
Superintelligence for Managers: What to Delegate and What to Keep
A practical guide for managers to split tasks an AI agent can prepare from decisions that stay with the accountable owner, with criteria and an example.
Want to apply this in your company?
Book a call with Juan →Superintelligence for area managers comes down to one practical decision: which tasks an AI agent can prepare on its own, and which decisions stay with the person who answers for the result. The agent can pull data, summarize history, draft replies and flag inconsistencies, but it should not approve spending, promise deadlines or change area rules without human review. The split becomes simple once each task is classified by the cost of a mistake and how easily that mistake can be undone.
Key takeaways
- Delegate the preparation: collecting, summarizing, drafting and consistency checks can go to the agent.
- Keep the decision: spending approval, policy exceptions, sensitive negotiations and priority changes stay with the manager.
- Classify each task by the impact of an error and whether the result can be reversed.
- Write down what the agent may read, what it may execute and when it must hand the conversation to a person.
- Start with one real workflow, with an action log and an agreed metric before widening the scope.
What the term superintelligence means for someone who runs a team
Marketing material uses the word loosely, often to describe any assistant that answers fluently. For someone who has to decide, the useful question is not whether the tool is superintelligent. It is what the tool does reliably, with which data, within which limits, and who answers for the outcome. A language model produces plausible text, but it does not know what is happening in your stock, your calendar or your discount policy unless that information is supplied from a controlled source. For a more careful reading of the vocabulary, see the superintelligence glossary for executives.
What to delegate: tasks the agent can prepare
Delegating does not mean taking the person out of the process. It means having the agent do the preparation work, which usually consumes time and rarely requires the judgment of the person who signs off. As a rule, these tasks are good candidates:
- Gathering information from authorized sources, such as a spreadsheet, a ticket system or the area's document folder.
- Summarizing the history of a customer, supplier or case before a meeting.
- Drafting replies, proposals, reports and meeting notes, always marked as drafts for review.
- Checking inconsistencies, such as an order without an invoice, an empty field or a deadline that has already passed.
- Organizing the backlog of pending items and suggesting priority according to a rule the team already uses.
- Filling in system records when the format is fixed and the data source is reliable.
What these cases have in common is that the agent's output can be checked before it has any effect in the real world.
What to keep: decisions that should not leave your hands
The decisions that stay with the manager share one trait: they affect money, a person, a promise made to a customer, or an exception that could become a precedent. Keep these with the responsible person:
- Approving spending, discounts, refunds or any exception to policy.
- Negotiating with a customer, supplier or employee in a sensitive situation.
- Changing a deadline, price or commercial term that has already been promised.
- Evaluating someone's performance or making any decision about their career.
- Responding to a serious complaint, a legal question or a request involving sensitive personal data.
- Authorizing the agent to trigger an external system with an irreversible effect, such as sending a collection notice or cancelling a contract.
The rule is not the size of the task. A short message can carry far more risk than a long report if it promises something the company cannot deliver.
Decision guide for classifying each task
Before handing a task to the agent, work through the four questions below in order. The first one that gets a yes sets the level of control.
- Can the task be undone without material cost, and would an error show up early? Then the agent may execute it, with a record of each action.
- Does the task affect money, a promised deadline or a person? Then the agent prepares the output and the manager approves it before anything is sent or posted.
- Does the decision depend on a judgment criterion the company has never written down? Then the manager decides, and only afterward is the criterion recorded so the next case has a basis.
- Is the data source incomplete, duplicated or out of date? Then the agent does not act on it until the source is corrected.
This guide works best when the team writes it in its own words, using real examples from the previous week.
Illustrative scenario: the purchase request queue
Illustrative scenario: an operations manager receives purchase requests from three departments over the course of a day. An agent reads each request, checks whether an equivalent item is already in stock, prepares a comparison of quotes already on file, and sorts the requests into a list of those inside the approved value range and supplier policy. Requests outside that range, or for items with no history, appear on a separate list with the reason written next to each one. The manager reviews the in-range list and approves it in a batch, and decides the remaining requests personally. Every approval is recorded with name, date and justification. The gain shows up in faster preparation and cleaner records, while the purchasing decision stays with the person who is accountable for it.
How to apply the map, limit, pilot and hand over method
In our approach, delegation starts by mapping the current process and a baseline: how many people handle the task, where it gets stuck and what already works. Next, we limit data and actions so the agent sees only what it needs and cannot execute anything outside its scope. Then we pilot one real workflow, with logs and a success metric agreed before the start. Finally, we hand the operation over to the area and decide whether to expand, adjust or stop. This sequence is an implementation practice, not a promise of results. For an example of how an agent operates today, without overstating its capabilities, see superintelligence in practice: what an agent does today.
Data, privacy and decision records
If the agent handles personal data of customers or employees, the manager needs to know which data enters the workflow, where it is processed and who can access it. Brazil's LGPD requires that clarity, and the national data protection authority (ANPD) publishes guidance that companies can use. When an automated decision affects a person's interests, the right to human review under Article 20 of the law should be built into the process design. See the details in LGPD automated decision-making and the right to review under Article 20. If you want to measure team performance with AI without turning measurement into surveillance, focus on process metrics rather than monitoring each individual conversation, as discussed in how to measure AI productivity without surveillance on your team.
Finally, keep a simple record of the decisions the agent prepared and of the ones the person changed. That history is what lets you adjust the workflow based on facts rather than impressions.
Frequently asked questions
Do I need a technical team to start delegating tasks to an agent? Not necessarily for the first workflow, as long as the area is clear about the process and has a named owner. The technical side comes in to connect data sources, control permissions and keep records. Without those pieces, results tend to stay limited to drafts produced outside the company's systems.
Can the agent approve small purchases without checking with anyone? It can, if the policy is written down, the value and item type fall within a defined range, and every approval is recorded with a reason. Even so, it is wise to start with human approval for a period and only then consider automatic execution for simple cases. Any change to the range should go back to the manager.
How do I know whether a task is ready to be delegated? It is ready when the input is standardized, a person can check the output in a short time, and an error would not cause irreversible damage. If the area still cannot describe the rule clearly, the problem is the process, and the agent will only make that gap more visible. In that case, mapping the process before automating it is the most useful step.
When should I stop a delegation pilot? Stop, adjust or expand based on the metric agreed at the start, not on the feeling that the tool is impressive. If the agent creates rework, hides errors or requires more review time than the manual process, the pilot should be revised or ended. Ending it is also a legitimate decision and should be recorded with the reason.