do. That is certainly an important question. Increasingly, however, it seems the more important question is whether organizations understand their own work well enough to redesign it intelligently. The organizations that benefit most from this transition will not necessarily be the ones that automate the largest number of tasks. They will be the ones who develop the clearest understanding of why those tasks existed in the first place.