After the Hype: The Real AI Questions Facing Enterprises

Few subjects generate more heat and less light than artificial intelligence. Across a series of recent, wide-ranging discussions among investors, founders and operators, the same arguments surface again and again: AI will eliminate work, or it will create it; models are about to become superhuman, or they are quietly commoditising; the technology should be tightly controlled, or set free. For anyone responsible for building systems that have to work in production, the noise is the problem. Beneath it sit a handful of questions that genuinely determine outcomes. This is our attempt to separate the signal from the spectacle.

The Jobs Debate Is the Wrong Debate

The argument over AI and employment has hardened into two camps. On one side sit the executives attributing layoffs to automation and warning of mass displacement. On the other sit those who frame it as ordinary creative destruction, pointing out that automating a task is not the same as eliminating a job, and citing the familiar precedents: bank tellers increased after the ATM, live entertainment grew after television. Both positions contain truth, and arguing them in the abstract settles nothing.

The more useful signal sits beneath the argument. The dividing line is not the technology; it is agency. People and organisations that treat these tools as a way to learn faster and do more are pulling ahead. Those who use them to avoid learning, or who refuse them outright, are exposed. As one veteran investor put it, the most durable form of job security is to become the most capable version of yourself the tools allow.

The best way to protect yourself from AI is to be the most AI-enabled version of yourself you can be.

What this looks like in practice is instructive. The marketable skill is not access to a model; it is the ability to extract value from one. In one example, a small team produced a genuinely useful daily briefing system not by writing code, but by iterating on prompts and a detailed instructions document until the output was sharp and contextual. The lesson is the opposite of the popular fantasy that you can drop AI into an organisation and watch value appear. Someone still has to direct it, supervise it, validate it and improve it every day.

That is why the operators who compound advantage tend to share a recognisable profile:

• Systems thinking — they understand the process they are automating well enough to judge the output.
• High agency — they reach for the tool first and learn by doing, rather than waiting for permission.
• Iteration as habit — they treat prompts, context and workflows as living artefacts that improve daily.
• Supervision and validation — they assume the model is probabilistic and build the checks that make it trustworthy.

A word on AI washing

Not every layoff blamed on AI is caused by it. A good deal of recent restructuring is, on closer inspection, a correction of years of over-hiring rather than the result of measurable automation gains. Honest operators draw that distinction carefully, because the discipline that matters is the same one that separates real adoption from theatre: measuring actual return rather than asserting it. It is also worth asking a question that rarely gets asked in these debates — whether the people doing the jobs under discussion actually want them. High turnover in many of the roles most cited as at-risk suggests the honest answer is more complicated than either camp admits. Displacement is real, it is uneven, and the people affected are not abstractions. The constructive response is to enable rather than to alarm.

The Trust Problem Is the Real Problem

If there is a single underappreciated risk in this moment, it is not capability — it is trust. Commencement speeches by prominent technologists have been booed. Surveys show real public anxiety. The resistance is not irrational. People intuit that a powerful technology controlled by a small number of actors can create outsized advantages for those actors and leave everyone else exposed, and the benefits that would answer their fears have not yet reached them. Some of the unease is older and deeper: a sense that AI is somehow anti-human, displacing people from the centre of the story. And there are credible arguments that some of the loudest opposition is actively encouraged by interests that would prefer to see progress slow.

The antidote is not louder optimism. It is to change who we listen to. The breathless commentary of model-makers tells you little about whether the technology works; the people using it tell you everything.

Stop asking the inventors of AI what they think. Ask the nurse, the factory worker, the scientist — and tell those stories.

Those stories exist, and they are concrete. In one account, a parent unwilling to accept a bleak prognosis used these tools to research a rare genetic condition and identified an existing, approved drug that materially improved their child's quality of life — with work now underway to tailor it further. Researchers have used the technology to revisit shelved drug candidates and to crack problems that had resisted progress for years. This is the part of the story that gets lost in the argument over headcount: when AI is grounded in a real problem and a capable human, it does things that were previously out of reach. For builders, the lesson is practical. Trust is earned at the level of outcomes, not announcements.

There Is No Single Best Model Anymore

A quieter but more consequential shift is the convergence of the frontier. Independent evaluations increasingly find that the leading models cluster within a fraction of a percentage point of one another on the same tasks. Read superficially, that looks like several systems reaching the same level of capability at once. Read strategically, it suggests the model layer is commoditising faster than the capital flowing into it.

The economics reinforce the point. As training stacks are rewritten closer to the hardware and domain-specific silicon arrives, the cost of producing a competitive model is falling by orders of magnitude. The era of the ten-billion-dollar training run is giving way to far cheaper alternatives, and the open-weight ecosystem keeps narrowing the gap with the closed labs.

Commoditised is not the same as finished, however. The next frontiers — recursive self-improvement, where a model contributes to its own training, and continual learning, where it learns from experience the way people do — could re-accelerate progress sharply, with some practitioners talking about order-of-magnitude gains year on year. Two further shifts matter for anyone deploying this technology. First, the harness around a model — the runtime that manages memory, state and integrations — is becoming as important as the model itself; capability and harness now have to be developed together. Second, the future is unlikely to be one enormous model but networks of smaller, often verticalised models that are cheaper to run and easier to govern.

For an enterprise, the implication is direct: do not anchor your business to a single frontier provider. The risks operators name most often are concrete.

• Leapfrogging — one provider pulls ahead suddenly, and you have bet on the wrong horse.
• Terms-of-service and political risk — a provider's policy can cut you off from a capability you depend on, sometimes for reasons that have nothing to do with you.
• Lock-in — pricing power and dependency accumulate quietly until they are expensive to unwind.

