Manager Skills in the Age of Superintelligence: What to Learn

Manager Skills in the Age of Superintelligence: What to Learn

Before any tool, the manager skills that matter in the age of superintelligence are four that do not lose value as AI improves: the judgment to decide what is worth doing, the ability to set goals that can be verified, the discipline to review results you did not produce yourself, and the willingness to take responsibility for what was decided. The more capable the system, the cheaper execution becomes and the more expensive it is to point it in the wrong direction. This article shows how to train each of these skills in your routine, without depending on any prediction about superintelligence.

Key takeaways

• Execution gets cheaper, so the bottleneck shifts to deciding what to do and why.
• A vague goal produces a vague result: people who can define success, limits and deadlines direct both staff and agents better.
• Reviewing well is a trainable skill, and it requires knowing what to look at when you would not do the task yourself.
• Accountability cannot be delegated: the name that signs the decision still belongs to a person.
• These skills are built in small pilots with records, not in generic courses about technology.

What changes for managers when AI becomes more capable?

First, an honest separation between three kinds of claims. There is what official documentation describes: AI systems already carry out text, analysis and tool-use tasks with growing autonomy. There is what implementation experience shows: these systems make errors that are not always obvious, and the cost of an error depends on who reviews the work. And there is the hypothesis that systems far beyond human performance in many domains will exist. On that last one, nobody knows the timing or the form, and building your management approach on it would be a bet. If you want to understand the difference between the concepts in circulation, read the article on the difference between AGI, ASI and AI agents.

What you can say without betting is the direction: tasks that once took a person hours are now done by an agent in minutes, with variable quality. That changes what a manager does in three ways.

• There is less work in producing and more work in choosing, prioritizing and checking.
• The volume of things done goes up, and so does the volume of things to review.
• Errors repeat at scale, because a poorly briefed agent makes the same mistake many times before anyone notices.

In other words, management stops being the craft of distributing effort and becomes the craft of directing and verifying.

Why does judgment matter more than execution?

Judgment is the ability to say, with reasons, that this is worth doing and that is not. When execution was expensive, scarce time forced prioritization. When execution gets cheap, it is easy to produce a lot of irrelevant output that looks like good work: reports, replies, proposals, automations nobody asked for.

Judgment has components you can train:

• Know the real process, not the designed one. Someone who has never followed a customer request or a service order from start to finish cannot tell where AI helps and where it gets in the way.
• Know what is irreversible. Sending a message to a customer, granting a discount or editing a record each have a different cost to undo.
• Tell judgment tasks from rule tasks. A rule can be a deterministic workflow. Judgment is where AI comes in, and where review carries the most weight.
• Recognize when an answer looks too good. Fluent text is not proof of correctness.

A good exercise is to log, for a few weeks, the decisions you made with one sentence on why. When AI suggests something, compare the suggestion with your log and you will see where your judgment was already clear and where it was only habit.

How do you set goals that a person or an agent can meet?

A bad goal is one you only know was met afterward, based on someone's opinion. With AI agents this becomes obvious: without a definition of done, the system delivers something plausible and the manager complains about something they never asked for. Anthropic's guidance on agents, in its article Building effective agents, recommends starting with the simplest solution that works and adding complexity only when measurements justify it, which assumes you know what you are measuring.

A checklist for writing a goal that survives contact with practice:

• The expected result in one sentence, without adjectives like good, fast or better.
• How it will be verified: who checks, with what sample, how often.
• Data limits: what information the system may see and what it may not.
• Action limits: what it can do alone, what needs approval and what is forbidden.
• A stop condition: when the pilot is ended, adjusted or expanded.
• A named owner: one person answers for the result, not a department.

None of this is specific to AI. It is the good management of always, except that now the absence of it shows up faster and in greater volume.

How do you review work you would not do yourself?

Reviewing is not the same as approving. Approving is saying yes. Reviewing is having a method to find what is wrong even when the text looks correct. This skill matters more as AI takes on tasks outside the manager's own specialty, because the manager ends up checking something they may not know how to produce.

A few practices help:

• Review by sample, with criteria defined beforehand. Choose easy, hard and unusual cases, not just the ones that arrived first.
• Keep reference examples. A small set of cases with a known correct answer works as a ruler. Anthropic describes this logic in Demystifying evals for AI agents, and the same reasoning applies to tasks that are not code. For machine-generated tests, there is a practical discussion in AI-generated tests and whether to trust them.
• Look at the process, not just the result. Decision logs show why the agent did what it did, and that is where systematic errors appear.
• Ask for evidence. When the AI states something factual, the review question is where the information came from.
• Separate review from production. The person who wrote the instruction tends to see what they intended, not what is actually there.

