The Threshold Effect: Why Physical AI Will Disrupt Jobs Differently Than You Think

Jobs don't disappear when AI automates tasks. They disappear when the last valuable task goes

A few weeks ago, David Oks published an essay that’s been making the rounds in economics circles. Its argument is strong: the famous story about ATMs not killing bank tellers is true — I am a big fan of the example in fact I have referenced it many times in presentations; but it’s only half the story. The other half is that another technology did kill them, and it wasn’t the one anyone expected. Not a robot, not a smarter ATM. It was the iPhone.

Oks’s point is that the ATM substituted tasks within an existing paradigm, so the institution adapted and the teller survived and evolved. The iPhone, he argues, did something different — it created an entirely new paradigm that made the branch itself irrelevant. The idea is that task substitution leaves humans in the loop. Paradigm replacement makes them obsolete. It’s a clean story, and it’s been widely shared for good reason.

But I think it’s an incomplete story. Or more precisely: it’s right about the outcome and incomplete about the mechanism. And the mechanism is what matters if you want to use the story to think clearly about what physical AI will do to jobs over the next decade.

It wasn’t a paradigm shift. It was a threshold.

Here’s another way to think about it. The iPhone didn’t create a new banking paradigm. People still have bank accounts. Still have institutional relationships with lenders. Still think of money as something managed by a financial institution rather than a protocol or a community. The paradigm — banking as a relationship mediated by a physical presence — didn’t change. What changed is that the physical presence ran out of relevant tasks to justify its fixed costs.

Think about what a branch visit represented: a bundle of potential tasks. Get cash. Deposit a check. Transfer funds. Ask a question. The ATM automated the first. Online banking — long before the iPhone — handled the third and fourth for most people. The iPhone, with its camera and its App Store and its payments layer, automated the second. And once check deposit moved to mobile, there was almost nothing left that required showing up. The branch didn’t become conceptually obsolete — it became economically indefensible. Fixed costs with no remaining task load to justify them.

This is a threshold effect, not a paradigm shift. Task automation accumulated slowly and mostly invisibly across four decades — ATM, then internet banking, then mobile — and at some point the bundle of remaining branch tasks became too thin to support the institution’s cost structure. When that threshold crossed, employment fell off a cliff. Fast, structural, apparently sudden. Which is why it looks like a paradigm shift. But the paradigm never changed. The economics did.

This distinction matters because threshold effects and paradigm shifts make different predictions. A paradigm shift is hard to see coming — it arrives from outside the existing frame. A threshold effect is traceable: you can watch task automation accumulate, you can monitor the fixed-cost structure of the institution, and you can ask which remaining tasks are still economically important. It’s not easy, but it’s a mechanism rather than a label applied after the fact.

In my experience, workflow changes in institutions happen very slowly, while technology moves fast. This is the gap that makes threshold effects dangerous. Technology automates tasks faster than institutions can reorganize around them; so often they don’t. For a long time, the institution absorbs the change, finds new roles, retrains workers, discovers that the automated tasks freed people for higher-value work. But that absorption has limits. When the last load-bearing tasks go, the institution can’t adapt fast enough, and the cliff appears.

Knowledge work is already mid-cliff

The threshold model didn’t wait for robots. It’s been running on knowledge work for over a decade, and in several cases the cliff has already appeared — we just haven’t named it correctly.

Medical transcription is perhaps the cleanest example, and it’s instructive precisely because of how narrow the job was. A medical transcriptionist did essentially one thing: listen to a physician’s recorded notes and convert them into structured text for the medical record. That’s not a simplification — the task bundle that comprised the job was thin. There was no secondary high value task to fall back on when voice recognition software matured. No adjacent function the institution needed humans to absorb. The threshold crossed fast because there was almost nothing between “the core task automates” and “the job disappears.” Between the late 1990s and the mid-2010s, a workforce of over 100,000 in the US shrank dramatically, and it has kept shrinking. The paradigm — physicians documenting patient encounters — didn’t change at all. The economics of paying humans to do the conversion did.

This points to something the threshold model predicts but often goes unstated: task bundle width is a vulnerability factor. Jobs comprised of many tasks have more runway. When one task automates, the institution can reorganize around the remaining ones, retrain workers, find new functions to justify headcount. The cliff, if it comes, comes later. Jobs with narrow bundles have no such buffer. The automation of a single core task can be immediately fatal to the economics, even if the surrounding institution is perfectly healthy.

