The most useful output of an AI calling campaign isn't the recording or even the summary. It is a one-word outcome for every call that you can count. A school-book publisher we work with scores each teacher call into outcomes like APPROVED, NOT_TEACHER, CALLBACK, WRONG_PERSON, DROPPED_EARLY or HOSTILE, and each becomes a dashboard column. The team sees approvals by region and class without listening to a single call. Here is how to set up automatic call scoring and keep it accurate.
Set up call scoring on your agent: Start free on Edesy — Rs 50 free credit, no demo needed.
Summaries vs scores
| Summary | One-word score | |
|---|---|---|
| Good for | Reading one call | Counting thousands |
| Consistent across calls | No | Yes |
| Filterable | No | Yes |
| Drives actions | Manually | Automatically (callbacks, CRM stages) |
You want both: summaries for context, scores for decisions.
How it works
- Define fields with a small, fixed set of allowed values and a clear description for each.
- After every call, the platform reads the conversation and fills each field.
- Fields become columns in the call log, reports and exports, and can be pushed to your CRM or a webhook.
On Edesy these are the agent's data fields, extracted after each call.
Example: the publisher's teacher calls
| Field | Values | Description |
|---|---|---|
| outcome | APPROVED, NOT_TEACHER, CALLBACK, WRONG_PERSON, DROPPED_EARLY, HOSTILE | What happened on the call |
| class_band | PRIMARY, MIDDLE, SECONDARY, SENIOR | Class the teacher teaches |
| subject | from a fixed list | Subject taught |
| whatsapp_ok | YES, NO | Agreed to continue on WhatsApp |
Read the full use case in how a publisher reaches teachers with AI calls.
Example: sales qualification
| Field | Values |
|---|---|
| lead_temperature | HOT, WARM, COLD |
| intent | BUYER, JOB_SEEKER, VENDOR, OTHER |
| timeline | THIS_MONTH, 90_DAYS, LATER, UNSURE |
| next_step | TRANSFERRED, CALLBACK_BOOKED, WHATSAPP_SENT, NONE |
See how a manufacturer qualifies leads with AI calls.
Writing good field descriptions
- Describe each value in one line. "CALLBACK: the person asked to be called at another time."
- Say what to choose when unclear. "If the call ended before the question was asked, choose DROPPED_EARLY."
- Keep values mutually exclusive. A call should fit exactly one outcome.
- Separate dimensions. Outcome, interest and language are different fields, not one long list.
Keeping scores accurate
- Spot-check in the first days: listen to 5-10 calls per outcome and compare.
- Tighten descriptions where the label and the call disagree.
- Recheck after script changes: a new question can shift how outcomes are labelled.
- Watch DROPPED_EARLY: a rise often means a problem with the opening line or silence on the call. See how to fix dead air.
Acting on scores automatically
- CALLBACK → create a callback task at the requested time.
- APPROVED / HOT → notify sales on WhatsApp, or transfer live during hours.
- WRONG_PERSON → capture the right contact and call them.
- HOSTILE / opt-out → add to the do-not-call list.
Scores can go to a webhook, Google Sheets, or HubSpot, Salesforce, Zoho CRM and GoHighLevel.
Try it
Create a free account, add an outcome field with five values to your agent, and run 20 test calls. For surveys, where every answer is a field, see AI phone surveys in Indian languages.