---
id: KB-MA-004
url: https://app.codecontract.io/help/marta/when-not-to-trust-the-answer
idioma: en
categoria: marta
audiencia: usuario
nivel: intermedio
actualizado: 2026-08-13
tambienEn: [es]
relacionados: [KB-MA-002, KB-MA-003]
citadoPor: [KB-MA-007, KB-CL-004, KB-MA-012, KB-MA-015, KB-MA-016, KB-MA-019, KB-MA-020, KB-MA-008, KB-MA-006]
---

# When not to trust the answer

_Where it is right, where to check, and why that does not make it useless._

**Responde a:** does the ai assistant make mistakes · can i trust what marta says · the ai gave me a wrong figure · verify the assistant's answer

An assistant that answers in seconds across hundreds of cases is useful precisely because it is fast. What you need to know is where that speed is paid for, so you check there and not everywhere.

## Where it is right and where to check

| Kind of question | Reliability | What to do |
| --- | --- | --- |
| "Which cases are stuck?" | High: it is a filter | Use it |
| "Whose insurance is expiring?" | High, if the dates are set correctly | Use it |
| "Summarise what happened on this case" | Good for orientation | Read the case before deciding |
| "How much do we owe this supplier?" | Depends on that data being here | Check it in your own system |

> [!IMPORTANT]
> Before a decision with consequences — a payment, granting site access, closing a case — check the fact at source. The assistant takes you there in seconds; that is its job, not to replace the check.

## Why it sometimes falls short

**En corto**

- It only knows what is here: what was agreed by phone is not.
- A misread value in a document will be repeated as is: it does not invent it, but it carries it along.
- If the question is ambiguous, so is the answer.

> [!NOTE]
> Asking with a specific timeframe and subject removes most misunderstandings. "What is missing?" is a bad question; "what does Muñoz Engineering still need to start on site?" is a good one.

> [!WARNING]
> If it gives you a figure that surprises you, neither dismiss it nor assume it: open it. There is almost always something real behind it — an expired document, a case nobody was watching — even if the reading was not exact.

**Does it invent data?**

If it cannot find something, it says so rather than filling it in.

**Can it see what I cannot?**

No. It inherits your permissions.

**Is it worth it if I have to check?**

Checking a fact it took two seconds to find is still far faster than hunting for it.

## Ejemplos

**The assistant says a supplier's insurance lapsed two months ago.**

- The document is opened
- The date had been misread: it expires next month

→ The value is corrected, the alert repositions, and the check was still worth it because nobody had looked.

**The answer sounds confident and is wrong.**

- Checks which documents it comes from

→ Verification is one click.

**The answer is used for an important decision.**

- Cross-checks against the original document

→ The decision rests on the source.

**The answer rests on a misread figure.**

- Corrects the figure in its document

→ The next answer is already right.

**Something outside the scope is asked.**

- Checks which files it can see

→ It becomes clear why the answer is partial.

**The answer is copied verbatim to a client.**

- Reviews the text before sending

→ What goes out carries your judgement.