Skip to main content

How Clara decides

In plain words

Before Clara answers anything, it asks three questions. Is this safe to answer? What is this person really asking? Is the answer we wrote actually good? Only when all three pass does the answer go out. Every step is written down.

This page is for everyone. No technical background needed.

The three questions​

Question 1: is this safe to answer?​

Clara checks the message against your rules and general safety rules. Examples of things that raise a flag:

  • The customer threatens legal action.
  • Answering would mean promising a refund in writing before it was approved.
  • The message contains card numbers or other private data.

A flag does not always mean silence. It can mean caution (answer carefully), escalate (a person should handle this), or stop (do not answer).

Question 2: what is really being asked?​

"Where is my order?" can mean "I am worried" or "I am about to demand a refund." Clara builds a short internal briefing: the facts, the mood, the risk, and the recommended way to handle it. On the higher service levels you can receive this briefing yourself, in plain English.

Question 3: is the answer actually good?​

The written answer does not go straight out. Independent reviewers, separate from the writer, look at it and vote: approve, caution, or stop. If they stop it, the customer does not see it, and the record shows the reasons.

What the outcomes mean​

OutcomeWhat it meansWhat happens
AnsweredEverything passedThe customer gets the answer
CautionFine, but flaggedThe answer goes out and the flag is recorded
EscalatedA person should decideYour team is notified with the briefing
StoppedNot safe to answerNothing goes out, the reasons are recorded

Why "no" is a feature​

An AI that never says no will eventually say something expensive. Clara saying "stopped, here is why" protects you. The record shows the refusal was deliberate and rule-based, not random. Ask us to see a real stopped example from the beta.