Legal document review follows a similar pattern, playing out more slowly because the bundle is somewhat wider and the institution — the law firm — has more fixed-cost slack to absorb change. The load-bearing tasks of a junior associate used to include research, document review, contract drafting, and due diligence. Each has been eroding at a different pace: legal research databases in the 1990s, e-discovery software through the 2000s, LLMs now doing credible first drafts of contracts and due diligence summaries. The law firm as an institution is intact. The billable hour survived. But the task bundle that justified entry-level associate headcount has thinned. The paradigm is fine. The threshold is approaching.

What these cases share is the same pattern: automation accumulated across many tools and many years, largely invisibly within the institution, until the remaining human tasks were too thin to support the old cost structure. Nobody saw or announced a paradigm shift in medical documentation or legal practice. The threshold crossed.

This matters for how we think about the next wave, because knowledge work provides a cleaner view of the mechanism than physical work does. The bundle components — data access, reasoning, language — matured in a relatively legible sequence, and we can trace which tasks went when. Physical AI is harder to read because the bundle is more complex and several components are still immature. But the dynamic is the same. And in some physical domains, it’s closer than it looks.

Physical AI is a bundle, not just a device

When it comes to physical AI, it’s a collection of technology and it looks different from the digital world.

The question isn’t whether physical AI will create a new paradigm for, say, warehousing or healthcare or food preparation. It probably won’t anytime in the foreseeable future— those institutional paradigms are deeply embedded and will likely survive in some form. The question is whether physical AI will automate enough tasks, fast enough, to push specific institutions past their economic viability threshold before they can reorganize.

To answer that question, you need to open the bundle. What people call “physical AI” is not a single technology any more than “the iPhone” was. It’s a cluster of components at different stages of maturity that all need to come together and deliver. Computer vision and spatial perception. Dexterous manipulation. LLM-based task reasoning. Cheap, reliable actuators at human scale. Edge compute powerful enough to run inference in the field. A power and battery layer that can sustain a full working shift. Not to mention a workflow integration layer — kind of like the App Store — that lets domain-specific applications be built on top of a general physical substrate without bespoke engineering every time.

Each of these is a separate development curve. The bundle’s readiness for any given job depends on which components that job requires and how mature those components currently are. This is more useful than asking “is physical AI ready?” — because it gives you a dependency map rather than a binary answer.

Let’s review the components in terms of rough maturity.

Computer vision and perception are mature. Systems that can identify objects, navigate known environments, and interpret visual scenes reliably enough for industrial use have existed for several years. Good enough to build on.

LLM-based task reasoning is mature and improving fast. Interpreting ambiguous instructions, handling exceptions, knowing when to escalate — no longer a serious bottleneck. The connective tissue between understanding and action is being closed rapidly by agentic frameworks.

Dexterous manipulation is improving but not there. This is the most important bottleneck in the bundle right now. Handling objects in unstructured environments — the chaos of a kitchen counter or a patient’s room rather than the controlled geometry of a warehouse shelf — remains hard. Progress from some startups is real, but the “good enough” threshold for most jobs hasn’t been crossed. Think of this as where the smartphone camera was in 2005: improving, but not yet the unlock.

Cheap, reliable hardware and power are close but dragged down by an energy problem that doesn’t get enough attention. The cost curves on actuators, sensors, and compute are moving in the right direction. But a humanoid robot that sustains two or three hours of real work before needing a recharge isn’t a viable labor replacement — it’s an expensive proof of concept. Battery energy density and power management for sustained physical work remain genuine constraints. The hardware was working long before all-day battery life for smartphones. We just adapted. Physical AI is somewhere in that same gap now but the cost and inconvenience of adapting isn’t worth the pain.

The workflow integration layer is nascent and underappreciated. This is the missing App Store — the platform that lets a hospital administrator or restaurant operator configure and deploy physical AI for their specific context without rebuilding from scratch each time. Without it, technically capable systems still require significant integration work that makes deployment slow and expensive. Physical AI is still waiting for the infrastructure that turns a device into a platform.

Trust, liability, and regulatory frameworks are genuinely missing in most domains. A robot that injures a patient, causes a construction accident, or contaminates food creates liability questions that no legal framework currently resolves cleanly. In some sectors — contained industrial environments — this will be sorted relatively quickly. In healthcare, it could delay the bundle’s assembly by a decade regardless of hardware progress.

Some examples

The maturity map becomes most useful when you lay it against specific jobs — and ask not just which tasks can be automated, but which high value tasks are keeping the institution’s economics alive.

