Three layers define today's AI stack — foundations, applications, and adoption. The third has been the most neglected, even though it determines whether AI replaces workers or amplifies them.
AI is here. But who benefits?
The answer is supposed to be “everyone.”
The evidence of the past 18 months tells a different story — one of accelerating divergence between the workers and organisations that can productively use AI, and those that cannot. At large, AI can replace the workers who use it, or it can amplify them. The same software, with the same access, can lead to vastly different outcomes.
The gap has a name. It is the adoption layer.
The three layers
There are three layers to the AI stack today, each with a different make up, and different winners.
The first is the foundation layer of frontier labs that capture the headlines, the geopolitics, and the unprecedented revenue growth.
The second is the application layer: products built on top of models to solve specific problems like customer service agents, legal research tools, design assistants, and AI tutors. The winners here are those who can identify and scale the most high-value use cases for AI to augment or replace the existing ways of doing things.
The third is the adoption layer: the human, organisational, and institutional capability to actually use any of this in a way that changes real world outcomes.
For too long the adoption layer has been treated as an afterthought, but it’s by far the most exciting and urgent challenge facing technology today.
The productivity gap between the most and least AI-fluent workers using the same tools is already 6x by some measures, and growing. For corporates, it is the difference between AI as a cost centre and AI as a productivity lever.
At a societal level, particularly in the polarised, low-growth democracies of the West, successful adoption is the difference between technology that lifts workers, elevating their roles, and technology that simply happens to them. Without broad distribution, the benefits accumulate to those already best placed.
This is the part of the AI stack that has been most neglected. It is also the layer where Britain and Europe have particular institutional advantages: highly educated workforces, strong absorptive capacity, mature systems of employer-led training, and public services that can act as natural laboratories for applied AI. The bones of an adoption infrastructure exist. The question is whether we grasp the opportunity in front of us.
The deployment problem
If you doubt that adoption is the binding constraint, the clearest evidence is what the AI labs themselves are now doing.
In the past few weeks, two leading frontier labs have raised multi-billion dollar professional services ventures, backed by some of Wall Street’s largest investors. Job ads for Forward Deployed Engineers — essentially a hybrid between consultants and software engineers — are through the roof.
It is a striking concession. For years, the pitch was that AI would automate away the cost of professional services. The labs are now standing up the largest professional services build of the decade not to replace consultants, but to do the work themselves.
The reason isn’t mysterious. Most organisations are not the clean-edged enterprise of the demo video. When it turns out most of your customer interactions fall into the category of edge cases, getting AI to deliver inside that reality requires more human labour, not less.
These services arms might reach the very largest enterprises, the Fortune 500, and the portfolio companies of the biggest private equity funds. But it is not a workforce-scale answer across a continent as diverse and complex as Europe, where 5.5 million UK businesses and roughly 25 million EU SMEs sit far below the labs’ line of sight. And even at the largest enterprises in the world: deployment does not equal adoption.
So what will actually work?
Some people have claimed that the key to AI adoption is simply trial and error, something that anyone can do on their own. This myth of self-teaching is easy to debunk. The issue is not so much using the tools as it is finding the right thing to apply them to; that requires support. And the idea that generic AI courses or YouTube videos explaining the technology will provide any meaningful gains to the adoption of the technology for workers doesn’t stand up to scrutiny. After all, trying to deliver outcomes with AI in the workplace often exposes the profound gap between what the technology can do in theory, and what is deemed permissible on the job by legal, IT, and HR departments in practice.
Multiverse has three principles for teaching; well-supported by the underlying learning science and our experience in delivering on the job learning to tens of thousands of workers, across nearly every sector and industry.
The first is that AI fluency is not primarily a technical skill. Looking at the difference between AI ‘power users’ and the rest, our learning scientists found that the determining factors are largely human: analytical reasoning to break a problem down, creativity to push beyond the obvious prompt, systems thinking to anticipate how the model will respond, scepticism and detail-orientation to catch its mistakes. Prompt-engineering is one of the smallest things that separates the most effective users from the least.
The second principle is that adults learn complex skills through application. Real work, in real contexts, with feedback from experts who have done the work themselves. They learn by trying, failing on small things, having someone more experienced explain why, and trying again. There is no shortcut around this loop. There is no ‘AI in 30 days’ course, no matter how slickly produced, that can compress it.
The third principle is that context really matters; and context is probably the defining word of the AI-era thus far. What a financial analyst at a global bank, a nurse working for an NHS Trust, or an operations manager at a fast growing startup can do with AI in their roles differs wildly. So you can’t sheep dip your way to progress here. You need a detailed understanding of an organisation’s goals and constraints, a clear understanding of function, team, and job specific priorities, and a thorough grasp of individual worker capability and experience. The context gap is one of the key reasons why even as expertise becomes commoditised by AI, humans still have a vital role to play in delivering work.
These three principles point to the same operational answer: learning embedded inside the job, guided by experts, with the learning coming from the weeks a worker spends tackling a real problem in their own organisation, developing the necessary skills in the process, and building organisational capability that otherwise wouldn’t exist.
This is, in essence, the apprenticeship model — older than AI, older than computing, and one of the things Britain, Germany, and the European continent, our core markets, have more institutional memory of than almost any other part of the world.
The output isn’t a certificate. It’s a measurable change in how a team operates.
Where we go from here
Most of this is obvious. Nearly every business leader is frustrated that their organisation isn’t adopting AI faster, and they’re well aware that the solution will never come from top down mandates.
It comes from the analyst who notices that three hours of every Tuesday is spent reconciling the same two reports. From the case officer who realises the bottleneck in their service is one repetitive judgement call.
There are many examples of how adoption is working in practice already. A service and selling coach at a national retailer built a ‘Distillation Bot’ that compresses hours of supplier training video into bite-sized micro-learning, cutting prep time by 87.5% and lifting partner quiz scores by 25%. An NHS consultant in paediatric intensive care designed an automated shift-safety alerting system that flags senior staffing gaps fourteen days in advance, swapping costly last-minute agency cover for planned cover. An aerospace ops officer built a stock-tracking dashboard for Aircraft-on-Ground events that, beyond eight hours saved a month, surfaces compliance risks worth up to $1m a year in penalties. A placements officer at a university built a structured job-title report for students choosing between placements, cutting analysis time from twenty minutes to three.
This is a quiet revolution happening right now. All of these individuals built things that matter to their organisations because someone trusted them to learn, and gave them the tools to apply that learning on the job. All of those individuals are frontline workers.
Multiply that by every council, every hospital, every factory, every back office, every team. That is the value of the adoption layer. It is also the surest way to tip the scales in favour of human agency, and away from the real risk of many workers being automated out of existence.
The civilisational bet
Because beneath all this is the biggest question: whether AI ultimately replaces the workers who use it, or whether it amplifies them.
Too often the default running through most of the AI conversation, particularly from big tech, is the former. Multiverse is one of the very few companies in this market whose explicit mission is the latter.
So far, the story of AI has been one of humans training machines to make machines smarter. We are working towards the inverse: using a blend of machines and our coaches to train humans to make humanity smarter, and to keep human judgement, creativity, and agency in the loop, even as the technology itself becomes more powerful.
Which outcome we get depends almost entirely on what we choose to invest in, and what we choose to prioritise.
This is, ultimately, why we built Multiverse the way we did. The next decade of AI will not just be won by those who built the underlying technology, but by those who built the workforce capability, and the institutions, to benefit from it.
