How to Build a QA Scorecard for Debt Collection Calls

A good QA scorecard does two jobs at once: it protects the company from compliance risk, and it coaches agents toward better outcomes. Most scorecards do only the first. Here's a structure that does both.

Start with hard compliance gates

Certain items should be pass/fail, not partial credit — because a single miss is a violation, not a coaching note. These typically include: the mini-Miranda disclosure delivered correctly, no false or misleading statement, hard-stop statements (attorney, bankruptcy, cease-and-desist, minor, identity theft) handled with an immediate stop-and-escalate, and no disclosure to an unauthorized third party.

A call that fails a compliance gate should score at or near zero regardless of how well the rest of the call went — that asymmetry is intentional, and agents should know it.

Score call-flow adherence

• Right-party verification completed before any account detail
• Purpose statement clear and calm
• Open-ended discovery on reason for delinquency (not interrogation)
• Resolution offered in the correct order: full payment, then structured plan, then settlement
• Terms read back with a confirmation reference

Score resolution quality, not just compliance

This is where most scorecards fall short. Add: was hardship met with genuine empathy and a sustainable option? Did the agent avoid coercive language? Did the agent stay professional through refusal or a difficult tone? Was the close professional regardless of outcome?

Weight it, don't average it

A common mistake is averaging every item into one number, which lets a strong 'resolution quality' score paper over a compliance miss. Instead: compliance gates are pass/fail and dominate the outcome; call-flow and resolution-quality items are weighted and summed separately, feeding coaching rather than the pass/fail gate.

Where AI QA changes the math

Manual QA can typically sample only about 2% of calls. AI-augmented QA — speech-to-text plus automated scoring — can apply this same scorecard to every call, flagging the ones that need human review instead of hoping a random sample catches them. The scorecard structure above works identically whether a human or a model applies it; what changes is coverage.