Why AI Employees Need Clear Limits and Human Approval

In a recent LinkedIn post, Luís Rodrigues discusses how an AI employee should handle incomplete data when a task is governed by explicit permissions and human approval. Rodrigues presents a Slack scenario in which an AI assistant, Viktor, misses a deadline because a finance export is incomplete. The post is a paid partnership, and its example makes a case for treating governance and oversight as part of AI deployment.

“What he didn’t do: wake anyone up, send anything, or change anything on his own.”

Holding a report can be the right decision

In Rodrigues’s account, Viktor is instructed to prepare work but not release it without approval. When the export arrives with missing data, the assistant holds the report, identifies a vendor maintenance window as the cause, documents a fix and waits. The report goes out later that morning, after a person approves it.

That sequence frames the delay as a safeguard rather than a failure. Rodrigues emphasizes that the leadership numbers were a few hours late but correct. The example illustrates a practical trade-off: in a consequential workflow, producing an answer on schedule may matter less than avoiding a report built on bad data.

Permissions shape the role of the model

Rodrigues connects the scenario to two principles he says he had discussed in earlier posts. One is that the system around a model determines what it can do and when a human must step in. As he puts it:

“The harness matters more than the model. Permissions decide what waits for a human.”

In this view, safe delegation depends on defined permissions, not simply on the model’s capabilities. The assistant can investigate and prepare, while actions with external consequences remain subject to approval. Rodrigues’s account therefore presents the operating rules as central to how an AI employee behaves.

Agree on goals and limits before delegation

The second principle in the post is to establish the objective and boundaries before work begins. Rodrigues writes:

“Agree on the goal and the limits before takeoff. Then let the team act inside them.”

The distinction matters in his example: Viktor is allowed to trace the problem and propose a fix, but not to send the report or change anything independently. That division gives the assistant room to make progress while preserving a human decision point.

Human ownership remains part of the workflow

Rodrigues argues that an AI employee does not eliminate the need for an owner. Instead, he suggests, a well-bounded system can bring a decision to that owner before a problem becomes cleanup. His post offers a compact illustration of that claim, though it remains a promotional example rather than independent evidence about performance across organizations.

For businesses considering AI in operational workflows, the scenario raises concrete questions: Which tasks may the system prepare? What information should stop work? Which actions require explicit approval? Rodrigues’s central point is that those limits should be set before delegation begins.