Sebastian Mallaby stood in front of a roomful of CEIBS executives and said something that should have been obvious but somehow wasn’t: the biggest obstacle to AI agents taking over work isn’t whether they can do the job. They can. The obstacle is whether we’ll let them.
Consider DeepMind. They built AlphaGo, watched it dismantle Lee Sedol in 2016, then cracked protein folding with AlphaFold in 2020. Two of the decade’s most stunning AI achievements. And yet when the Transformer architecture emerged and large language models started reshaping everything, DeepMind was late to the party. Not because their engineers were incompetent, but because success had created what Mallaby calls cognitive self-enclosure. They’d won so decisively with reinforcement learning that the signal from a different approach got filtered out as noise.
There’s a pattern here that reaches beyond AI labs. Einstein spent the last decades of his life rejecting the probabilistic interpretation of quantum mechanics, not because he couldn’t understand the math but because it violated something deeper in how he thought the universe should work. “God does not play dice,” he famously said. The man who revolutionized physics couldn’t accept the next revolution because his own framework had become too successful, too comfortable, too all-encompassing.
Path dependence doesn’t just trap individuals. It traps entire fields. And right now, we’re watching it play out in reverse. The AI community has pivoted so hard toward agents and autonomous systems that we’ve stopped asking the uncomfortable question: what if the technology is ready but the world isn’t?
The Trust Problem Nobody’s Solving
You’d let a delivery driver pick up your package from your doorstep. You wouldn’t let them manage your retirement account. The difference isn’t capability. Plenty of delivery drivers could learn basic portfolio management. The difference is trust infrastructure.
When you hand your package to a stranger wearing a uniform, you’re not trusting them personally. You’re trusting a system: background checks, GPS tracking, insurance policies, customer reviews, legal liability, corporate reputation. Remove any piece of that scaffolding and the transaction falls apart.
Now imagine an AI agent that monitors your email, schedules your meetings, negotiates contract terms, and maybe reallocates some assets based on market conditions. The agent is technically capable of all this. The models exist. The APIs are there. But where’s the scaffolding?
If the agent misreads an email tone and torpedoes a client relationship, who’s liable? If it moves money based on a pattern that looked like market momentum but was actually a data glitch, who compensates the loss? If it signs you up for a subscription you didn’t want because it misunderstood context, how do you get your money back?
Traditional product liability law assumes predictable failure modes. Your toaster either works or it doesn’t. If it catches fire, there’s a clear chain of responsibility from manufacturer to retailer to you. But agent behavior is emergent. The same agent that perfectly handled 10,000 tasks might make a bizarre decision on task 10,001 because of some unexpected interaction between its training data, its current context, and your specific situation. You can’t recall it like a defective brake pad. You can’t patch it like a software bug without changing how it handles everything else.
This isn’t a technical problem waiting for a technical solution. This is a social infrastructure problem. We need the equivalent of business licenses, malpractice insurance, professional certifications, and contract law, but for entities that don’t fit into any existing category. They’re not employees. They’re not contractors. They’re not products. They’re something else, and we don’t have the legal or social framework to handle “something else” at scale.
Why Your Company Won’t Use Agents Next Year
In 1910, the first practical tractors started appearing on American farms. By 1920, the technology had proven itself. Tractors were faster, more powerful, and more efficient than horses. Any rational cost-benefit analysis said farmers should switch immediately.
They didn’t. Most American farms didn’t fully mechanize until the 1950s and 1960s. It took half a century for an obviously superior technology to achieve widespread adoption. The delay wasn’t about the technology. It was about friction.
Friction with existing infrastructure: barns built for horses, not machines. Friction with knowledge: fathers who knew horses inside and out but had never touched an engine. Friction with finance: banks that understood livestock as collateral but were skittish about lending against mechanical equipment. Friction with culture: farming communities where your reputation was partly tied to how well you managed your animals.
The tractor effect is playing out again with AI agents, but the friction patterns are more complex because they’re less visible. Some domains have almost no friction. Software development, content creation, data analysis, these are already being transformed because they exist entirely in digital space. An AI that writes code or analyzes spreadsheets doesn’t need to interface with physical reality. It doesn’t need regulatory approval. It doesn’t need to convince anyone it won’t cause harm. It just needs to work, and increasingly, it does.
