Superintelligence Company: 10 Signs of a Prepared Operation

A superintelligence company, in the useful sense of the term, is not one that owns an AI superior to humans. It is one that has built an operation able to improve every time a better model arrives. That shows up in observable signs: its own memory, evaluations, minimal permissions and human approval at the points of risk. No company needs to claim superintelligence to have these signs, and this article does not claim it for anyone.

Key takeaways

• What makes an operation prepared is not the model it picked. It is what exists around the model: knowledge, tests, limits and records.
• Memory, evaluations, permissions and approval are the four axes that let you swap models without starting over.
• A manager with no technical background can check each sign with simple questions and concrete evidence.
• Prepared operations treat a model swap as a tested change, not as a bet.
• Starting with one real workflow and an agreed metric is safer than trying to prepare the whole company at once.

What does it mean to say a company is prepared for superintelligence?

The term has different uses. In research and philosophy, superintelligence means a system that would outperform humans at practically any cognitive task. As far as public documentation allows us to say, that is not something a company can buy or declare it has. To tell the concepts apart, read the difference between AGI, ASI and AI agents and what it implies for what your company actually runs.

In operational use, "superintelligence company" is a working metaphor: an organization whose processes absorb new AI capability with little friction. When a more capable model becomes available, it can test the model, measure whether it improved and put it into production under the same limits. That is operations engineering, not science fiction.

To be honest about where the claims come from, what follows combines three different things. The first is official documentation from vendors and from bodies such as NIST, cited in the sources. The second is implementation experience, meaning patterns that tend to work in agent projects. The third is hypothesis: the idea that an operation like this tends to benefit from future models. Nobody can promise that benefit. You can only lower the cost of verifying it.

Which memory and context signs show the operation is ready?

The first group of signs covers what the company knows and where that knowledge is kept.

Sign one: company knowledge lives outside the model. Policies, catalogs, support rules and exceptions sit in versioned sources with a named owner, and they are handed to the model at the moment of the task. If the knowledge exists only inside a giant prompt pasted into a tool, changing models means rewriting everything. The technical discussion of how to deliver the right context at the right time appears in Anthropic's engineering writing on effective context engineering for AI agents.

Sign two: memory has an owner, a time limit and a forgetting rule. If the agent stores customer preferences or conversation history, someone decides what may be kept, for how long and how it gets deleted. In Brazil this connects directly to the LGPD: personal data without a defined purpose is a risk, not an asset. Memory without a disposal rule is not accumulated intelligence. It is accumulated liability.

Verification questions for this group:

• If we switched models tomorrow, what would we have to rewrite?
• Who approves a change to a business rule the agent consults?
• Where is it recorded what the agent may remember about a person, and for how long?

Which evaluation signs separate testing from impressions?

The second group covers how the company knows something got better.

Sign three: there is a set of real cases with expected answers. It does not need to be large. It needs to be representative: common questions, ambiguous questions, attempts to leave the scope, and cases where the correct answer is to hand off to a person. Anyone who has read about evaluations of AI agents knows the idea: without evaluation, every change turns into an argument over opinions.

Sign four: a model swap runs against that set before going live. The result is compared with the current model, and the decision is recorded: adopt, adjust or keep. A new model can be better on average and worse on a case that matters to your business. Only evaluation shows that.

One practical detail: part of these evaluations can run against simulations, without touching real systems. The article on how to test AI agents with mocks and a sandbox explains how to do this without risk to customers or data.

Which permission and approval signs protect the operation?

The third group covers what the agent can do and who decides when the risk goes up. This is where most improvised operations fail, because a more capable model with broad permissions is simply a more capable mistake.

Sign five: minimal permissions per tool, with reading separated from writing. An agent that looks up orders should not, by default, be able to cancel orders. The OWASP authorization guidance on least privilege applies to agents just as it does to people and systems.

Sign six: risky actions require human approval with written criteria. A refund above a limit set by the company, a contract change, a mass send, a data deletion: each item has a clear rule on when the agent proposes and the person decides. Decisions that affect data subjects also call for attention to the right to review, the topic of the article on automated decisions under the LGPD.

Sign seven: handoff to a human is part of the design, not an emergency exit. The agent recognizes when it should not continue, passes along the full history, and the person takes over without asking the customer to repeat everything. On WhatsApp, this is usually the point that weighs most in the customer's perception.

