Why Full Factory Automation Is Always Five Years Away

Prediction markets are not particularly popular in the manufacturing industry, but if they were, anyone who had consistently bet against “full automation within five years” would have retired wealthy many times over. The prediction has been made with confidence in every decade since the 1960s. It has been supported by genuinely impressive technology demonstrations each time. It has not been borne out.

This is not an argument that automation does not happen or does not matter. The manufacturing sector is dramatically less labor-intensive per unit of output than it was fifty years ago. That transformation is real and consequential. The argument is about a specific claim — that human labor will be removed entirely from the factory floor within a near-term horizon — and why that claim has proven so persistently wrong, and what that persistence tells us about the nature of manufacturing work.

The 1960s vintage

The first serious wave of “automation will eliminate factory workers” predictions coincided with the introduction of numerically controlled machine tools. NC machining, which allows a lathe or mill to be programmed to cut complex shapes automatically, was genuinely revolutionary. It eliminated the most demanding hand-work in precision machining: the expert craftsman who could manually cut a complex contour was replaced by a programmer who could specify that contour in code and a machine that could execute it consistently.

The prediction that followed was that this technology would generalize to the entire factory. If complex machining could be automated, why not assembly? Why not material handling? Why not inspection? The prediction was not unreasonable based on the evidence available. What it missed was the difference between automating a well-defined, isolated task and automating the totality of factory work, which consists of thousands of micro-tasks with rich interdependencies.

The compounding difficulty of integration

Here is a clean way to think about why “full automation” is so much harder than “partial automation.” If automating 50% of factory tasks requires 100 units of engineering effort and capital, automating 80% does not require 160 units. It requires something closer to 600. The last 20% costs something like 3,000 units to achieve, if it is achievable at all.

This is because the hardest tasks are hard precisely because they involve the exceptions, edge cases, and contextual judgments that the automation of the easy tasks creates and concentrates. Automate the repetitive assembly steps and you have a production line that reliably produces parts — but the exception handling that humans previously distributed across many tasks is now concentrated in the few remaining human roles. Remove those humans and you need systems that can handle the full tail of exceptional situations that the automated portions cannot manage.

The combinatorial explosion in edge cases is not a failure of imagination by automation engineers; it reflects the genuine complexity of physical manufacturing processes. A human assembler, encountering a situation they have not seen before, applies general intelligence, physical intuition, and contextual knowledge about what “good” looks like to resolve it. Replicating that generalized capability in a machine is orders of magnitude harder than replicating any specific task that the assembler performs routinely.

The moving economic target

There is also an economic argument that is less discussed but important. Automation is not free, and its cost-effectiveness depends on the labor costs it displaces. As automation technology advances, it displaces the lowest-hanging-fruit tasks first — the ones where the technology works well and the economics are compelling. This raises the bar for the remaining tasks in two ways: the remaining tasks are technically harder (otherwise they would already be automated), and the economic calculation is more demanding because the remaining human workers’ wages are set in a market where their lower-skill competitors have already been displaced.

In practice, this means that the economic threshold for automating the next tranche of tasks keeps rising even as the technology advances. The manufacturing worker who survives the first round of automation is generally more skilled, more adaptable, and more productive per labor hour than the one displaced. The automation system needed to replace them is correspondingly more capable and more expensive. The payback calculation becomes harder, not easier.

This dynamic is visible in the data from 2025-2027. Labor costs in the manufacturing sectors that have seen the most automation pressure have increased faster than economy-wide wage growth. The workers who remain on factory floors in high-automation environments command significant premiums over workers in manual environments. The economics of automating those premium workers is much less attractive than the economics of automating a minimum-wage task that was available fifteen years ago.

The AI wave: genuinely different, structurally similar

The current AI-driven automation wave has real differences from its predecessors. Machine learning-based vision systems can handle part variability that defeated earlier approaches. Large language models have enabled new kinds of human-machine interaction that reduce programming burden. Reinforcement learning in simulation has allowed robots to develop manipulation strategies that were not explicitly programmed. These are genuine advances, not incremental improvements.

But the structural constraints that have defeated previous waves remain in place. The combinatorial edge-case problem is still there. The economic targeting dynamic is still there. The tacit knowledge problem is still there. What has changed is where along the capability curve the current wave sits — it is further along than any previous wave, and the tasks it is reaching are more complex than what previous waves reached. Whether it is “further along enough” to finally close the gap is genuinely uncertain.

The specific claim from AI automation optimists in 2025 was that large language models, combined with improved robotics and vision, would enable a new class of general-purpose manipulation systems capable of handling arbitrary manufacturing tasks with minimal programming. The claim was not entirely wrong — systems like this have been demonstrated in research settings and have made real progress — but the gap between “research demonstration” and “production-reliable at automotive volumes” has proven wider than the 2025 predictions suggested.

The five-year mechanism

Why specifically five years? There is something almost sociological about the timeframe. It is short enough to be exciting — five years is within a planning horizon — and long enough to defer accountability. Technology demonstrations that are genuinely impressive but not yet production-ready create a specific kind of narrative gravity: the next five years will see the remaining problems solved.

The mechanism that keeps the prediction moving forward is not dishonesty (usually). It is the genuine difficulty of predicting how long it takes to bridge from “works in the lab” to “works in production at scale.” Researchers and engineers who have solved the hard technical problem — and the hard technical problem is real — correctly observe that the remaining barriers are more about engineering robustness, integration, and cost reduction than fundamental capability. Those barriers seem, from the inside, more tractable than the capability barrier that was just overcome. They consistently prove harder than expected.

There is also a selection effect in prediction-making. The predictions that get published and cited are the optimistic ones, from the people who have just achieved a breakthrough and are most excited about its implications. The skeptical predictions — made by people who have spent years watching previous breakthroughs fail to generalize to production environments — receive less attention and are remembered less.

What this wave will and will not achieve by 2032

The manufacturing automation landscape in 2027 allows a reasonably confident set of predictions about the next five years (acknowledging the irony of making a five-year prediction after arguing that five-year predictions in this domain are systematically overoptimistic).

The tasks where significant automation progress is likely by 2032: light assembly of electronic components (further generalization of current systems), quality inspection across a broader range of defect types and product categories, intra-factory logistics in structured environments, and some dexterous assembly tasks in well-characterized manufacturing environments where the economics justify development investment.

The tasks where progress is likely to be slower than current predictions suggest: wiring harness assembly, maintenance and repair in production environments, final assembly of complex products with flexible or fragile components, and any task that requires judgment about novel situations that were not present in training data.

The honest expectation for 2032 is not lights-out manufacturing at scale. It is continued progress on the same trajectory as 2022 to 2027: meaningful automation of the next tranche of well-defined tasks, real but limited progress at the dexterous frontier, and a factory floor that is more automated than today but still substantially populated with humans performing the work that automation cannot yet reliably do.

This is not a failure. It is the normal pace of industrial technology diffusion, which has always been slower and more path-dependent than the predictions made at the technology’s inception. The prediction has always been five years away. The transformation has always been slower. The result, over fifty-year timescales, has always been more dramatic than anyone predicted in real time. These are not contradictions; they are what progress actually looks like from inside it.