Automate without losing control: how to decide what to delegate to AI

When a company starts using AI day to day, the first question is usually what it can do. The second, more important one comes soon after: what should it do without anyone reviewing it?

The problem isn't AI, it's the level of autonomy

An assistant that drafts an email and one that sends it to the client use the same technology, but the risk is very different. Treating every task the same leads to one of two extremes: either everything is reviewed and automation saves no time, or nothing is reviewed and a mistake reaches the client.

Three levels of delegation

• Look up: the AI searches, summarises and answers internal questions. It changes nothing.
• Propose: it prepares a draft, a reply or a task, and a person approves it.
• Execute: it acts on its own on narrow, repetitive tasks that are easy to undo.

The useful question isn't "can the AI do it?", but "what happens if it gets it wrong?".

How to choose the level for each task

Three questions help you decide:

1. Does the result leave the company? If it reaches a client or a supplier, start at "propose".
2. Can it be undone? Moving a task to another column takes a second to fix; a payment or a sent email does not.
3. Is there a clear criterion? If two people on the team would answer differently, the AI still needs review.

The usual approach is to start with more supervision and move up a level as the team confirms the results are reliable.

Control is context too

Delegating with confidence requires limits defined in advance, not decided on the fly. That's why at Unelia control is one of the layers of context: you define what each person or agent can do, and the AI works within those limits.

The goal isn't for AI to do more things, but to do the right ones with the team's trust.