
Superintelligence and Manager Careers: Skills to Build Now
Management, review and process design skills that tend to gain value as AI agents handle more work, and where a manager should start learning.
Want to apply this in your company?
Book a call with Juan →When people talk about superintelligence, the useful question for a manager is not when it arrives but which skills keep mattering when agents carry out more of the work. The short answer: mapping processes accurately, setting limits on actions and data, reviewing results with discipline, and deciding what stays with people and what can be automated. These skills do not depend on predicting the future, because you already need them to run the AI agents that exist today.
Key takeaways
- Mapping a process with its inputs, outputs, exceptions and decisions is worth more than knowing any specific tool.
- Limits on actions, data and permissions stop an agent from deciding beyond what the company accepts.
- Reviewing agent outputs requires quality criteria defined before the run, not after.
- Choosing what is automated, what is prepared by the agent and what is decided by a person is a management skill.
- A team transition works better with agreed metrics, a record of decisions and room to adjust.
What changes in the manager's role when agents do the tasks
An support or operations agent queries systems, follows steps and answers people. As it becomes more capable, the bottleneck stops being execution and becomes decision making: who sets the rule, who approves an exception and who answers when something goes wrong. Anthropic, in its guidance on building effective agents, recommends starting with the simplest solution possible and adding complexity only when it is justified. For a manager, the core competence is knowing where complexity pays off and where it only adds risk.
It helps to separate three things that often get mixed together: what a tool's official documentation says, what your team has tested in practice, and what is still a hypothesis. Anyone making a decision based on a hypothesis should treat it as a hypothesis, with a deadline to verify it.
Skill 1: map the process before automating it
Before any agent, the process needs to exist in a written, verifiable form. This sounds basic, but it is the step that most often exposes what the company does not know about its own work. A short list of questions usually helps:
- What are the inputs, the outputs, and who makes each decision at each step?
- Which exceptions show up every week, and which ones are rare but cause the most trouble?
- Where does work loop back, wait on someone or get redone?
- What is a fixed rule that can be written as a condition, and what depends on judgment?
The last question matters most. Flows with fixed rules can be built with deterministic automation, with no AI involved. The part that depends on judgment is where a model can help, and where human review has to be designed in. To see how to organize the information that feeds this mapping, read Data Ready for Superintelligence: What to Standardize First.
Skill 2: define limits on actions, data and permissions
An agent should only have access to what the task requires. This principle, known as least privilege, is old in information security and remains the best foundation for agents. OWASP, in its list of risks for applications built on language models, names excessive agency as a specific problem: the system gets more tools or more decision power than it needs, and a mistake or a malicious instruction turns into a real action.
In practice, a manager should answer four questions in writing before any pilot:
- Which actions can the agent take on its own?
- Which actions require approval from a person?
- Which data can it read, and which personal data should never enter its context?
- When must it stop and hand the case to a human?
In Brazil, personal data falls under the LGPD, the national data protection law. That covers purpose, minimization and, in some cases, international transfer. A manager does not need to become a lawyer, but needs to know which questions to put to the technical team and to legal counsel. For AI in meetings, for example, the requirements on notice and legal basis are covered in Recording Meetings With AI: Consent and Privacy Law (LGPD).
Skill 3: review results against criteria defined in advance
The most common review is also the weakest: someone reads a few answers, decides they look fine and approves. That does not scale, and it does not show whether the agent is improving or getting worse after changes. The alternative is to define evals, the term technical teams use for structured tests, with real cases, expected answers and acceptance criteria. Anthropic has published material on evaluating agents that describes this process step by step, and its central idea is that evaluation starts with cases that reflect real use, not with examples chosen to look good.
Illustrative scenario: a building materials store wants an agent to answer delivery time questions on WhatsApp. The manager defines a set of typical questions, the correct answer for each, the cases where the agent should say it does not know, and the rule for handing off to the counter staff. Before going live, the team runs this set in a test environment and checks every answer. When an answer fails, the failure becomes a new test case. The manager learns the same thing they would learn reviewing a new salesperson, but with a method that can be repeated.
Skill 4: decide what stays with people
Not every task should be automated, even when the technology allows it. Deciding what stays with the team is a management decision, and a simple test can guide it:
- If the decision is repeatable, has a clear rule and errors are cheap to fix, it can be automated with a log of every action.
- If the rule is clear but an error is costly or hard to reverse, the agent prepares the proposal and a person approves it.
- If the decision depends on relationships, negotiation or judgment about an exception, a person decides and the agent supports them with organized information.
- If it involves sensitive data, a legal obligation or a direct impact on a person, the rule must be written and approved before any automation.
This guide is not fixed. It should be revised after the pilot shows where the agent got it right and where it got it wrong.
Skill 5: lead the adoption and measure without surveillance
The human side of the transition is often the hardest part. Teams resist when they feel replaced or monitored. A manager who handles this well makes clear what the agent does, what it does not do, who is accountable for decisions and how each person's work changes. Measuring process results, such as response time, rework and customer satisfaction, is different from measuring each person by how many messages they sent. For that balance, read How to Measure AI Productivity Without Surveillance on Your Team and, on the leader's role, Superintelligence for Team Leaders: What Changes in Your Daily Routine.
Study plan for the coming months
If you want to prepare without waiting for the market to settle, a practical path can follow these steps:
- Pick a real process in your area, ideally one with known volume and exceptions, and draw the current flow on paper.
- Mark where decisions happen, where exceptions happen and where personal data appears.
- Define three quality criteria for the output of any agent and name who reviews it.
- Build a set of real cases and test a version in an isolated environment before exposing customers to it.
- Record what the system decided, what was corrected and why. That history is the team's learning material.
In our own work, this path follows the sequence map, limit, pilot, transfer. First you map the process and its baseline. Then you restrict actions and data. Next you run a real flow with logging and a metric agreed in advance. Finally the company decides whether to expand, adjust or stop. The sequence also works for someone studying alone, because it does not depend on any specific tool.
Some priorities tend to return little right now: memorizing model names, following predicted dates or buying products that promise to replace an entire process. To separate marketing claims from real capabilities, the piece on Is ChatGPT Superintelligence? What It Is and What It Is Not helps frame the conversation with the board.
Frequently asked questions
Do I need to learn programming to build these skills? Not necessarily. Mapping processes, setting limits, writing review criteria and leading a team are management competencies. Understanding the basics of how an agent receives data and executes actions helps you talk with the technical team, but it does not replace the decision about what should be done.
What should a manager review first in a customer support agent? Start with answers that involve money, deadlines, personal data or a commitment the company makes. Also check whether the agent knows when it does not know and hands the case to a person. These points carry the greatest operational risk and usually show up in the first tests.
How do I know if a process is ready to be automated? A process is ready when its inputs and outputs are written down, its exceptions are known and someone is responsible for each decision. If there is still debate about how the work should function, the priority is to organize the process, not to put an agent on top of it. A small pilot with logging and an agreed metric shows quickly whether the automation makes sense.
Will these skills keep their value regardless of how models evolve? The safer bet is yes, because mapping, limiting, reviewing and deciding are needs of any operation that has customers, data and legal responsibility. Tools change quickly, but the ability to say what should be done, by whom, within which limits, and how to know whether it worked remains a management responsibility.