Medium friction domains are where things get interesting. Law, medicine, education, these fields have the technical capability for agent integration right now. An AI could review contracts, suggest diagnoses, or tutor students with increasing sophistication. But there’s regulatory friction (who certifies an AI doctor?), liability friction (who gets sued when the diagnosis is wrong?), and cultural friction (how many parents are comfortable with their child’s primary teacher being a machine?).
Then there’s high friction territory: manufacturing, surgery, infrastructure, anywhere the digital has to reliably interface with the physical world under high-stakes conditions. An agent that optimizes factory schedules is one thing. An agent that controls the robots building your car is another. The gap isn’t technical capability. It’s the compound friction of safety certification, physical reliability, legal liability, insurance models, and the simple fact that when things go wrong in the physical world, people can die.
This is why predictions about agent adoption are mostly wrong. They extrapolate from low-friction domains and assume the curve continues. But friction doesn’t scale linearly. A technology that spreads like wildfire through software companies might take decades to penetrate manufacturing, not because the technology gets better slowly, but because the social systems surrounding manufacturing change slowly.
The One-Person Company Delusion
There’s a seductive narrative making the rounds: AI agents will enable one-person companies to do the work of hundred-person teams. The solo founder who builds, markets, sells, and scales a global business with nothing but their laptop and a subscription to the right APIs.
It’s possible. There are already examples. But they’re not the future for most businesses, for the same reason that most successful companies aren’t solo endeavors even when technically they could be.
AI agents amplify execution. They don’t amplify judgment. They can draft the marketing copy after you define the positioning. They can’t tell you what positioning will resonate with a market you don’t understand yet. They can analyze customer data and spot patterns. They can’t tell you which patterns matter and which are noise. They can prototype features faster than any human team. They can’t tell you which problems are worth solving.
The scarce resource in business has never been the ability to do tasks. It’s been the ability to figure out which tasks matter. To make decisions under uncertainty. To read situations that don’t fit any pattern in the training data. To know when the data is lying to you.
One-person companies work in domains where judgment scales, where one person’s understanding of the problem space is enough to make all the critical decisions. They fail in domains where judgment requires diverse perspectives, where understanding emerges from collision between different mental models, where the feedback loops are too long or too subtle for one person to track everything that matters.
AI doesn’t change this. It just makes the execution part cheaper. And in most businesses, execution was never the bottleneck. The bottleneck was knowing what to execute.
Running Fast in the Fog
We’re somewhere in the middle of what historians will probably call AI’s wild growth period. Not the beginning, the technology is already too mature for that. Not the end, too much is still chaotic and undefined. The middle, where the ground is still shifting but some patterns are starting to solidify.
The pattern that matters most: the gap between what AI can do technically and what society will let it do is growing, not shrinking. Every capability breakthrough widens the gap because it outpaces our ability to build trust infrastructure, update regulations, retrain workforces, and adjust cultural expectations.
This creates a strange dynamic. The people and companies winning right now aren’t necessarily the ones with the best AI. They’re the ones who can navigate the gap. Who can deploy capable agents in low-friction domains while everyone else is still arguing about whether it’s safe. Who can build trust scaffolding faster than competitors. Who can read which social frictions are temporary (people getting used to the idea) versus structural (real regulatory or safety concerns that won’t fade).
The skill isn’t pure velocity. It’s not about running faster than everyone else toward the obvious future. It’s about maintaining risk awareness while moving fast enough that you don’t get left behind. Knowing which barriers are real and which are phantom. Which friction will melt away in six months and which will still be there in six years.
Because here’s what Mallaby understood that most people miss: the agent era is already here for some things and might never arrive for others. The question isn’t when agents take over. The question is where, and for how long the gap persists between technical possibility and social readiness.
The companies that figure this out won’t be the ones with the most powerful AI. They’ll be the ones who understand which doors are actually open.
Stay updated with our latest AI insights
