Tesseract SA

Superintelligence: what the word means for a small business in 2026

No superintelligence system exists in 2026. What the word covers, who is behind it, and the five decisions a small business can make now about its tasks, its data and human approval.

Matte 3D illustration on a light grey background: a staircase of five rounded white steps with one red sphere resting on the second step.

What is superintelligence?

Superintelligence is a hypothetical AI system that would outperform the best humans in practically every cognitive field: science, strategy, creative work, human relations. It goes further than artificial general intelligence, which would only match a capable person. In 2026 no such system exists, and its exact definition is still debated.

Philosopher Nick Bostrom popularised the word in a 2014 book to describe an intellect that greatly exceeds human performance in almost every field that matters. Since then it has become a fundraising argument, a lab slogan and a cover story. For the owner of a small business, the useful question is not "when will it arrive?" but "what does this word force me to decide today?". This article answers the second one.

Superintelligence, general AI, generative AI: three different things

The confusion comes from three very different ideas sharing one vocabulary. The table below separates them and shows what matters for a small business.

NotionWhat it isWhere things stand in 2026Effect on a small business
Generative AIModels that produce text, images or code from a promptIn production everywhere, including in your own toolsDirect: this is what you can plug into your repetitive tasks
General AI (AGI)A system as versatile as a capable person across varied tasksNo consensus: some call it close, others call it poorly definedNone until it is available and verifiable
Superintelligence (ASI)A system that far exceeds humans in almost every fieldHypothetical: no complete system existsNo direct effect. An indirect one on public debate, regulation and supplier pricing

Take this reading away: everything you can buy, install and measure today belongs to the first row. The other two rows are research hypotheses, not products.

Who talks about superintelligence, and why the word is everywhere

Several players have made it their banner. Safe Superintelligence, founded in June 2024 by Ilya Sutskever, Daniel Gross and Daniel Levy, states a single goal: building a safe superintelligence. Meta announced a group of labs carrying the name. OpenAI published a reflection on governing superintelligent AI as early as 2023, suggesting an international authority comparable to the International Atomic Energy Agency.

A sceptical camp reads part of this language as communication. Belgian broadcaster RTBF and the Blog du Modérateur relay experts who see it as a way for AI companies to frame the debate in their favour. In a podcast published in April 2026, KPMG France presents superintelligence as a narrative as much as a scientific fact, and points out that the risks already present lie elsewhere: overconfidence in answers, personal data, loss of skills.

Our view, as a studio that puts AI inside small businesses: both camps are partly right, and it does not change your decision. Models improve fast. Nobody knows where it stops. A serious plan has to hold whatever the answer is.

When will it arrive? Nobody knows

Dates on offer range from "within ten years", according to statements by OpenAI leaders reported by the Swiss press, to much more distant horizons, or to a notion judged too vaguely defined to date, among more cautious voices. That spread is itself the useful information: a forecast that varies by a factor of ten is no basis for a budget.

The practical rule for a small business is simple. Do not postpone any automation decision while waiting for "the next generation". Every quarter of waiting is a quarter of repetitive work paid at full price, while current models already handle reading, sorting, preparing and first drafts well.

What it changes in practice for a small business in 2026

Four effects, all indirect, all actionable.

1. Supplier dependency becomes a management risk

If capabilities move in jumps, the model you use today will be replaced, renamed or billed differently. A process built around one specific model ages badly. A process described by its inputs, rules and outputs moves from one model to another. Describe the task first, the tool second.

2. The data question becomes central

The more powerful a model, the greater the temptation to hand it complete files. Yet everything you send to an online service leaves your perimeter. Before sending anything, read what to check before sending data to a language model, and separate the data that can leave from the data that must stay with you.

3. Human approval stays the fixed point

Whether the model is average or excellent, a mistake that reaches a client or commits money remains your mistake. The principle "AI prepares, a person approves" does not depend on model power. It depends on the cost of an error.

4. The media noise will grow

Every lab announcement will be followed by predictions about your trade. Decide in advance what triggers a review of your choices: a change measured on your own tasks, not a headline.

