---
id: KB-MA-013
url: https://app.codecontract.io/help/marta/what-marta-does-without-asking-and-what-it-does-not
idioma: en
categoria: marta
audiencia: usuario
nivel: intermedio
actualizado: 2026-08-13
tambienEn: [es]
relacionados: [KB-MA-005, KB-CL-007]
citadoPor: [KB-MA-016, KB-TL-029]
---

# What Marta does without asking and what it does not

_The line between looking things up and acting, and who decides where it sits._

**Responde a:** can the ai do things on its own · what actions does the assistant execute · does the assistant send emails by itself · limits of the ai agent

The question always comes, and rightly: if the assistant can read everything of yours, what can it do with it? The short answer is that looking things up is one thing and acting is another, and the second is subject to what you decide.

## The three layers

| Layer | Examples | Who authorises |
| --- | --- | --- |
| Consulting | Summarising, searching, comparing, explaining | Your own permissions: it sees what you see |
| Preparing | Drafting, building a process, proposing a list | The same, but nothing leaves here |
| Acting | Sending, requesting from a third party, closing, approving | Configured, and by default it asks |

> [!IMPORTANT]
> The first layer is bounded by your permissions, not the assistant's. If you cannot see a file, it does not appear in an answer either: there is no back door through which the AI sees more than you.

## Where to draw the line

1. **Let it consult and prepare, always** — That is where nearly all the time saving is and it can break nothing.
2. **Let it act only on the reversible** — Opening a task, sending a reminder, flagging something for review.
3. **Have it ask on anything going outward** — Anything reaching a third party is better with a person present.
4. **And never on what decides** — Approving, rejecting, blocking a supplier. Someone with a name answers for that.

> [!WARNING]
> Point three is the one most quickly relaxed "because it's going well", and the one that goes worst when it fails: a wrong message to two hundred suppliers is not undone by a correction. What goes outward deserves a human glance even at thirty seconds' cost.

## What gets recorded

**En corto**

- Every executed action is in the activity log, marked as done by the assistant.
- What it consults leaves a trace too, exactly as if you had opened it.
- And what it executes consumes the same: an automatic action costs what a manual one costs.

That last point has a practical consequence: a rule firing a thousand times spends a thousand, and it is worth checking the breakdown in the first month it is switched on.

> [!NOTE]
> You can start with no ability to act and widen it later. That is the recommended path: first check its answers are reliable in your case, and only then let it execute anything.

**Can it delete things?**

Not something worth delegating, and by default it does not.

**Does it learn from our documents?**

It uses them to answer you; see the article on what the AI does with your documents.

**Can I see what it has done?**

Yes, in the activity log, like any user.

## Ejemplos

**A company wants the assistant to send reminders and approve simple documentation.**

- Lets it handle reminders and tasks
- Keeps approval with a person

→ Recovers the chasing time without delegating any decision it answers for.

**It is expected to launch requests unprompted.**

- Checks what it does on its own and what it does not

→ No unrequested actions are assumed.

**It is asked to write to a supplier and nobody reviews.**

- Reviews the text before sending

→ What goes out carries your judgement.

**It is assumed a figure has been corrected.**

- Checks the figure in its document

→ The real state is known.

**It is asked something that requires a decision.**

- Takes the decision and uses the answer as support

→ Each does their own job.

**It is expected to act on files it cannot see.**

- Checks its scope

→ Expectations are set.