Sane contact center quality assurance does not listen to every recording. It uses a weighted sample, one rubric for humans and bots, and escalation when a bad score meets risk. Hearing every thread is a capacity fantasy, not a control.
Sample unevenly on purpose
Take more from complaints, refunds, fraud, and reopened threads. Take less from FAQs whose deflection is already honest. Pure random sampling drowns risk in “opening hours.” On WhatsApp, which Meta said at WABS on 11 August 2026 is used by 71 percent of people in Indonesia every day, volume kills even QA.
Write the weekly sample size and who scores. A supervisor scoring only their own team without calibration will inflate marks. Calibrating two scorers on ten shared cases once a month is enough for a small team.
A rubric a bot cannot game
Correct, complete, civil, and finished. “Finished” must match honest first contact resolution: not reopened inside your window. A polite, wrong bot still fails. Kompas on 11 August 2026 wrote that 79 percent of respondents found Meta AI responses useful. A survey of usefulness is not your internal QA score.
Add a handoff line. If a fraud or UU PDP trigger is ignored, zero the case even if the language is pretty. ANTARA cited 83 percent saying AI agents help them get what they need faster. Fast and correct can diverge. The rubric must split them.
Chat, voice, and transcripts
For chat, QA reads the thread, not a snippet. For voice, a transcript plus audio on risky cases. Do not judge emotion from text alone when tone is the complaint. Store the score reason on the ticket, not in a private sheet.
OJK’s IASC, as reported by Kompas on 8 September 2026, blocked Rp 771.2 billion tied to fraud. “My transfer vanished” threads belong in the QA sample, not in the random pile.
Since 1 July 2025 WhatsApp templates are billed per message. QA may also flag a mis-categorised template. That is cost and compliance, not only tone.
Minister of Manpower Regulation 7/2026 took effect 30 April 2026. Vendor QA does not replace QA on core agents who should sit inside the company. A quality metric is not a labour-law loophole.
Frequently asked questions
Must QA hear every recording?
No. A weighted sample on risk and reopens is more honest and fits a small team.
How many cases per agent per week?
Enough to see a pattern, not enough for theatre. Keep a calibration round. There is no sacred number here.
Do bots use the same rubric?
Yes for correct, finished, and handoff. Tone can be adapted. Do not write a looser “bot rubric.”
What about UU PDP privacy?
Limit QA access, limit retention, and log who opened a recording. QA is not a reason to keep every chat forever.
What happens to a bad score?
Coaching, a knowledge-base fix, or turning an intent off. A score with no action is decoration.
Does CSAT replace QA?
No. CSAT is feeling. QA is whether procedure and facts were right.
Book a demo and see your own contact flows running on ConverseAIght
ConverseAIght