How to read a superintelligence announcement in four questions

Every week brings an announcement, a ranking or a prediction. To avoid reacting to each one, put four questions to it, in this order.

  1. Who is speaking, and what are they selling? A lab raising money does not have the same interest as an independent researcher or an audit firm.
  2. Is a system available and testable? A demonstration, a press release or a roadmap goal is not a product you can try on your own files.
  3. What is being measured? A score on a public test set says nothing about your invoices, quotes or contracts.
  4. What does it change on one of your three tasks? If the answer is "nothing measurable", the announcement asks nothing of you.

This grid is deliberately plain. It also avoids the opposite trap of automatic scepticism: if a new model does better on your own tasks, you will see it in your own numbers, and that is where to decide. Keeping a short log of these tests, with the date, the task, the model and the error rate, is enough. After two or three review cycles you hold something no press release can give you: a track record on your own work, which is the only benchmark that predicts how a model will behave in your business.

What a prudent small business does differently from a large one

A large group can fund a research team, negotiate a contract with a lab and absorb a bad bet. A small business cannot, and should not try to imitate it. Its advantage is speed and proximity: the owner knows which tasks hurt, which clients are sensitive and which mistakes are expensive. That knowledge, not a forecast, is what makes an automation choice sound. Use it, keep your choices reversible, and let the models improve underneath a process you control.

Five decisions to make now, in this order

  1. List three repetitive tasks whose rules fit on one page. These are your automation candidates whatever happens next.
  2. Sort your data into three levels: public, internal, confidential. Only the first can go to an online service without prior thought.
  3. Choose where the model runs for each level: an online service for public data, a Swiss host or local AI for confidential data.
  4. Put a human approval step on every output that commits money or touches the client relationship.
  5. Set a review test: every six months, replay your tasks on a newer model and compare the error rate with the old one.

None of these five decisions requires a forecast about superintelligence. They hold if progress stops tomorrow, and they benefit from progress if it continues.

Local AI or online model: the trade-off the debate hides

The superintelligence debate suggests that only the largest models matter. For a small business the real trade-off lies elsewhere. An online model generally gives the best raw quality and is quick to set up. A local model, installed on your premises or with a Swiss host, keeps data with you and depends on no subscription, at the price of sometimes lower quality on hard tasks and hardware that must be sized.

At Tesseract Studio, the base we install ships with local AI by default: the method, one base then modules describes how data stays with you and how each action is logged. It answers a confidentiality question. It is not a position on superintelligence.

When this approach is not the right choice

Three cases where you should do otherwise. First: your data is not sensitive and your volumes are low. A well-chosen online service is enough, and installing local AI would cost more time than it saves. Second: the task demands the best reasoning quality available, for example a complex legal or medical analysis. Neither a small local model nor an automation replaces a specialist there. Third: the process is not described. No model, even a hypothetically superior one, repairs a task nobody has defined. To sort these cases, also read what a small business should not automate.

What we take away

Superintelligence is a serious subject for research, public policy and safety. It is not a purchasing criterion for a small business in 2026. Your decision rests on three measurable things: which tasks to automate, which data to protect, who approves. To put a figure on the first, the express diagnostic in five questions gives an order of magnitude in a few minutes.

Frequently asked questions

What is artificial superintelligence?

A hypothetical AI system that would outperform the best humans in practically every cognitive field. Nick Bostrom popularised the term in 2014. In 2026 no such system exists.

What is the difference between AGI and superintelligence?

General AI would match a capable person across varied tasks. Superintelligence would far exceed humans in almost every field. Both remain hypotheses with no agreed definition.

Does superintelligence exist yet?

No. Current models are generative AI, very strong on specific tasks, but no company offers a complete superintelligence system. Several labs have set it as a goal, though.

Should a small business wait for superintelligence before automating?

No. No date is reliable, and current models already read, sort and prepare documents well. Automate tasks whose rules fit on one page, with human approval on outputs that commit money.

Do I need local AI to be ready?

Only if your data is confidential. In that case a model installed with you or with a Swiss host avoids sending files to an online provider. If your data is public and volumes are low, an online service is enough.

Sources
Read next