By now, you may have already read that the AI job apocalypse is probably not going to happen. This view is gaining mainstream popularity over the past month, with Sam Altman walking back his earlier predictions, articles by the NY Times, and even AI investors like Andreessen Horowitz weighing in. **

**So what changed? Perhaps it's simply the recognition that job apocalypse narratives are creating a marketing sentiment that is actually hurting adoption in their own industry. Today, Gen Z's are publicly booing talks about AI, and optimism about AI is very low. People are avoiding AI, because why would anyone want to train their own replacement? **

As someone who works in helping enterprise organizations adopt AI, threats of AI replacement were always clearly overblown. That's because if you look at the actual job and adoption data, AI isn't actually replacing humans — it's changing how humans work.

Are you sure that engineering jobs will reduce?

This follows a predictable pattern of technology adoption across history, now popularly called the Jevons Paradox, where the introduction of technology counterintuitively increases the work and jobs of the people it was meant to replace. **

An early example was when the first ATMs were introduced in the US. Everyone predicted the job of the bank teller would reduce by up to 75% in the coming years. But not only did the teller job not reduce, the number of tellers in the three decades after ATM introduction increased 2X, faster than the average job in the US.

Why? When technology makes the underlying service easier and less expensive, people apply it more. In the years following the introduction of ATMs in the 1970s, the number of bank branches increased because ATMs made it cheaper to open a branch. Similarly today we see the proliferation of software and AI into every field — people everywhere developing their own apps and websites, even those with no coding background.

And as the technology takes over one task, human roles evolve to take over the tasks that technology cannot do. Tellers evolved into customer service and answering complex questions. In software engineering, around 30% of code at Microsoft and Google is now AI-generated, and the engineer today is similar to a software architect — reviewing code, understanding what needs to be built, making the engineering systems decisions that AI cannot. And as any vibe coder will tell you, they always need to go to a real engineer before deploying their apps to production.

Sure the number of tellers per branch did fall by 30%, similar to how fewer engineers may be required for the same team today, but since the overall demand for bank branches went up, the total number of teller jobs went up as well. With the enhanced demand for software and AI, the outlook for software engineering seems positive as well.

And the data shows this: although roles with entry-level requirements and highly rules-based “production” tasks have dropped over the past few years, roles with orchestration and judgement — owning the workflow — have increased and are actually being paid more.

Why AI capability alone doesn't predict how AI will impact work

Software engineering is the most extreme example of how significantly AI can impact jobs, because AI is very capable at the core of the job — coding, a highly technical task with a rigid set of rules and clear expectations of good and bad. But capability across other types of tasks may be quite different.

For those in plumbing, equipment repair, or other skilled trades, AI simply cannot replace the core tasks of these jobs. Physical AI has made significant progress over the past year, but it will unlikely deal with the ambiguity and dexterity required for these roles in the near future. Similarly, AI will never be capable of tasks involving real interpersonal human skills. **

But AI capability is still not the only factor. For instance, consider customer service and mental therapy — AI is highly capable at both. It can talk like a human, often replies with more empathy than the average person, and can take orders and process them effectively without any lunch breaks.

And yet, Taco Bell paused the rollout of its AI drive-throughs after an ambitious AI rollout that brought the technology to 500 drive-through locations. People complained about errors, and some deliberately sabotaged the AI by ordering 18,000 cups of water.

And despite AI's prowess at responding with empathy, research suggests human therapy consistently outperforms AI interventions — humans overwhelmingly prefer talking to another human for therapy. (For good reason — therapy is more about reaching a personal connection with someone who pushes back at the right time, a human connection AI cannot simply replace.) **

Anthropic's own data shows AI usage lags behind theoretical capability in most jobs — for computer and math workers, AI is theoretically capable of handling 94% of their tasks, yet Claude currently covers only 33% of those tasks in observed professional use. Adoption barriers explain why. **

The PRIME Framework for the biggest barriers to AI adoption

Policy — Are there government regulations preventing adoption, or would using AI threaten longstanding governmental or societal policies? As an example- If AI self-driving trucks were to be introduced tomorrow, would our societal policies and regulations around driving let it?

Resources — How truly expensive is it to deploy AI in a role? Companies are quickly discovering AI is not that cheap. __Uber burned through its entire 2026 Claude budget in just four months, __with some executives now realising that humans are cheaper than AI at work

Inertia — Incorporating AI at work requires culture change. How quickly will a company integrate new tech like AI into its processes, systems, and workflows?

Market Acceptance — As discussed in the examples above, will people and consumers accept what AI is producing?

Error Tolerance — AI still hallucinates, and may always continue to. How tolerant is that job or task of errors? Remember that people are much more accommodating of errors by humans and demand of technology, especially if the work has serious consequences, like in healthcare.

In conclusion, AI is not here to replace anyone. Its impact on work is highly uneven, depending on its capability and the number of barriers to AI adoption in your specific task or job. What's clear is that it's looking like the people who will thrive are those who own workflows, use judgment to understand where and where not to trust AI, and use it to do better and more meaningful work.

Ani Bajaj is a Product Manager at Microsoft and the author of Indispensable: Your Career Guide for the age of AI, which analyzes 100+ US occupations for how AI will change work, and how you can use AI in your job today.