The problem
Manual review reaches only a small part of what happens on the phones, and two reviewers often score the same call differently. Coaching then rests on whichever calls happened to be picked.
The recordings already exist, frequently in several systems. What is missing is a consistent way to read them and a way to know how far to trust the result.
Who it is for
QA leads: Reviewing far more conversations with one consistent rubric.
Contact centre managers: Seeing what changed and who needs attention.
Team leads: Coaching from specific calls, with the reason behind each score.
Teams running AI agents: Holding AI and human agents to the same measures.
What it handles
Transcription
Recordings from uploads, S3-compatible storage, FTP or SFTP, Google Drive, links or the API, and calls made in Kilevel.
Scoring
Twelve ready-made metric packs, or your own structured outputs. Judgement fields can carry a written reason citing what was said.
Measured metrics
Talk share, dead air, interruptions, speaking rate and longest monologue, taken from the timing of the recording.
Scorecards
Weights, targets and auto-fail criteria, reported as their own rate so one serious breach is not averaged away.
Briefings and alerts
A daily summary of what changed, and rules checked every night.
How it works
Bring the conversations in
Calls and chats in Kilevel are already there. Point a transcription job at your archive for the rest.
Attach analysis
Use a metric pack or your own fields. New transcription jobs start with the core pack.
Identify the agents
Imported recordings need the agent's name or ID in the file name, a spreadsheet or the API call.
Adjust the scorecard
A standard one is built from the metrics your account produces. Change the weights to match what your programme rewards.
Calibrate
Score a sample by hand and compare it with the model before relying on the numbers.
What you need
- Conversations: calls in Kilevel, or recordings in a place Kilevel can read
- A way to tell which agent handled each recording
- Your QA criteria, or a preset to start from
- A few reviewers for calibration
- A decision on how people should appear, if a works council is involved
What it connects to
Recording sources: S3-compatible storage, FTP and SFTP, Google Drive, public or pre-signed links, and the API.
Your systems: Results come back through the API and webhooks.
Any contact centre: Imported recordings count alongside conversations that ran through Kilevel.
What to expect
- A consistent reading of far more conversations than manual review can reach.
- Coaching based on specific calls, with the reason for each score.
- An early signal when a measure moves, rather than at the end of the quarter.
- A figure for how closely the model agrees with your reviewers.
Limits
- Scores are a model's judgement. Calibrate them against human reviewers before you act on them.
- Averages come from the conversations that were actually analysed. Check coverage before treating a figure as complete.
- Transcription quality depends on the audio. Fixed language settings are the most accurate when every recording is in one language.
- In a workplace with a works council, per-person measurement is subject to co-determination. The privacy modes support an agreement but do not replace one.
- A pseudonym is not anonymity: over a small team it can still identify someone.
How to evaluate it
Import a real batch: Use recordings you already know, including some difficult ones.
Score twenty by hand: Around twenty completed reviews is enough for the agreement figure to mean something.
Read the disagreements: A criterion people disagree about is usually worded ambiguously. Rewrite it.
Check coverage: Make sure you know how much of your volume the numbers describe.
Questions
Does it work on calls not made in Kilevel?
Yes. A transcription job takes audio from your archive and produces records of the same shape as a Kilevel call, so they count together.
Can we use our own QA criteria?
Yes. Scorecards use the metrics you choose, with your weights and targets. An account starts with a standard scorecard that you can edit.
How accurate are the scores?
Do not take that on trust. Review & calibration measures how often the model and a human reviewer land within ten points of each other on your own conversations.
Does it score AI agents and human agents?
Both. Calls, chats and imported recordings are counted per agent, and you can filter by source and by kind of agent.
What about works councils in Germany?
Besides showing names, you can use pseudonyms or aggregate figures only, set minimum group sizes and log every reveal. Whether that fits your agreement is for you and your works council to decide.