
Decision Making With Superintelligence: A Manager's Checklist
A practical checklist for managers to test an AI recommendation before acting, separating evidence, premises and accountability, with clear limits.
Want to apply this in your company?
Book a call with Juan →Before accepting an AI recommendation, a manager needs to separate three things: what was verified against a source, what was assumed as a premise, and who answers if the decision goes wrong. Without that separation, a well-written, confident answer looks like a finished decision, even when it depends on incomplete data or on an assumption nobody reviewed. The checklist below works for recommendations on pricing, inventory, customer support or hiring, including when the system is presented as superintelligence.
Key takeaways
- A recommendation is a working hypothesis until the evidence behind it has been identified and checked.
- Premises must be written down, because that is where recommendations usually fail without anyone noticing.
- Responsibility for the decision stays with a named person, never with the system.
- Every relevant recommendation should leave a record of what was asked, what was used and who approved it.
- When the check does not close, the right move is usually a small pilot, not full acceptance.
Why separate evidence, premise and accountability
The word superintelligence suggests a capability beyond what the team can judge. In practice, though, an AI system works with what it receives: data, instructions, recent history and a few assumptions built into the request. The quality of the recommendation depends far less on the tone of the answer and far more on the quality of those inputs.
Evidence is what can be checked: a sales report, a spreadsheet exported from the system, a service record, a signed contract. A premise is what the recommendation assumes without showing, such as "demand this month will look like last month" or "the customer accepts the standard lead time". Accountability is the answer to "who signs this decision and who answers for it". Mixing the three makes a weak premise look like a fact and leaves the decision without an owner.
Checklist before acting
The checklist works best when it is done in writing, on one page, and reviewed by the person who will carry out the decision. The steps can be these:
- Record the exact question asked of the system, with the context and the data sent.
- List the sources the recommendation used and confirm they exist, are current and cover the case.
- Identify each explicit premise and each implicit premise, such as period, audience, region or exceptions.
- Check the calculation or reasoning against at least one real case the manager knows well.
- Confirm that the recommended action is within what the system can do and what the company allows.
- Define who approves, who executes and who follows the result over the next few weeks.
- Decide what will be measured to know whether the recommendation worked, and by when.
If one of these steps cannot be completed, that is not a reason to discard the recommendation. It is a reason to reduce the size of the action.
How to spot a hidden premise
Hidden premises are the most fragile point of any recommendation. They appear when the system uses a number nobody supplied, treats a rare case as the rule, or ignores a constraint the team knows about but never wrote into the request. A few questions help uncover them:
- Does this recommendation depend on a specific period that may not repeat?
- Does it assume that all customers or products behave the same way?
- Does it account for legal, contractual or operational constraints that are outside the data sent?
- Does it rely on a figure entered by hand that may be wrong?
- If one of these premises is wrong, what changes in the decision?
The last question is the most useful. If the answer is "almost nothing changes", the premise is secondary. If the answer is "the whole decision changes", it must be verified before any action.
Illustrative scenario: restocking a building materials store
Illustrative scenario: a building materials store receives a recommendation from its AI tool to buy more of one item ahead of a seasonal peak. The recommendation cites sales history from recent months, which is evidence. But it assumes that a large customer's construction project will continue at the same pace, and that information is not in any system. It exists only in the head of the responsible salesperson. During the check, the manager flags this premise, asks the salesperson and learns the project has finished. The purchase is reduced, and the rest is approved with weekly follow-up. No part of the process depended on trusting the system blindly, and no part depended on distrusting it on principle.
Who answers for the decision
Responsibility needs a name. Inside a company, an AI recommendation should be treated as the suggestion of a colleague who cannot legally answer for their own actions. The person who approves is the one who answers. For that reason, it helps to separate the role of whoever proposes, whoever reviews and whoever decides. In operations that handle personal data, this separation also connects to Brazil's LGPD. When an automated decision affects a person, such as credit, access to a service or an evaluation, the data subject has the right to understand the criteria and request a review. That point is covered in detail in the article on LGPD automated decision-making and the right to review. If the recommendation involves sending data outside the country, international data transfer when using AI providers also belongs in the check.
To get an overview of terms and categories before taking the topic to the board, the superintelligence glossary for executives helps align the conversation without jargon.
When to decline, reduce or test before accepting
Not every recommendation has to be accepted in full. Three paths are usually safer than full acceptance:
- Decline when the evidence does not exist, when the main premise cannot be verified, or when the action goes beyond what the company has authorized.
- Reduce the size of the action when the recommendation is plausible but depends on a premise that has not yet been confirmed.
- Test at small scale, with a success criterion defined in advance, when the recommendation looks good and the cost of being wrong is controllable.
A test can be simple, such as applying the recommendation in one region, to one product or to one group of service requests, and comparing the results with what would have happened without it. What matters is that the result is measured with the same yardstick defined at the start.
How Juan Carlo's method applies to this check
In consulting and deployment work for agents and automations, Juan Carlo's method follows four stages: map the current process and its baseline, limit the data and actions the system can touch, pilot a real flow with records and an agreed success metric, and hand the operation to the team, deciding whether to expand, adjust or stop. This sequence follows the same logic as this checklist. Mapping reveals the premises already present in the process, limiting defines what the system can touch, the pilot turns a recommendation into evidence of its own, and the handover defines who answers. Managed operations are delivered through Be Human (be-human.ai).
Frequently asked questions
What should a manager check first in an AI recommendation?
Start with the exact question that was asked and the data sent to the system, because many problems begin with a poorly framed request. Next, identify the main premise and verify whether it holds in the specific case. Only then assess whether the action is permitted and who will approve it.
Can a recommendation with a good justification be accepted without verification?
No. A well-written justification shows internal coherence, but it does not prove that the data is correct or that the premise applies to your case. When the decision is significant, the justification should be checked against an independent source or against a case the team knows well.
How do I know whether to test before accepting?
Test when the recommendation seems reasonable but depends on a premise that has not been confirmed, or when an error would be expensive to reverse. Define in advance what will be measured and over what period, so the result cannot be interpreted according to what anyone hoped would happen. If the test cannot be compared against anything, the manager still does not have enough evidence.
Who is responsible when a decision based on AI turns out wrong?
Responsibility stays with the person or area that approved the decision, not with the system. That is why recording the request, the sources and the approval is essential for understanding what happened and correcting the process. In cases involving personal data, the company must also follow LGPD obligations on automated decisions and review.