One caution: review should not turn into surveillance of people. Measuring the quality of a process is different from monitoring every move of the team, and the article on measuring AI productivity without surveillance covers that boundary.

Who answers when AI decides?

Accountability is the least technical skill and the hardest to outsource. When an agent replies to a customer, calculates a quote or suggests a refusal, the company has replied, and inside it someone must be able to explain the decision. The OECD AI Principles treat transparency, robustness and accountability as pillars, and the NIST AI Risk Management Framework organizes risk management around governing, mapping, measuring and managing.

In Brazil there is also a legal side. The LGPD (Brazil's data protection law) provides, in Article 20, the right to request review of decisions made solely on the basis of automated processing that affect the data subject's interests. The article on automated decisions and Article 20 explains what this means in practice. For a manager, the consequence is simple: before automating a decision, know who reviews it and how the customer asks for that review.

A decision guide in list form, for each task you are thinking of handing to an agent:

• If the error is easy to undo and low impact, the agent can act alone, with sampled review.
• If the error can be undone but is visible to the customer, the agent prepares and a person approves before sending.
• If the error is irreversible or affects a right, credit, health or a contract, the decision stays with a person, and the AI only supports with information.
• If nobody can explain the decision afterward, the task is not ready for automation.

How do you practice these skills in your routine?

Management skills are learned by doing, with feedback. Here is a plan that fits in a busy manager's calendar:

• Pick one small, real workflow, such as answering recurring questions or preparing a draft proposal.
• Write the goal with the checklist from the earlier section, before opening any tool.
• Define what the AI can see and do. If personal data is involved, confirm the legal basis and how it is processed. A company policy for ChatGPT use helps give the team a common rule.
• Run it for a short period, logging each decision and each human correction.
• Gather the team and review the errors: was the failure in the instruction, the data or the judgment?
• Decide in writing: expand, adjust or stop.

The central point is that learning comes from reviewing mistakes, not from successes. A manager who only sees what worked does not train judgment.

Illustrative scenario: the operations manager and the quotes

Imagine an operations manager at a distribution company who uses an agent to prepare quote drafts from customer messages on WhatsApp. This is an illustrative scenario, not a real case.

In the first week, she notices the drafts look good, but nobody can say whether they are right. Instead of approving everything, she picks a sample of easy and hard cases, checks each one against the official price table and notes where the agent went wrong. She finds that the errors cluster in orders with mixed units of measure.

She then changes three things: she rewrites the instruction to handle units, she decides that orders with more than one type of unit always go to a person, and she names someone each week to review a sample. The agent did not get smarter. The management got clearer, and the results improved because of it.

How does the map, limit, pilot and hand over method apply?

The same exercise can be done in four moves. Mapping is describing the current process and recording a starting point, so you can tell later whether anything improved. Limiting is defining the permitted data and actions, with human approval where an error is costly. Piloting is running a real workflow, with records and a success metric agreed beforehand. Handing over is passing the operation to whoever will maintain it, with documentation and clear criteria to expand, adjust or stop.

Notice how each step exercises one of the skills in this article: mapping trains judgment, limiting trains accountability, piloting trains goal setting and review, and handing over trains clarity about who answers. To see how this kind of work is usually carried out, read about how Juan Carlo works with companies.

Frequently asked questions

Does a manager need to learn to code to work with AI?

Not necessarily. You need to understand enough to describe a process precisely, define limits and read decision logs. Coding helps in some contexts, but judgment, goals and review carry more weight in most management roles.

Will superintelligence replace the manager?

Nobody can claim that on the basis of evidence, and treating it as certain in either direction is speculation. What we can observe is that decisions affecting people, contracts and reputation still need someone to answer for them. Preparing well means strengthening the skills that stay necessary in any scenario.

Where should a manager start preparing?

Pick one small, real workflow, write the goal with a way to verify it and clear limits, run a short pilot with records and review the errors with the team. Doing this once teaches more than reading a lot about the subject. Then repeat it with a different workflow to test whether the learning holds.

How do I know if I am reviewing AI work well?

One sign is being able to say, before looking at the result, what would count as an error. Another is having reference examples with known answers and comparing the agent's output against them periodically. If you approve almost everything without finding flaws, doubt your review method before concluding the system is perfect.

Sources

• Anthropic: Building effective agents
• Anthropic: Demystifying evals for AI agents
• NIST AI Risk Management Framework
• OECD AI Principles