> ## Content Index
> Fetch the complete content index at: https://futureinuse.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# The risk with AI agents is permission to act
- URL: https://futureinuse.ai/permissao-agentes-de-ia-risco/
- Published: 2026-07-29T15:59:00.000Z
- Updated: 2026-10-22T12:20:40.000Z
- Description: With agents, the question stops being which model a company uses and becomes what each agent is authorized to do. In the author's view, control has to be built into the architecture.
- Author: Rafael Cordeiro
- Tags: Rafael Cordeiro, colunista, #historico-linkedin

The biggest risk of AI in companies may not be that it makes mistakes. It is that it has permission to act when it does.

## From software that shows to agents that act

For a long time, we used software as a tool. The system showed a piece of information. A human analyzed it. A human decided. A human executed.

With AI agents, that logic starts to change. Now the system can interpret a situation, make a decision, access other systems and carry out an action.

That completely changes the conversation about AI adoption.

## What is this agent authorized to do

The question is no longer just which model we are using. It becomes what this agent is authorized to do.

- Can it access the CRM?
- Query financial data?
- Change an order?
- Send a proposal?
- Approve a payment?
- Run code?
- Hire another agent to do part of the task?

And perhaps the most important question. Who stops everything when something goes off script?

## Control as architecture

The more autonomy we give agents, the more we need to think about identity, permissions, policies, observability, auditability and execution limits.

Not as a compliance layer added later, but as part of the company's own operating architecture.

## Reconstructing the path of a decision

There is an interesting analogy with something I learned working for years with data and measurement.

In Retail Media, it was not enough to know a campaign was running. What really mattered was being able to connect exposure, action and outcome to understand the incremental impact it generated.

With agentic systems, I see a similar evolution. It is not enough to know that the agent executed 2,000 tasks or consumed millions of tokens.

We need to be able to reconstruct the path that runs from intent to decision, passes through the agent, the model, the tool and the action, and arrives at the outcome.

And understand not only what happened, but why it happened, under which authorization and what outcome it produced.

## From the assistant era to the execution era

We are leaving the era of the AI assistant and entering the era of AI execution.

Companies that want to operate this way will need something that still gets far less attention than the models themselves. An operating layer able to control, observe and evaluate that execution.

Buying AI is relatively easy. Giving it access to systems is too.

Building an organization able to delegate work to agents without losing control, context and accountability will be a completely different discipline.

### In short

- The biggest risk may not be that AI makes mistakes, but that it has permission to act when it does.
- Agent control is part of the operating architecture from the start.
- Delegating work to agents without losing control, context and accountability is a new discipline.

> **What has changed since July 29, 2026 · Editor's note**  
> On September 25 and 30, Microsoft announced a [central registry of IT-approved plugins](https://futureinuse.ai/microsoft-copilot-agentes-plugins-governanca/) for Copilot, with spending policies and a choice of model families by group.

**Further reading**

- [AI Gateway, the control layer for scaling AI in the enterprise](https://futureinuse.ai/ai-gateway-camada-de-controle-ia-empresarial/)
- [Microsoft swaps agents for plugins in Copilot](https://futureinuse.ai/microsoft-copilot-agentes-plugins-governanca/)
- [Agents across companies and the paradox of invisible decisions](https://futureinuse.ai/agentes-entre-empresas-decisoes-invisiveis/)

---

*Originally published on LinkedIn on July 29, 2026\.* [*See the original post*](https://www.linkedin.com/feed/update/urn:li:activity:7488261898975580160/?ref=futureinuse.ai)*. This version has been revised and expanded with context and sources. The views are those of the author, a co-founder of OperOS AI.*