Superintelligence for Founders: What to Decide Yourself and What Not To

An AI system can analyze pricing, staffing and risk exposure with more depth than a founder can manage alone. But decisions that commit money, people and reputation should stay with the founder, who answers for the result and knows the context the system cannot see. Superintelligence, in the sense of systems that outperform people on most cognitive tasks, remains a speculative topic rather than a tool you can buy today. In practice, the useful question is which part of the decision AI can prepare and which part requires a named person with authority.

Key takeaways

• AI can produce the analysis, but the decisions on pricing, hiring and risk still belong to the founder.
• Decisions that affect people, money above a set limit or personal data require an identifiable human owner.
• A well-written recommendation is not proof that its assumptions are correct, so those assumptions must be checked.
• Agents should receive narrow permissions, written rules, a decision log and a clear escalation path.
• Before expanding use, run a pilot with an agreed metric and a clear rule for adjusting or stopping.

Why more sophisticated analysis does not transfer responsibility

AI models organize text, compare scenarios and find patterns in spreadsheets at a speed no team can match. That fluency creates a trap: the recommendation looks complete, and then it starts to feel like a decision. An analysis, however, is only as good as the premises it receives, and much of a business's knowledge sits outside any database. Examples include an informal arrangement with a supplier, a manager's read on the team, or the sensitivity of a long-standing customer.

The NIST AI Risk Management Framework treats human oversight, clear accountability and monitoring as part of managing the risk of AI systems, not as an implementation detail. The OECD AI Principles also emphasize accountability for those who develop and operate these systems. In practical terms, responsibility stays with whoever leads the organization, and no tool can take over that role.

What stays with the founder: pricing, people and risk

An AI system can simulate demand sensitivity, compare margins by product and point out where competitors charge more. Positioning, the promise you make and what the brand is willing to lose, however, are choices. For that reason, no agent should change prices on its own, not even in a WhatsApp conversation, where a discount given by mistake becomes a commitment. An agent can apply the current price list and route any condition above a written ceiling for approval.

Hiring, promotion, pay and dismissal involve employees' personal data and directly affect someone's life. Article 20 of Brazil's LGPD gives data subjects the right to request a review of decisions made solely on the basis of automated processing that affect their interests. Even when AI helps organize resumes or evaluations, the final decision must be human, documented and explainable. This topic has its own implications, covered in automated decisions under LGPD and the right to review.

Risk includes legal exposure, data leaks, reputation crises, dependence on a single supplier and entry into a regulated market. These decisions should not be delegated to an agent, because mistakes are expensive and not always reversible. AI can map signals, organize facts and prepare scenarios. The choice belongs to the founder.

When AI can act without approval

Not every decision needs a human at the door. Some routine actions can be executed safely by an agent when the rules are written down:

• Answering questions with information the company has already published, such as opening hours, address and delivery times.
• Rescheduling an appointment within open calendar slots and advance-notice rules defined in advance.
• Drafting proposals, reports and replies that a person reviews before sending.
• Classifying incoming messages and routing them to the correct queue.
• Recording objective facts from a conversation in the system, without interpreting sensitive intentions.

The common criterion is that the action is reversible, low impact, governed by a written rule and logged. If any of these is missing, the AI prepares and the person decides.

A four-question check before delegating

Before giving an agent any decision-making power, answer these questions together:

• Can the decision be undone within a useful timeframe, without harm to a customer or employee?
• Does it involve people, money above a defined limit or personal data?
• Is there a written rule the system can follow without having to interpret it?
• Is there a named owner who can see the record that led to the recommendation?

If the answer to the second question is yes, or if the answer to the last one is no, the decision stays with the founder or a designated manager. This rule applies to small teams too: without a named owner, the agent should have no power to act.

Illustrative scenario: a residential maintenance company uses an agent on WhatsApp to quote services. The agent applies the current price list and can offer a discount within a ceiling the founder has set. When a customer asks for terms beyond that ceiling, the agent does not negotiate. It creates an approval request with the conversation history, the service, the justification and the expected response time, and tells the customer that an answer is coming. The founder approves or declines from a phone, and the record stays available for periodic review. AI handles the volume and the record-keeping, while the commercial decision remains with the person accountable for the cash flow.

How to run a pilot without handing over the wheel

A well-designed pilot shows what AI does well without transferring the decision to it. Follow this sequence:

• Map the current process and record who decides each step today and on what criteria.
• Sort decisions into three groups: the AI executes, the AI recommends and a human decides, and the human decides alone.
• Limit data and actions: the agent accesses only what the task requires and executes only what is on the approved list.
• Agree on the success metric before starting, along with the criteria for adjusting or stopping the pilot.
• Log every recommendation, approval, refusal and exception, so the review is based on facts.
• Test in a controlled environment with known cases before connecting the agent to real customers.
• Review with the owner and decide whether to expand, adjust or stop.

This is the approach we follow in AI projects for companies: map the process and the baseline, limit data and actions, pilot a real workflow with logs and an agreed metric, then hand over the operation and decide whether to expand, adjust or stop. The method does not change with the sophistication of the model. It exists precisely so that the final decision stays visible to the person who must make it. To test an agent before connecting it to customers, see how to test AI agents with mocks and a sandbox.

Common mistakes when delegating analysis to AI

The first mistake is accepting the analysis without checking its premises, such as the margin used in a calculation or the reference date of the data. The second is using the AI's recommendation as a shield for an unpopular decision, which weakens leadership and accountability. The third is giving the agent broad access because it is more convenient during implementation. The fourth is sending customer and employee data to services without checking where it is processed and under what conditions. That last point has its own rules under the LGPD, detailed in international data transfer when using AI providers. Brazil's National Data Protection Authority (ANPD) publishes official guidance on the subject, and it is worth following.

Frequently asked questions

Can an AI replace the founder in strategic decisions? No. Responsibility for the outcome, the context and the relationships with people and customers cannot be transferred to a system. AI can prepare scenarios, compare alternatives and flag risks, but the final choice and its justification must be human and recorded.

Can an AI agent set prices? An agent can apply a current price list, calculate terms within a ceiling and route exceptions for approval. Changing the pricing policy, defining positioning or accepting a discount outside the rules should remain with the founder, because those choices involve margin, brand and commitments to customers.

How do I record that the decision was the founder's and not the AI's? Keep the recommendation received, the premises used, who approved it, the date and the human reasoning for accepting or rejecting it. This record protects the company in a review and lets you learn from mistaken decisions without blaming the tool.

What changes when models become more capable? The need for limits, records and a named owner will likely remain, because the central question is not only technical capability but who answers for the consequences. Discussions about superintelligence help frame long-term governance, but day-to-day company operations depend on written rules and supervision starting now.

Sources

• NIST AI Risk Management Framework
• OECD AI Principles
• ANPD: Autoridade Nacional de Proteção de Dados