Examples

Scam check

Is this message likely to be a scam?

A message arrives by text, email or chat. Before a customer acts on it, or before your team does, check whether it looks like a scam.

The request

{
  "model": "myles-g-12b",
  "state": {
    "channel": "text message",
    "sender": "unknown number",
    "message": "Parcel service: your parcel is held. Pay the 1.99 redelivery fee within 24 hours: parcel-redeliver-uk.co/pay"
  },
  "questions": {
    "scam": {
      "type": "noul",
      "instructions": "Is this message likely to be a scam?",
      "criteria": {
        "yes": "It tries to get money, card details, passwords or personal data by deception",
        "no": "It is a genuine message or harmless"
      }
    }
  }
}

Reading the answer

"scam": {"type": "noul", "noul": 0.97}

Warn or block above a high threshold, such as 0.95. Below 0.05, let the message through. Between them, show a gentler caution or send it for review.

What to watch

• Genuine messages flagged as scams: each one is a customer you have worried, so keep the upper threshold high.
• New patterns: scams change quickly. Review a sample of the middle band every week.
• Put the sender, channel and any link in the state as separate fields; they carry much of the evidence.