
Human Approval in Superintelligence: A Risk-Based Decision Scale
A risk and impact scale to decide when your AI agent acts alone and when a person must approve, with practical business decision examples.
Want to apply this in your company?
Book a call with Juan →Human approval in superintelligence means defining, before any automation, which actions an agent may take without review and which ones need a person to authorize them. The practical answer does not depend on how advanced the model is. It depends on two questions: how much damage a wrong decision would cause, and how hard it is to undo. A scale that combines risk, impact and reversibility separates what the agent does alone from what goes back to the person responsible.
Key takeaways
- An agent's permissions should follow the impact of each decision, not a general trust in the model.
- Irreversible decisions involving money, personal data or someone's rights stay with a person.
- Routine, low-impact, reversible actions can be executed by the agent, with logging and periodic review.
- The scale must be written into the tool permissions, not only into the prompt instructions.
- Before releasing any band of autonomy, test the workflow with simulated data and a controlled pilot.
What changes when we talk about superintelligence
Superintelligence is still a subject of research and debate, not something a company buys today. What businesses run day to day are AI agents with access to tools, systems and customers. The difference between the two scenarios is one of scale: the more a system can act, the more it needs a clear design of who decides and who answers for the result.
That is why the scale I propose works for a WhatsApp support agent as well as for any system that may gain more autonomy later. The principle stays the same: autonomy proportional to risk, with the right person at each decision point. To understand the conceptual difference between these terms, read the difference between AGI, ASI and AI agents.
The scale: risk, impact and reversibility
The question "is the model reliable?" is the wrong one for deciding approval. Even a model that is right most of the time can make a mistake in a case that is costly, and average confidence does not protect the company from that case. The useful question is different: if this action goes wrong, who is affected, how much does it cost to fix, and can it be undone?
Risk is the chance of error for that type of decision, considering ambiguous data, incomplete information and requests outside the script. Impact is the consequence of the error for the customer, the company or third parties, in financial, legal and reputational terms. Reversibility is the ability to undo the action without significant harm, such as refunding a wrongful charge or resending a corrected message. The NIST AI Risk Management Framework treats AI risk as a continuous process of governance, mapping, measurement and treatment, and that logic fits well with a scale you can revise.
Four decision bands
In practice, the scale becomes four bands. Each agent action falls into one of them, and the band defines what happens before it runs.
- Band 1, low risk and low impact, reversible: the agent executes on its own and logs the action. Typical examples are confirming a slot that is already free or sending a reminder covered by the service playbook.
- Band 2, low risk but high impact: the agent prepares the action and a person approves it. Example: drafting a discount offer for a known customer.
- Band 3, medium or variable risk with limited impact: the agent acts within explicit limits, such as a maximum amount or a defined case type, and a sample of its actions is reviewed afterward.
- Band 4, high impact or irreversible: the person decides and the agent only recommends, organizing the information needed for the decision.
The scale does not need to be perfect to start. It needs to be clear enough that two people on your team classify the same action the same way.
Decisions that should never run alone
Some decisions stay with a person even when the agent is well evaluated. These include authorizing payments and refunds above the agreed limit, cancelling contracts, changing prices, denying a service or a right, responding to a public complaint, and any processing of sensitive data.
When decisions made solely by automated means affect a person's interests, Brazil's data protection law (LGPD) gives the right to request a review under Article 20. The design therefore needs a real path to human review, not just a form that never gets answered. For the legal side, the article on LGPD automated decisions and the right to review explains what to check.
Step by step: build your own scale
Follow this sequence to build the scale in your company. It avoids starting from the tool and makes sure the decision comes from the process.
- List the actions the agent can execute today, checking the real tools and integrations, not only the ones you imagine.
- For each action, record the possible damage from an error and who would be affected.
- Mark whether the action is reversible and how long an error would take to correct.
- Assign a human owner to each class of decision, by name and role.
- Set concrete limits, such as maximum amount, customer type, topic or channel, that separate one band from the next.
- Test the scale in a simulated environment before releasing any live action.
- Review the scale after a period of operation, based on logs and the exceptions found.
Our method follows the same logic: map the current process and baseline, limit data and actions, pilot a real workflow with logs and a success metric agreed in advance, then hand over the operation and decide whether the workflow should be expanded, adjusted or stopped. The approval scale is one of the first pieces that enters that limiting stage.
Illustrative scenario: layered approval in purchasing
Illustrative scenario: a building materials store uses an agent to compare quotes from suppliers already registered. Up to an agreed amount, the agent can issue the order to an approved supplier, because the action is reversible within the cancellation window and the impact is contained. Above that amount, the agent prepares the comparison and sends it to the purchasing manager, who approves or rejects it. When a new supplier appears, the decision always goes back to a person, because registration, delivery and quality risks cannot be judged from the available data alone. A similar case appears in an AI agent for procurement that compares quotes and asks for approval.
Recording approvals and common mistakes
An approval only has value if the person has enough information to decide. Each approval request should show what the agent proposes, the rule that triggered the review, the data used and the recommended alternative. It is also necessary to record who approved, when and with what decision, so the company can audit the process later.
There are three frequent mistakes. The first is asking for approval on everything, which causes fatigue and leads people to approve without looking. The second is giving the agent a tool with overly broad permissions and relying only on the prompt to stop it from using that tool. The third is ignoring that messages from customers can contain instructions that try to change the agent's behavior. The OWASP guidance on prompt injection describes this risk and recommends keeping important controls outside the text the model reads. For permissions, the principle of least privilege is the starting point.
Frequently asked questions
Is superintelligence something a company needs to approve today? Not yet, in the sense of systems with general capability above human level. What companies operate are AI agents with tools and permissions, and the approval scale already applies to them.
Who should approve an agent's action? It should be the person accountable for the result of that process, such as the finance lead for payments or the support lead for customer exceptions. In small companies one person may hold several roles, but the owner of each class of decision must be defined in writing.
Will human approval slow down customer service? It only delays the decisions that truly need review, which are usually a minority when the scale is well built. Low-risk actions stay automatic, and the review arrives with the context already prepared, which shortens the time to decide.
Do I need a lawyer to define these rules? For decisions that affect customers' rights, such as review of automated decisions and processing of personal data, it is worth consulting a legal professional. The operational scale can be designed by the operations team, but the legal framing should be validated before production use.