Verification questions for this group:

• What is the worst action this agent can take today, on its own?
• Who last reviewed the permission list, and when?
• What happens when the agent is not sure?

Which observability and model-swap signs complete the list?

The last group covers seeing what happens and keeping the operation flexible.

Sign eight: every relevant step leaves a traceable record. Which question arrived, which sources were consulted, which tool was called, what was decided and by whom. The observability concepts from OpenTelemetry help frame this, even if the company starts with something much simpler, such as a well-structured decision log in a spreadsheet. The point is to be able to reconstruct a decision weeks later.

Sign nine: the model layer is replaceable. The workflow does not depend on a trick that only works with one specific vendor, and there is a plan for when the model becomes unavailable, changes behavior or stops meeting a privacy requirement. For companies with data restrictions, this includes considering alternatives, as described in running Ollama local models in a company, and understanding what the LGPD requires when data leaves the country.

Sign ten: there is an agreed success metric and an explicit power to stop. Someone in the company can pause the agent without asking permission from whoever built it. Review happens on a set date, with three possible outcomes: expand, adjust or end. An operation that only knows how to expand is not prepared. It is just enthusiastic.

Illustrative scenario: swapping the model in a support operation

The scenario below is illustrative. It does not describe a real customer or a measured result.

Imagine a building materials distributor that uses a WhatsApp agent to look up order status and answer delivery questions. The agent only reads order data, changes nothing, and hands any exchange or refund request to a person. The delivery rules live in a document with a named owner. Conversations are logged, with the source consulted for each answer.

A new model becomes available. The team does not switch it on right away. It follows these steps:

• Runs the set of real cases against the new model, in a test environment with simulated data.
• Compares the answers with those of the current model, looking first at the cases where the correct move is to hand off to a person.
• Checks that the new model respects the same tool and scope limits.
• Turns the model on for a small share of conversations, with human review of the first answers.
• Decides in a short meeting whether to adopt, adjust or keep the previous model, and records the reason.

None of this requires the model to be extraordinary. It requires that memory, evaluation, permissions and approval are already working. If the new model is better, the company captures the gain at low risk. If it is not, the company finds out before the customer does.

How do you apply this without trying to prepare the whole company?

The safest path is to pick one workflow and go through four verbs, which are also Juan Carlo's working method: map, limit, pilot and transfer.

Mapping means describing the current process and recording a baseline: who does the work, with which tools, where errors and rework occur. Without that, any improvement is an impression. Limiting means deciding which data the agent accesses and which actions it can perform, starting with the minimum. Piloting means running a real workflow, with records and a success metric agreed before starting. Transferring means handing over the operation with documentation, evaluations and internal owners, and deciding together whether the next step is to expand, adjust or stop.

To see the technical structure behind these signs, the article on the company superintelligence layer in five building blocks shows how memory, tools, evaluations and governance fit together. For the risk view at organization level, the NIST AI Risk Management Framework offers a useful vocabulary for talking with legal and security teams.

A quick self-test: if you can answer yes to most of the verification questions in this article for a single workflow, the operation already has a real foundation. If you cannot, the cheapest gap to close is usually evaluation, because it can start with a list of real cases and a spreadsheet.

Frequently asked questions

Does a superintelligence company already exist? There is no public evidence that allows us to say any company operates a superintelligence in the technical sense. What exists are companies with operations well designed to use AI in a controlled way. Be wary of anyone who promises otherwise without showing how to verify it.

Do I need a frontier model to have these signs? No. The signs describe the operation around the model, and they work with models of different sizes, including smaller or local ones. The advantage of having memory, evaluations, permissions and approval is precisely being able to change the model when it makes sense, instead of depending on a single one.

Where do I start if the company has none of the ten signs yet? Choose a workflow with reasonable volume and low risk, such as status lookups or triage of common questions, and first build the list of real cases and the minimal permissions. Then record decisions and agree on a success metric. Only after that think about widening the scope.

Do these signs replace legal analysis under the LGPD? No. They help organize the evidence, such as records, purpose and retention rules, but the assessment of legal basis, contracts and data transfer should involve whoever is accountable for it in the company. The ANPD, Brazil's data protection authority, is the official reference for guidance on the subject.

Sources

• Anthropic: Demystifying evals for AI agents
• Anthropic: Effective context engineering for AI agents
• OWASP: Authorization guidance
• NIST AI Risk Management Framework