Contact-centre QA
Transcribe each finished call, bleep card numbers and IDs from the recording and mask them in the transcript, score the call against your QA rubric with an AI agent, and send low scores to a supervisor.
The problem
Supervisors can only listen to a few calls per agent, and recordings with card numbers and dates of birth sit in storage for years. Sending raw transcripts to a model for scoring would spread that data further.
How UserTasks solves it
The call is transcribed with speakers, card numbers and IDs are bleeped at the times they were spoken and masked in the transcript, and only the masked transcript reaches the QA agent. A decision table sends low scores, negative customers and unverified account changes to a supervisor.
Architecture
In practice
Post the call id and a signed recording URL when the call ends; the rest runs without a UI until a supervisor is needed.
curl -X POST "$USERTASKS_URL/api/v1/workflows/start" \
-H "Authorization: Bearer $USERTASKS_API_KEY" \
-H "Content-Type: application/json" \
-d '{"workflow_id":"<id>","variables":{
"call_id":"CALL-0042",
"recording_url":"https://telephony.example.com/rec/0042.mp3",
"agent_name":"Casey Example","queue":"Billing"}}'
The flow
Your platform posts the call id, recording URL, agent and queue.
Speakers are separated; card numbers and IDs are bleeped and masked.
The QA agent scores the masked transcript; thresholds decide on a review.
Supervisors coach on flagged calls; redacted copies are archived with retention.
Ready-made package
Each finished call is transcribed, card numbers and IDs are bleeped and masked, an AI agent scores it on your QA rubric, and low scores go to a supervisor.
Headless / API · needs Media AI provider, AI model, Assignees · about 20 minutes to set up · signed, free to install
No tool does that on its own. Bleeping card numbers from recordings and masking transcripts reduces where card data is stored, but your compliance depends on your whole process, including the original recordings in your telephony platform.
Detection depends on the transcript. Supervisors listen to flagged calls, and you can tighten the info types or add custom patterns and terms.
Not by this package: it reads the recording from your platform’s URL and archives only the redacted copy and masked transcript.
Meeting minutes
A recording becomes a transcript with speakers, minutes the organiser checks, and a task for every action owner.
Read the use casePublic records and evidence redaction
FOIA and public records requests: deadlines, AI-assisted redaction of video, photos, audio and documents, an exemptions log and legal approval before release.
Read the use caseAI agent orchestration
Agents as governed workflow steps with granted tools, structured outputs, human review and full run history.
Read the use caseBuilding blocks:Complex workflows Integrations and AI Forms and attachments Document actions Images, audio and video Package gallery Deploy anywhere
Open source under Apache-2.0. Use the cloud, or self-host it next to your own systems.