Industries
CallScribe for contact centres
AHT, FCR, CSAT, and QA scoring — pulled directly from Arabic call audio.
Last updated: April 2026
A GCC contact centre with 200 agents handles roughly 80,000 to 120,000 Arabic calls per month. Manually QA-sampling 5% of those calls — the industry norm — yields a few thousand audited interactions, and that sample is biased toward calls QA managers happen to pick. CallScribe transcribes 100% of traffic, runs sentiment and disposition models on every call, and turns AHT, FCR, CSAT, and QA scoring into measurements rather than estimates.
Average Handle Time, accurately measured
AHT in most GCC contact centres is reported from PBX data — call connect to call disconnect — and that conflates productive talk-time with hold-music, transfer-routing, and dead air after the customer hangs up. CallScribe distinguishes pure speech time (using VAD on the transcript), hold time (silence or hold-music detection), and post-call wrap (often present in single-channel recordings). Real AHT, broken down into its components, frequently shifts what looks like an agent-performance issue into a routing or process issue.
First-Call Resolution detection from transcripts
FCR is hard to measure because it requires linking call N+1 to call N — same customer, same issue, within a window. CallScribe surfaces FCR signals directly from transcript content: explicit re-contact phrases ("بكلمكم تاني" — I'll call again), unresolved-state markers ("ما زال", "لسة في المشكلة"), and post-call dispositions linked across calls by customer ID. FCR scores produced this way are typically 8-15 percentage points lower than the optimistic dispatch-system metric, which is exactly why operations leaders ask for it.
CSAT scoring from sentiment, not just surveys
GCC call-centre survey response rates are notoriously low — often 5-10% on post-call IVR satisfaction surveys. The 90% of calls without a survey are an information gap. CallScribe scores conversation sentiment trajectory (start, mid, end) and end-state customer affect across every call. Combined with the survey responses you do receive, sentiment-derived CSAT is calibrated against the survey ground truth and used as a wide-coverage proxy on unsurveyed calls.
QA scorecard automation
A typical QA scorecard has 15-30 questions. "Did the agent greet by name?" "Did the agent verify the customer?" "Was the call closed with a courtesy phrase?" Most are direct transcript searches. CallScribe runs a pre-built QA scorecard plus customer-defined rules against every transcript and produces per-call scores so QA leads spend their time auditing edge cases rather than reading every transcript end-to-end. Random sample for human QA is then drawn from disagreement zones (low-confidence model scores) rather than uniformly.
Agent coaching evidence packs
Coaching sessions land harder when agents see exact transcript snippets, not management abstractions. CallScribe assembles a per-agent monthly evidence pack: 5 strongest calls (high CSAT, low AHT, full QA score), 5 weakest calls (low sentiment, missed scorecard items), and aggregate trend data for talk-listen ratio, filler frequency, and dead-air. Evidence-based coaching is what coaching research actually finds effective.
Regulator interest in call recording
GCC regulators increasingly require call recording and retention. The UAE Personal Data Protection Law (Federal Decree-Law No. 45 of 2021) treats voice recordings as personal data; UAE TDRA telecom regulations expect operator call-recording for dispute resolution. Saudi Arabia's Communications, Space and Technology Commission (CST) and Saudi Central Bank (SAMA) regulations on financial-services calls require retention windows ranging from one to seven years. CallScribe stores transcripts on EU infrastructure with audit logs and supports retention rules per project.
At a glance
- ✓AHT decomposed into talk, hold, wrap
- ✓FCR from transcript content, not just dispatch metadata
- ✓Sentiment-derived CSAT for unsurveyed calls
- ✓Auto-scored QA scorecards on 100% of traffic
- ✓Per-agent coaching evidence packs
FAQs
How does CallScribe integrate with our existing PBX or ACD?▾
CallScribe accepts call recordings via SFTP drop, S3 push, or REST API. We do not need to be in the call path. Most customers feed the daily recording dump from Avaya, Genesys, Cisco, or Mitel into a CallScribe project; transcripts and analytics appear within minutes per call.
Can we run CallScribe alongside our existing speech-analytics tool?▾
Yes — many customers run CallScribe on Arabic traffic specifically because their existing English-first speech analytics tool fails on Arabic, and keep the existing tool for English. Transcripts export as JSON for downstream platform integration.
How is agent privacy handled on coaching transcripts?▾
Personally identifiable information (PII) — credit card numbers, ID numbers, phone numbers — is masked at transcription time using a regex-and-NER pipeline. Agent identity is preserved for coaching but can be anonymised for aggregate reporting.
What sample size do you need to calibrate sentiment-CSAT?▾
A few thousand surveyed calls is enough to fit a calibration curve mapping sentiment scores to CSAT. We have a calibration helper that runs against your existing IVR-survey data, then applies the curve to all calls.
Does CallScribe handle multi-language contact centres?▾
Yes — Arabic dialects (Khaleeji, Levantine, Egyptian, MSA) are auto-detected per call. Mixed-language traffic with English, Hindi, or Urdu is supported. The dialect-ID probe runs in the first 5 seconds and routes to the appropriate model.
What is the typical setup time for a 200-seat operation?▾
Typically two to three weeks. Week 1 is sample audio onboarding and QA scorecard configuration. Week 2 is calibration of sentiment to your existing CSAT data and definition of agent groups. Week 3 is rollout. Most of the work is on the customer side defining what counts as "good" rather than CallScribe configuration.
5 min/mo free · No credit card · 8-12% WER on Khaleeji