The Real Moat Is Context, Not the Model

The single most repeated insight from people actually shipping these systems is that the model is not the differentiator. AI is probabilistic; it has to be grounded in real data and a semantic layer — a single source of truth — or it produces confident nonsense. The advantage belongs to whoever can feed the model the right context, governed and current, not to whoever happened to call the cleverest model this quarter.

Architecturally, this is pushing serious organisations away from off-the-shelf vertical applications and towards horizontal platforms on which they build their own workflows. It favours headless products and a control plane that can hot-swap one model for another as the frontier shifts. And in regulated industries — finance, healthcare, anything touching personal data — it makes on-premise deployment, data residency, auditability and compliance non-negotiable rather than nice to have.

This is where most of the value, and most of the failure, lives. The widely shared story of a large enterprise spending heavily on tokens for minimal results is not an indictment of the technology. It is an indictment of deploying it without grounding, governance or measurement. The systems that work are the ones built on a foundation of trustworthy context.

Sovereignty Becomes a First-Class Requirement

Privacy used to be a question of data sovereignty — who can see your information. The emerging question is intelligence sovereignty: whose model interprets your world. It is one thing to keep your files private; it is another to decide who gets to analyse them and tell you what they mean. As that distinction sharpens, the ability to run capable models on your own hardware, and to govern them yourself, moves from a technical preference to a strategic safeguard.

This is why the open-source and open-weight ecosystem matters beyond cost and customisation: it is the backstop against a future in which participating in the modern economy means surrendering control to a single provider. The same concern animated a recent and widely discussed papal encyclical on AI, which framed the central question plainly.

Will AI be used to concentrate power in the hands of a few, or will it serve everyone?

Regulation cuts both ways here. The instinct to clamp down on a powerful technology is understandable, but power handed to a single authority is rarely handed back, and over-empowering one body to approve or constrain models risks producing the very concentration it sets out to prevent — the old problem of who guards the guardians. It is worth remembering that meaningful checks already exist: a competitive, decentralised market lets customers walk away from a provider that oversteps, and the courts already hold developers accountable when their systems cause harm. For European and Swiss organisations in particular, jurisdiction, data residency and the right to run and govern your own models are becoming board-level questions rather than infrastructure footnotes.

Cooperation Over Decoupling

The geopolitics of technology are moving in a similar direction. Recent high-level engagement between the United States and China has revived an old idea: that economic entanglement is the shortest path away from conflict. The logic is straightforward. In a world where productivity and output are expanding, there is less reason to fight over a fixed pie; cooperation lets both sides raise living standards without taking from the other. In a resource-static world, the incentives invert.

Applied to technology, the argument is that diffusion lowers the probability of conflict while restriction raises it — a parallel that several observers draw explicitly with the nuclear era, where an uneasy balance proved more stable than monopoly. It points towards narrow, sensible guardrails that rival powers could share, such as basic know-your-customer checks on the most capable models to keep them out of the wrong hands, rather than blanket attempts to bottle the technology up. The same reasoning is reshaping debates over chip exports and even the strategic weight of contested regions as domestic fabrication scales on all sides.

For builders, the takeaway is not a political position but a design assumption: capability will diffuse rather than stay bottled up, the world will remain multipolar and interconnected, and resilient systems are built with optionality rather than on the assumption of a single dominant supplier. The same conversations also turn to the fragility of the physical layer beneath all of this — power, data centres, supply chains, even the case for infrastructure beyond any single government's reach — a reminder that the systems we build are expected to hold up precisely when conditions are not benign. That is the bar that matters: working when it counts, not only when everything is calm.

The Economics Are Becoming Real

For all the talk of bubbles, the economics are starting to firm up. Leading model providers are reaching very large revenue run-rates at healthy margins on inference, and at least one has reportedly turned profitable — which changes the tenor of the return-on-investment debate. At the same time, consumption has outrun expectations. Generous flat-rate plans created the impression that tokens are free, and organisations responded by leaving the hose running, with usage compounding unnoticed and occasional reports of spectacular, unmanaged spend.

The result is a coming wave of efficiency. Expect intelligent routing between models, reserving expensive frontier calls for the work that genuinely needs them; cheaper inference as architectures are rebuilt closer to the hardware; and a more sober view of infrastructure, where the useful life and financing of compute matter as much as raw capability. The durable question the market is learning to ask is no longer how capable a system is, but what return its spend produces. Operators who can answer that with evidence will keep their budgets and their credibility. Those who cannot will watch both contract as the scrutiny arrives.

What This Means for Builders

Strip away the predictions and the posturing, and the debates resolve into a small set of durable principles. They are unglamorous, which is exactly why they are reliable.

• Build with agency. The advantage goes to teams that adopt, iterate and supervise — not to those who wait for certainty.
• Ground everything in context. The moat is trustworthy, governed data and a clear source of truth, not the model of the month.
• Earn trust through outcomes. Show the result to the people who use the system, not the promise to the people who fund it.
• Preserve sovereignty and optionality. Keep the ability to run, govern and switch models, and to keep data where it belongs.
• Measure outcomes, not activity. Instrument cost and value, and be able to prove the return rather than assert it.

This is the discipline behind agentic systems that work when it matters. It favours those who read the primary sources rather than the headlines, who assume the world will stay competitive and interconnected, and who build for volatility instead of pretending it away. The loudest forecasts will keep changing. The principles will not.