Superintelligence for Companies

Superintelligence for companies is not a product you can buy today. It is the name for a hypothetical AI that could outperform people at most intellectual tasks. What an owner can do now is prepare the business with mapped processes, controlled data and measurement, all of which already pay off with current systems. Five questions help separate that useful preparation from hype before you spend anything.

Key takeaways

• Superintelligence is a hypothesis, not a catalog item. Anyone selling it as a product is selling a narrative.
• Useful preparation is the same in any scenario: a clear process, data with an owner, limited actions and a metric agreed before you start.
• Every expense should be justified by a real workflow in your company, not by a forecast about dates or markets.
• Every automation needs a human owner, a record of its decisions and a simple way to stop it.
• If an investment only makes sense when a future technology arrives, it is not ready to be made.

What the word superintelligence promises and what it does not deliver today

The term became popular because of the debate about systems that would broadly exceed human ability. It is a legitimate debate, but it is about research, risk and policy. It is not about which tool your support team should adopt soon.

The systems available to companies are language models, agents and automations that do well-defined tasks: drafting, classifying, consulting documents, filling in fields, suggesting next steps. They make mistakes, invent information with apparent confidence and depend on well-built context. To understand how the concepts differ, read about the difference between AGI, ASI and AI agents and the superintelligence myths that get in the way of decisions.

The practical point: you do not need to decide whether superintelligence will arrive in order to decide whether AI is worth the investment. You need to decide whether one of your processes improves today with what already exists. The five questions below are built for that.

Question 1: Which concrete process do you want to improve?

If the answer starts with "be ready for the future," there is no project yet. There is an intention. A concrete process has a start, an end, an owner and a known volume: answering quote requests, checking documents for a registration, triaging support tickets, updating the CRM after a meeting.

To test whether the process is well defined, answer in writing:

• Who does it today and how long does it take, by your team's honest estimate?
• Where does the work stall, go wrong or depend on a single person?
• What would an acceptable result look like, and what would count as a serious error?
• What happens if the AI does not know the answer?

If nobody can answer, the first investment is not technology. It is mapping the current process and recording the baseline, meaning how the operation works before AI. Without that, any improvement is just opinion. This step usually costs little and almost always reveals that part of the problem can be solved with a simple rule, no AI needed.

Be wary of proposals that skip this question and offer a generic platform "for the whole company." A good start is usually small and specific.

Question 2: Which data and which actions can the AI touch?

This is the question that most separates serious preparation from improvisation. A useful AI needs access to information and, often, to tools: reading a spreadsheet, querying a system, sending a message. The wider the access, the greater the potential damage when something goes wrong.

Make two short inventories. The first is data: what information the workflow needs, which of it is personal or sensitive, where it is stored and where it goes when the model processes it. If data is sent to providers outside Brazil, it is worth understanding what the LGPD requires for international data transfer when using AI. The ANPD is the official reference for data protection guidance in the country.

The second inventory is actions: what the AI can read, what it can suggest and what it can do on its own. A practical rule is to start with reading and drafting, and only allow execution for reversible, low-impact actions. Sending a quote without review, changing a customer record or promising a delivery date are actions that need human approval at first.

There is also the technical risk of manipulation. Text received from customers or external documents can try to push the agent into ignoring its instructions. The OWASP Top 10 for LLM applications describes this and other risks, and it is a good checklist for anyone buying or building. To keep up with the regulatory context, see what to watch in Brazil's AI regulation.

Question 3: How will you know it worked?

An AI investment without a success criterion agreed in advance ends in a debate about perception. Someone thinks it got better, someone thinks it got worse, and nobody has records to argue from.

Define a metric tied to the process, not to the technology. "How many messages the AI answered" is activity. "The answers were correct and the customer did not have to repeat the request" is a result. Pick one or two measures, such as time to first useful response, the rate of edits made by the team, or the number of cases that needed correction. Compare them with the baseline from question 1.

Then keep records. Every agent decision should leave a trace: what came in, what went out, which source was used and whether a human stepped in. This lets you audit errors and improve the workflow. For more formal evaluation, Anthropic publishes material on how to structure evaluations of AI agents, and OpenAI has an evaluations guide. Both help turn "it seems good" into a repeatable test.

