---
title: "Human approval or a mandate with limits?"
url: https://knowyouragent.ch/en/knowledge/comparison/approval-or-mandate
lang: en
updated: 2026-10-17
license: © Innopulse Consulting GmbH
---

# Human approval or a mandate with limits?

Many companies start cautiously with AI agents: a person approves every action. That is safe as long as the numbers stay small. After that, a framework is needed in which the agent may act on its own.

[Leutrim Miftaraj](https://knowyouragent.ch/en/about/leutrim-miftaraj), Founder, Innopulse Consulting GmbH
Published 17 October 2026

## Key points

- A human approval decides case by case; a mandate sets the limits in advance.
- Approvals without a mandate do not scale and lead to blind confirmation.
- Both work best together: routine in the mandate, exceptions for approval, every approval as a signed decision.

## What this is about

With human approval, often called [human in the loop](https://knowyouragent.ch/en/knowledge/glossary/human-in-the-loop), the agent presents every action and a person decides. This makes sense for new deployments, because it builds trust and shows mistakes early.

A [mandate](https://knowyouragent.ch/en/knowledge/glossary/agent-mandate) sets the limits in advance: which actions, up to which amount, with which counterparties, for how long. Within these limits, the agent acts on its own, and every action is checked against the mandate before execution.

## The comparison

| Feature | Human approval | Mandate with limits |
|---|---|---|
| Decision | Case by case | In advance for a framework |
| Speed | Depends on availability | Milliseconds |
| Scaling | Limited by people | Any number of actions |
| Typical risk | Blind confirmation with many requests | Limits set too broadly |
| Evidence | Only if recorded as a decision | Mandate, decision and receipt |
| Suited for | Exceptions, new deployments | Routine with clear rules |

## When to use what

- **Approval** in the early phase of a deployment, for high amounts, new counterparties and unusual patterns.
- **Mandate** for recurring actions with clear rules, once the deployment runs stably.
- **Combination** as the permanent state: the mandate covers routine, and a person decides above it.

## How the two work together

The decision service checks every action against the mandate. If it lies within it, it is executed. If it exceeds it, it goes to a person, and the approval is recorded as a signed decision. Analyzing the approvals shows whether the framework fits: if almost all exceptions are approved, the mandate is too narrow; if many are denied, this points to mistakes or attacks.

## Common misconceptions

- **“Approving every action is safest.”** Only with small numbers. With many requests, attention drops and approvals become a formality.
- **“A mandate means nobody looks anymore.”** The mandate sets limits; analysis of decisions and receipts remains.
- **“An approval by click is evidence.”** Only if person, time and action are recorded and signed.

## The role of KYA

KYA checks whether an action lay within the mandate or was covered by a decision. An exception without a recorded approval is shown in the certificate as not proven. The comparison [Decision before or log after?](https://knowyouragent.ch/en/knowledge/comparison/decision-before-or-log-after) describes how decisions before execution work.

## Decision guide in three questions

- Does the action recur frequently with clear rules? Then it belongs in a mandate.
- Is the action rare, expensive or unusual? Then it needs approval by a person.
- Must it later be shown who decided? Then every approval is recorded as a signed decision.

## Frequently asked questions

### Should a person approve every action of an AI agent?

At the start of a deployment, this can make sense. In the long run, it does not scale; a mandate for routine and approval for exceptions is better.

### Which is better: approval or mandate?

Neither alone. A mandate covers routine quickly and verifiably; human approval remains for exceptions and high risks.

### How do I set the threshold for approvals?

By risk and experience. Many start low and raise the threshold once analysis shows that routine runs stably.

### How does an approval become provable?

As a signed decision with the approving person, time and action. A click in an interface without this evidence is hard to prove later.

### What does analyzing the approvals show?

Whether the mandate fits. Many approved exceptions point to limits that are too narrow, many denied ones to mistakes or attacks.

### How many approvals per day are reasonable?

That depends on complexity. When people just click through approvals, the number is too high, and the mandate should cover more routine.

### Who should grant approvals?

People with the appropriate authority for the action, for example from the company's authority rules. An approval should never come from someone who triggered it.

### What happens if nobody is available for an approval?

The action waits or expires after a defined time. Urgent cases need a substitution rule, never an automatic approval.

### Can an approval be granted afterwards?

Subsequent approval is possible but does not prove that the action was allowed beforehand. For actions with consequences, approval should come before execution.

## Sources

- NIST: Least Privilege: [https://csrc.nist.gov/glossary/term/least_privilege](https://csrc.nist.gov/glossary/term/least_privilege)

---
Quelle: https://knowyouragent.ch/en/knowledge/comparison/approval-or-mandate · KYA – Know Your Agent