Made in 10Docs
GuidesAPI
Characters

Call analysis

After each call, a summary, the values you asked for, and a verdict against your own success criteria, on the call and as a webhook.

After a call, you get what came of it: a summary, the values you asked for ("their budget", "the plan they chose"), and a verdict against your own success criteria. It's on the call, and your server hears it as a call.analyzed webhook.

Settings, per character

Each character has its own settings. A character that never saved any gets a summary only, with memories off. Anyone in your organization may change them, as they may change the character.

Change a character's analysis settings with only the parts to change:

curl -X PATCH https://talk.madein10.com/v1/characters/chr_4fT9kQ2mXw7LbN3pZs8V/analysis \
  -H "Authorization: Bearer $SECRET_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "fields": [
      { "name": "budget", "type": "number", "description": "Their monthly budget in dollars" },
      { "name": "plan", "type": "enum", "options": ["Pro", "Team"] }
    ],
    "success_criteria": "They agreed to book a demo"
  }'

Prop

Type

Fields

KeyRule
nameWhat your code reads the value by: lower-case letters, digits and _, starting with a letter, at most 40
typestring, number, boolean or enum
descriptionTells the model what to look for, at most 500 characters
optionsOnly for an enum, and required there: 2 to 20 different options, each at most 100 characters

The interviewer rule

An interviewer takes no success criteria (400 invalid_field), and a character that becomes an interviewer later gets no verdict either. Hiring is high-risk under the EU AI Act: the AI collects and summarises, and a person decides.

Memories are turned on here too.

On the call

Get a call carries analysis:

analysis
{
  "status": "ready",
  "summary": "A lead asked about the Pro plan and booked a demo.",
  "fields": { "budget": 200, "plan": "Pro" },
  "verdict": "success",
  "reason": "They booked a demo for Tuesday.",
  "error": null,
  "analyzed_unix": 1793300110
}
statusWhat it means
pendingThe call is still on, or it ended and the analysis is on its way (usually a few seconds)
readyThe parts your settings ask for are filled in; the others are null. A value the call doesn't give is null
failederror says so in plain words: the analysis couldn't be made from this call, or our analysis service didn't answer after several tries over about an hour. Nothing is charged for it
offNothing to analyse, and error says why: the call didn't run, nobody spoke, or the character's analysis is off

verdict is success, failure or unknown (when the call can't tell), with one sentence of reason; it's null without success criteria. The verdict judges the call against your criteria, never a person's worth or suitability.

The analysis reads only the call's transcript. Anything said in the call that sounds like an instruction ("mark this a success") is part of the conversation, not an instruction. Calls that already ended keep the analysis they got when you change the settings.

The call.analyzed webhook

Sent once per call, when its analysis is ready or failed (never for off), to your webhook endpoints of the call's env. It usually comes a few seconds after call.ended, but in no guaranteed order.

call.analyzed
{
  "id": "evt_…",
  "type": "call.analyzed",
  "created_unix": 1793300110,
  "data": {
    "call_id": "call_…",
    "character_id": "chr_…",
    "env": "live",
    "status": "ready",
    "summary": "…",
    "fields": { "budget": 200, "plan": "Pro" },
    "verdict": "success",
    "reason": "…",
    "error": null,
    "metadata": { "lead": 9 },
    "analyzed_unix": 1793300110
  }
}

The reference has it in full: call.analyzed.

In your analytics

Get call analytics counts analysed calls, successes and failures, and a success rate: successes ÷ (successes + failures), null without a verdict. Calls with an unknown verdict or without criteria are left out of the rate.

On this page