Finally, agree in advance on what happens at the end of the pilot: expand, adjust or stop. Stopping is also a valid result, and it saves more than pushing on.

Question 4: Who answers when the AI gets it wrong?

The AI will get something wrong at some point. The management question is who notices, who fixes it and who answers to the customer. If the answer is "the vendor," the design is incomplete.

Assign an internal owner to each workflow, with the authority to pause the agent. Also define the handoff point to a human: when the conversation involves a complaint, a billing issue, sensitive data, a request outside scope, or any situation where the agent shows uncertainty. On channels like WhatsApp, this handoff must be clear to both the customer and the team, with history available to whoever takes over.

If the workflow makes decisions that affect people, such as credit, hiring or access to a service, the LGPD provides the right to request review of automated decisions. The article on automated decision-making under the LGPD explains what that means in practice.

A simple test: imagine the agent sent a wrong answer yesterday. Can you find out which one it was, how many customers received it, and correct it today? If not, you are missing records and you are missing an owner.

Question 5: What still holds if the technology changes?

Models change, pricing changes, vendors change. A good investment is one where most of the value remains when you swap the model for a better or cheaper one.

What usually stays: the mapped process, the organized knowledge base, the written business rules, the evaluation tests, the decision records and the well-defined permissions. What usually gets lost: configurations tied to a single product, data trapped in a proprietary format, workflows nobody in the company understands.

Ask the vendor, or your own team: if we switch models later, what do we have to redo? If the answer is "everything," the dependency risk is high. The block-based architecture described in the company superintelligence layer helps separate what is durable from what is replaceable.

For organizing risk more broadly, NIST publishes an AI risk management framework, useful as a shared vocabulary between leadership, legal and technology. Anthropic also describes, in its piece on building effective agents, the principle of starting with the simplest solution and adding complexity only when needed. It is a good buying criterion.

How to apply the five questions: a step plan and an illustrative scenario

The method I use in consulting work follows four moves: map, limit, pilot and hand over. Mapping is describing the process and the baseline. Limiting is restricting data and actions. Piloting is running a real workflow with records and an agreed metric. Handing over is giving the operation to the team, with documentation, and deciding whether to expand, adjust or end it.

A step plan for the owner to use before signing any contract:

• Choose a single process and write down on one page how it works today.
• List the data it uses and mark what is personal or sensitive.
• List the possible actions and classify them as read, suggest or execute.
• Define the success metric and what will be recorded for each decision.
• Name an internal owner and the handoff point to a human.
• Agree on the pilot length and the criteria to expand, adjust or stop.
• Write down what would need to be redone if the model were replaced.

Illustrative scenario: an electrical supplies distributor receives quote requests on WhatsApp and loses time on repeated questions about lead times and availability. Instead of buying a broad platform, the owner chooses only the initial triage. The agent just reads the request, checks the approved product list and drafts a reply that a salesperson approves. No price goes out without review, and requests with a technical question go straight to a person. At the end of the pilot, the team compares response time and the corrections made against the previous situation and decides the next step. Nothing in it depends on a future technology.

If your company wants help with this design, the page on how Juan Carlo works with companies describes the process without promising results.

Frequently asked questions

Does superintelligence already exist? There is no consensus that it does, and current commercial systems do not match what the term describes. They are useful for well-defined tasks, but they make mistakes and need supervision. Treat any offer that uses the word as a product with caution and ask to see the concrete workflow.

Do I need to wait for superintelligence before investing in AI? No. The preparation that matters, such as a mapped process, organized data, metrics and access control, is useful with today's technology and stays useful if it changes. Waiting only delays your team's learning.

How much should I invest in a first project? There is no universal figure, and any number given without knowing your process would be a guess. A safer approach is to set a cap proportional to the problem the pilot solves and release more resources only after recorded evidence. The idea of unit cost per task, discussed by the FinOps Foundation, helps with comparison.

Does the LGPD change when the AI is more advanced? The principles are the same: purpose, necessity, security and transparency. What changes is the attention to what the system accesses and decides. Consult the ANPD guidance and involve legal counsel before putting personal data into any new workflow.

Sources

• NIST: AI Risk Management Framework
• OWASP: Top 10 for LLM Applications
• Anthropic: Building effective agents
• Anthropic: Demystifying evals for AI agents