The warehouse picker. Vision mature, navigation in semi-structured environments mature, manipulation of standardized objects are close, workflow integration with inventory systems is close. Liability tractable. The bundle is nearly assembled — Amazon already deploys versions at scale. But more importantly: the tasks that remain human in a warehouse are precisely the manipulation and exception-handling tasks that are closing fastest. When those go, there is no remaining task load that justifies the fixed cost of a human workforce at scale. Here, the threshold is close.

Now take a restaurant cook on a line. Similar vision and reasoning requirements, but manipulation is much harder — food prerparation involves enormous variety in texture, weight, and geometry in a fast-moving, high-temperature environment. The workflow integration problem is also harder: a kitchen is a {kind of} choreographed human system, and a non-human actor requires re-engineering the surrounding process, not just replacing a node in it. The bundle is two or three hard to create components away. But the institution’s economics are also more fragile — thin margins, high labor cost as a share of revenue — which means the threshold may be closer than the technology alone suggests. Or there’s a motivation for the institution to adapt workflow faster than they have historically in order to get labor costs down.

How about a home care nurse? Manipulation in unstructured environments: hard. Workflow integration with healthcare systems and family care plans: extremely complex. Trust and liability: unresolved. Several components missing simultaneously. But look at the institution: home care economics are already under severe strain from labor shortages and cost pressures. The threshold question here isn’t purely technological — like food preparation it’s whether the institution will restructure around partial automation, using physical AI for the subset of tasks it can handle while humans retain the critical/higher risk ones. That restructuring could preserve employment even as significant task automation occurs, because the economics of the institution don’t collapse — they adapt.

I believe that jobs aren’t threatened when many of the tasks are automated. They’re threatened when the last economically difficult tasks get automated, and the institution can no longer justify its cost structure. That’s a different vulnerability profile than either “this job is easy to automate” or “this job requires human judgment.”

Then there’s unpredictability

One thing the threshold idea shares with Oks’s paradigm framing: neither fully protects you from surprise.

The bank teller story is only surprising in retrospect. In 2005, almost nobody would have named a consumer smartphone as the technology most likely to hollow out branch banking. The connection between a camera app and a teller’s job security wasn’t visible until it was. The components assuming were invisible in plain sight, but nobody was mapping them onto the institution’s cost structure and asking which tasks were still load-bearing.

Physical AI will almost certainly produce at least one equivalent surprise to what happened with the iPhone. There is probably a job right now sitting at the intersection of already-mature bundle components — vision, reasoning, some manipulation capability that already exists in industrial contexts — where nobody has yet asked the threshold question seriously. The disruption will look sudden and structural when it arrives, the way teller employment looked when it fell off a cliff. But the automation will have been accumulating for years.

Here’s a proposed way to think about it: environments that are more structured than they appear, institutions with high fixed costs and thin remaining task loads, and roles where the “highly human” tasks turn out to be less critcal than assumed. Radiologists — medical imaging AI plus remote diagnostics plus a contained liability environment — is one candidate already in progress. Legal document review is arguably past the threshold already, but even there open questions still exist on whether customers will accept legal interactions beyond “review”. There will be others less obvious.

Conclusion: watch economics, not technology

The standard question in AI-and-jobs conversations is “can AI do this task?” Oks improves on it: he says the better question is whether AI creates a new paradigm, not just automates existing tasks. But it’s still looking in the wrong place.

The mechanism isn’t paradigm replacement. It’s threshold economics. Task automation accumulates — slowly, across many technologies, often invisibly — until the institution built around those tasks can no longer justify its fixed costs. Then employment falls fast, which looks like a paradigm shift but isn’t. The paradigm often survives. The economics don’t.

This means the right questions are: which tasks in this institution are still economically high value? How fast is the critical automation bundle assembling? And how much fixed-cost slack does the institution have to absorb task loss before it crosses the threshold?

Physical AI is a real and assembling bundle. Some institutions are close to their threshold — warehousing, standardised food preparation, parts of logistics. Others have genuine buffers, either because key bundle components are missing (manipulation, power, the integration layer) or because their remaining load-bearing tasks are genuinely hard to automate. But the buffer is in the economics, not in some intrinsic human quality of the work.

The teller wasn’t protected by anything irreplaceable. The threshold crossed, and the cliff appeared. For physical AI, the question is which institutions are quietly approaching the same point — and whether anyone is watching the right numbers.

Crawford Del Prete is a Senior Advisor at PSG Equity and former President of IDC. He writes about enterprise software, AI, and technology market dynamics. All opinions are his.