Skip to main content

Overview

Call analysis listens to your recordings for you. You write an analyzer: plain instructions that say what to look for (“flag any wrong refund information”). 60db transcribes each call, checks it against the instructions, and returns the problems it found with the exact quote, the time in the audio and the correct information. Each problem has a category (wrong_information, missed_step, compliance, customer_unresolved, behaviour, other) and a severity (high, medium, low). Each call gets a verdict: problems, needs_review or ok.

Where the recordings come from

  • Files you upload (mp3, wav, m4a, mp4, ogg, opus, flac, webm, up to 100 MiB each)
  • Links that 60db downloads for you
  • Your 60db AI agent calls and Dialer calls, picked one by one or watched so new calls are analyzed automatically
To use Call analysis with an API key, the key needs the Call analysis permission. New keys have it; for older keys switch it on in Developers → API keys. See the API reference.

Fix your AI agent

For a call made by a 60db AI agent, 60db can suggest better agent instructions from the problems it found. You review the suggestion, apply it, and undo it if you do not like it.

Self healing

Turn on Self healing on an analyzer that watches AI agents and it stops only reporting problems: it turns problems that repeat into a change to the agent’s instructions. Dialer calls have no agent instructions to change, so dialer healing is not part of this release.

When a draft is made

  • The same category and severity (medium or high) appears in at least 3 different calls among the agent’s last 20 analyzed calls. One-off problems and low severity problems never trigger a change.
  • Calls that were already the evidence for an earlier change are not counted again.
  • Drafts are free. They do not use credits.

Guardrails

  • One open change per agent. While a change is waiting for review, or is being measured, no other healing change is drafted for that agent, even from another analyzer.
  • Cool-down and daily limit. No new change within 24 hours of the last change to the agent, and at most one draft per agent per day.
  • Permission is re-checked. Automatic mode runs as the person who turned it on. If that person can no longer edit agents, the analyzer drops back to Suggest fixes and the change waits for review. If the permission check cannot be made, nothing is applied.
  • Nothing is overwritten. If someone edits the agent’s instructions in the meantime, the change is marked Superseded and neither applied nor undone.
  • Measured and rolled back. The change is measured on the next 10 calls that start after it. The share of calls with the targeted problem is compared with the 20-call window that triggered it: improved, no change, or worse. A worse result (or a new high severity problem) restores the previous instructions automatically.
  • Undo any time. Every applied change can be undone with one click while the instructions are unchanged.
  • A failed draft, a refused validation or a failed update never stops call analysis and never leaves the agent half-changed.

Who can use it

Viewing the Healing list needs view access to the analyzer. Approving, dismissing and undoing need manage access to the analyzer and permission to edit agents. Choosing Apply automatically also needs permission to edit agents. Healing endpoints are for signed-in users and are not available to API keys. See Get Sources and Set Sources for the setting, and List Healing Changes for the changes.

Use it from the API

Everything above works with a normal API key. Start here:

Analyze your first calls

A step-by-step curl walkthrough, including file upload

API reference

Endpoints, permissions, limits and errors

MCP tools

Let an AI assistant run analyzers for you
Analysis uses speech-to-text and model credits from your workspace. You confirm the recording count before every run.