AI, AGI and ASI – What the Terms Mean and What Actually Matters for SMEs

Few fields suffer as much from sloppy language as artificial intelligence. In the same presentation you will hear about "AI", about "AGI" and sometimes straight away about "superintelligence" – and nobody in the room is quite sure whether all three mean the same thing. For vendors, that vagueness is convenient: a chatbot that drafts quotes sounds far more impressive if you can place it somewhere near a machine that will supposedly soon surpass every human thought process.

For decision-makers in an SME, that same vagueness is expensive. It leads either to inflated expectations – "the system actually thinks along with us" – or to blanket rejection, because the debate about superintelligence and loss of control obscures what genuinely works today. In this article I sort the three terms out properly, explain why AGI in particular has never acquired a workable definition, and derive from that what a Swiss company can practically do with all of it.

The most useful way to sort the terms is along a single question: how broad is the range of tasks a system can handle? Not how impressive it is at one discipline – pocket calculators have been superhuman at arithmetic for decades, and nobody calls that intelligence.

The decisive part of this table is the "Status" column. Everything a company can buy, deploy and productively use today falls, without exception, into the first row. The other two rows belong to a research and societal debate – not to a procurement decision.

You would expect an industry investing hundreds of billions to at least agree on what its stated goal actually is. The opposite is true. OpenAI's founding document defines AGI as highly autonomous systems that outperform humans at most economically valuable work – an economic definition. Others use cognitive criteria: transferring knowledge, learning from few examples, coping with genuinely novel problems. Others again consider the term fundamentally unusable, because "general intelligence" is not a cleanly measurable construct even in humans.

The most useful attempt to defuse the argument so far came from a Google DeepMind research team in 2023 with the paper Levels of AGI. The idea: AGI is not a switch you flip, but a grid with two axes – performance (from "Emerging" through "Competent" and "Expert" to "Virtuoso" and "Superhuman") and breadth (narrow versus general). Within that grid you can say precisely where we stand. For narrow tasks, superhuman systems have existed for years – chess engines, protein folding, medical image analysis in tightly defined questions. For general tasks, today's language models sit on the lowest rung, "Emerging AGI": they cover a surprising amount of ground, but at fluctuating quality and without reliability.

The definitional gap has an unpleasant side effect: because nobody has authoritatively fixed what AGI is, any vendor can claim to be approaching it. "We are on the path to AGI" is not a verifiable statement, it is a marketing device. For assessing a concrete product it is worthless.

One rung above sits superintelligence: a system that surpasses humans not in individual disciplines but comprehensively. The underlying concern is not that such a machine would turn "evil", but that it would pursue goals that diverge from ours while being more effective at reaching them than we are. Over the past few years that debate has moved out of its niche and into politics.

One indication: in October 2025 the Future of Life Institute published a statement calling for a prohibition on the development of superintelligence – until there is broad scientific consensus that it can be done safely and controllably, and strong public buy-in. By early 2026 more than 133,000 people had signed, among them AI figures such as Geoffrey Hinton, Yoshua Bengio and Stuart Russell. On the other side stands, for example, OpenAI's Sam Altman, who wrote back in 2024 that superintelligence was conceivable "in a few thousand days".

And what does the broader field say? A large survey of AI researchers put the probability of superhuman AI at roughly 10 per cent by 2027 – and 50 per cent by 2047. Read correctly, that is remarkably honest: the profession itself does not know. A two-decade spread between scenarios is not a forecast, it is an admission of deep uncertainty.

Here comes the part that surprises many people: for an SME's operational IT decisions, the AGI and ASI debate is practically irrelevant. Not because it is unimportant – socially and legally it matters a great deal – but because not a single action for the next twelve months follows from it.

All of the measurable value companies get from AI today comes from the first row of the table: drafting quotes faster, making contracts and minutes searchable, triaging support requests, extracting data from receipts, making internal knowledge findable. These are unglamorous but calculable effects. They do not depend on whether AGI arrives in 2030 or 2060 – they depend on whether your data is findable, whether your processes are defined, and whether your people know how to use the tool.

What genuinely changes for an SME is the speed of the tools, not the nature of the work. An assistant that summarises mediocrely today will summarise well in two years. Whoever gets their processes, data and responsibilities in order now benefits automatically from every improvement. Whoever waits for the great leap will still have no clean foundation when it arrives.

Because the terms are used so loosely, a short reality check in vendor conversations helps. You will hear the following sentences often – here is what usually lies behind them and the question that creates clarity.

One point that tends to get lost in the superintelligence debate but has immediate practical consequences: what gets regulated is not hypothetical AGI, but today's concrete applications.

In the EU the AI Act applies in stages; obligations for providers of general-purpose AI models have been in force since August 2025, while the requirements for high-risk systems were pushed back under the so-called Digital Omnibus. Switzerland is taking a different route: in February 2025 the Federal Council decided to adopt the Council of Europe's AI Convention into Swiss law and otherwise regulate sector by sector rather than create a comprehensive AI act. A consultation draft covering transparency, data protection, non-discrimination and oversight is announced for the end of 2026.

For an SME this means that most obligations already follow from the revised Swiss Data Protection Act and from sector-specific rules – not from some future AI statute. If you document cleanly which data flows into which tool, you are prepared for both.

So that this article does not stop at terminology, here are the measures worth taking whichever way the AGI debate turns out:

• Put your data house in order. AI tools are only as good as the information they can reach. Scattered, outdated or unstructured documents are the most common reason for disappointing results.
• Pick two or three use cases with measurable value. Not "introduce AI", but "cut quote preparation by X hours". Without a metric you cannot judge later whether it paid off.
• Write a short usage policy. Which tools are approved, what data may go in, who reviews the output? Two pages are enough – but they have to exist, otherwise shadow IT fills the gap.
• Build competence instead of handing out licences. The difference between disappointed and enthusiastic users almost never lies in the product, but in whether someone showed them what it is good for and what it is not.
• Choose your dependency deliberately. Which data sits with which provider, and how would you get out if you had to? Better asked before deployment than after.

AI, AGI and ASI describe three completely different things: a mature technology, a loosely defined research goal, and a hypothetical scenario. Public debate blends them together – partly out of enthusiasm, partly out of sales interest, partly out of genuine concern. Keeping the three apart means you lose neither the societal dimension nor your footing.

For a Swiss SME the practical consequence is this: follow the debate about the future with interest, but keep it out of the investment case. Decisions are made about concrete tools, concrete processes and concrete data. Get those right, and you are better prepared for every stage of the technology than anyone waiting for the great leap.