Superintelligence for Founders: A Weekly Reading and Decision Routine

Superintelligence for Founders: A Weekly Reading and Decision Routine

A weekly routine for founders to read AI agent summaries, separate fact from hypothesis, and log decisions with written criteria and human review.

•

Want to apply this in your company?

Book a call with Juan →

In this article, the superintelligence lens does not mean tracking future systems. It means treating AI agents as sources of work at large scale. They summarize customer service, sales and operations faster than any team can read them, so they need their own routine for reading and deciding. The proposal is simple: once a week, the founder reads summaries that cite their sources, checks them against real data, and records each decision with its reasoning, the alternative that was rejected, and a review date.

Key takeaways

  • The routine is short, fixed on the calendar, and always uses the same summary and decision template.
  • An agent summary is input for a decision, not a verdict. Every relevant claim must point to a source you can verify.
  • Each decision becomes a record with evidence, accepted risk, an owner, and a criterion for reversing it.
  • The superintelligence lens helps you sort what you delegate, what you supervise, and what stays with people.
  • Without a success metric agreed before the pilot, the reading becomes a pleasant habit with no effect on operations.

What the superintelligence lens changes in reading

At the current stage, no system that a Brazilian company uses for customer service or operations is superintelligent. Agents work on bounded tasks, with defined tools and permissions. The risk for a founder is therefore not a machine that decides on its own. The risk is a fluent, well-organized summary that sounds certain and gets accepted without checking, because its form persuades before its content is verified.

The working hypothesis behind this routine is that the more operations depend on agents, the more the quality of the founder's reading functions as an operational control. Reading well means asking where each figure came from, what was left out, and what would happen if the conclusion were wrong.

To separate what is a chatbot, what is an agent, and what gets called superintelligence in public debate, read Is ChatGPT Superintelligence? What It Is and What It Is Not.

Which summaries belong on the weekly agenda

Not every summary deserves the founder's time. The agenda should be short and tied to decisions the company actually makes. Usually this includes summaries of customer service with handoffs to people, sales funnel and follow-up summaries, complaint and incident summaries, and a summary of what the AI tools cost to use. Sources that would not change any decision stay off the agenda until there is a reason to add them.

Before accepting any summary onto the agenda, check these points:

  • Which period and which sources were used to generate the summary.
  • What the summary left out, stated explicitly.
  • Which claims are backed by a conversation, a record, or a spreadsheet.
  • Which personal data appears, and whether it is truly needed for the decision.
  • Which action the summary suggests, and who would be responsible for carrying it out.

If the summary does not answer these questions, it goes back to the agent with an instruction to redo it. This return may feel bureaucratic, but it is what keeps the routine from turning into comfortable reading.

Step by step: the weekly routine

Always book the same time slot and use the same format. The session works best when the material arrives ready, and the reading starts with the exceptions, not the highlights.

  • Before the session, the agent generates the summaries from a fixed template, with sources, period, and limits noted.
  • The founder reads the exceptions section first, because what falls outside the norm is where errors and opportunities usually sit.
  • For each relevant claim, open the original conversation, record, or spreadsheet and confirm it before deciding.
  • Compare the result with the metric agreed for the pilot. If no metric exists yet, the first decision of the week is to define one.
  • Record each decision using the format in the next section, including the decision to do nothing for now.
  • Set the review date and the name of the person responsible for execution.
  • Give the agent back only what it is allowed to do: request more data, run another comparison, or prepare a draft. Direct action toward a customer requires human approval.

The last step is the one most often skipped. A record without a review date becomes a dead file, and the routine loses its reason to exist.

How to record a decision so you can check it later

A good record lets the founder, a few weeks later, find out whether the decision was right and why. Use the same template every time so records can be compared with each other:

  • Decision, in one clear sentence.
  • Evidence used, with source and period.
  • Alternative rejected and the reason it was dropped.
  • Accepted risk and who accepted it.
  • Owner of execution and review date.
  • The criterion that, if not confirmed, leads to reversing the decision.

This record also protects the company. When an agent makes a mistake, the question stops being who trusted it and becomes what the agreed criterion was and whether it was followed.

Illustrative scenario: a services company uses a customer service agent on WhatsApp. In one weekly summary, the report indicates that complaints about delivery times have dropped. When the founder opens the conversations cited, they see that the agent had started closing chats without transferring them to the team whenever a customer asked about delivery time, because the closing rule had been written too broadly. The recorded decision is to suspend automatic closing for delivery-time questions, keep the handoff to the team, and request a separate summary of those cases for the following week. The record notes that the rejected alternative was only rewording the summary, and that the review criterion is whether those transferred cases get a response the same day. The error was in the flow, not in the report, and it surfaced only because someone read the conversations.

When to delegate, supervise, or keep with people

Not every decision needs the same level of care. The bands below are a starting point, to be adjusted to your business and your risk tolerance:

  • Delegate to the system, with a record: reversible, low-impact tasks such as organizing a queue, suggesting a standard reply, or scheduling a reminder.
  • Supervise with human approval: messages outside the script, refunds, deadline commitments, and changes to customer records.
  • Keep with people: dismissals, performance evaluation of staff, decisions that affect someone's access to a service, and any automated decision that cannot be reviewed.

Decisions that affect people based on automated processing of personal data call for legal care. Article 20 of the LGPD provides the right to request a review of decisions made solely on the basis of automated processing of personal data that affect the data subject's interests. The topic is covered in more detail in LGPD Automated Decision-Making: The Right to Review (Art. 20).

How Juan Carlo's method fits this routine

In AI agent projects, we follow the order map, limit, pilot, and transfer. Applied to the founder's routine, mapping means listing the summary sources and writing down what each one can and cannot claim. Limiting means defining what the agent reads, what it writes, and which actions require approval. Piloting means choosing a real flow, with records and a success metric agreed in advance, and keeping the routine long enough to see whether it actually changes decisions. Transferring means handing the record template and criteria to the team, then deciding from the records whether the routine expands, adjusts, or stops.

Frequently asked questions

Do I need to understand superintelligence to use this routine? No. The routine is about reading, evidence, and recording decisions, practices that apply to any AI agent in use today. The term is useful as a reminder that volume and speed call for more control, not less. For the terms explained without jargon, the Superintelligence Glossary for Executives, Without the Jargon is a useful reference.

How much of the founder's time should the routine take? Ideally, the session fits a short window fixed on the calendar. If reading takes more time than is reasonable, the problem is usually the format of the summaries, not the number of decisions. Ask the agent to shorten the agenda and keep only the sources that change decisions.

What if the agent makes a mistake even with checking? Errors happen, and checking reduces them without eliminating them. The record lets you identify at which stage the failure occurred: the source, the criterion, the summary, or the execution. Without a record, the error becomes an opinion and the fix is guesswork. Adjust the template or the permission, and log the change as a new decision.

Which decisions should the founder always make personally? Decisions about people, material financial risks, sensitive data, and changes to the agent's scope should stay with the founder or with someone they designate. The agent can prepare the analysis and the options, but the choice must be human and recorded. For a broader view of the topic, see Superintelligence for Founders: What to Decide Yourself and What Not To.

Sources