AI QA in the call center: 100% quality assurance versus 1-3% sampling

What is AI QA in a call center?

AI QA in a call center is an automated quality assurance system that transcribes, scores, and surfaces every call against a defined rubric in real time, replacing the industry-standard 1 to 3 percent manual sample. A 100 percent AI QA stack catches script drift, compliance leaks, and conversion-killing patterns within minutes of the call ending. Coaching loops run on real data, not anecdotes.

The traditional QA model was built around scarcity. A human analyst could realistically score 30 to 60 calls per shift against a 20-question rubric. With a typical agent producing 60 to 120 calls per shift, the ratio gave you a sample of 1 to 3 percent of calls, scored 3 to 14 days after the conversation. The pattern was inherited from the early 2000s contact-center QA discipline, became the BPO industry default, and is still treated by most procurement teams as the standard against which vendor quality should be measured. The discipline has not aged well. Automatic speech recognition, large-language-model scoring, and cheap compute now make scoring every call cost less than the QA-analyst headcount the old model required. The right comparison in 2026 is not "your sample rate" but "do you score every call, and how fast does the breach reach the supervisor."

This pillar walks the full picture: why the 1 to 3 percent standard breaks in 2026, what 100 percent AI QA actually means technically, the four-piece tech stack, per-vertical rubric design, a seven-question vendor checklist, the cost math, the compounding coaching effect, and an honest FAQ. The activity-level CFG operational pattern is in the nearshore fronter perimeter and the regulatory companion piece sits at fronter vs licensed agent.

Why 1-3% manual sample QA fails in 2026

The 1 to 3 percent sample model has three structural problems that compounded as call volume grew and regulatory pressure tightened.

Statistical blindness. A 1 percent sample on a 120-call shift returns roughly one scored call per agent per day. On a 60-call shift it returns roughly half a call per agent. The sampling rate is below the noise floor for any agent-level performance signal. A QA score derived from one call is more about which call the analyst happened to pull than about the agent's actual performance distribution.

Latency. A breach captured in a manual sample surfaces 3 to 14 days after the call. By the time the coaching session lands, the agent has produced 200 to 1,000 more calls, several of which likely contain the same breach. The damage compounds and the coaching fires too late to prevent it.

Compliance exposure. A 1 percent sample on a Medicare AHIP-script program with 100 percent compliance requirement is mathematically incoherent. The buyer is paying for a sample that, by construction, cannot tell them whether the other 99 percent of calls hit the same standard. In regulated verticals (CMS MCMG, FDCPA Section 805, TCPA, HIPAA, NAIC), the gap between sample QA and actual compliance posture is large enough that procurement teams have started writing 100 percent QA into vendor SLAs.

What "100% AI QA on every call" actually means technically

100 percent AI QA means every call recording is transcribed, scored against the rubric, and surfaced to the supervisor dashboard within minutes of disconnect. The phrase is sometimes used loosely to mean "we score more than the industry average" or "we run AI-assisted sampling." Those are not the same thing. The honest definition has four requirements.

1. Every call, no sampling. Coverage is 100 percent of completed calls, not 100 percent of "qualifying" calls or 100 percent of calls above a duration threshold. If a call routes to voicemail or disconnects under 30 seconds, the AI QA system still logs the disconnect reason for separate analysis.
2. Automated transcription. Automatic speech recognition (Whisper, Deepgram, Google Speech-to-Text class) generates a transcript with speaker diarization (who said what) and timestamps. The transcript is the substrate the scoring layer reads.
3. Rubric scoring by model. A large language model scores the transcript against a structured rubric (compliance checks, scripted-language adherence, tone, intent capture, consent capture, deflection of out-of-scope questions). The output is a structured score plus the specific transcript span that triggered each flag.
4. Supervisor dashboard with sub-5-minute latency. The score, the breach flags, and the relevant transcript span land on the supervisor dashboard within 5 minutes of disconnect, typically in 90 seconds to 3 minutes. The supervisor can see the breach in time to coach the agent before the next call.

If any of the four pieces is missing, the system is not 100 percent AI QA. It is sample QA with a faster tooling layer, or it is bulk transcription without a scoring layer, or it is end-of-week reporting that does not reach the floor in time to change behavior.

The AI QA tech stack

The CFG AI QA stack runs on four operational layers. The layers are not unique to CFG. Any properly built 100 percent AI QA program runs a version of the same four. The stack is what the marketing claim "100% AI QA on every call" actually points at.

The pipeline runs on n8n workflow orchestration, with Supabase holding application data and the QA portal serving live results to supervisors. The stack is operational today and runs on every voice program CFG operates, regardless of vertical. For the operational mechanics on the broader CFG ops stack, the companion piece is tech-enabled BPO outsourcing 2026.

Per-vertical AI QA rubrics

The base rubric (pronunciation, fluency, grammar, communication, intent capture, consent capture) applies to every voice program. Regulated verticals add vertical-specific overlays. The overlays are what convert generic call quality scoring into compliance-grade quality scoring.

Medicare AEP (AHIP-aware)

Overlays: CMS MCMG scripted-language adherence, AHIP-restricted statements (no plan-specific or premium-specific language from a fronter), warm-transfer setup script integrity, consent capture under the September 2024 FCC declaratory ruling, deflection of out-of-scope plan questions. A breach on any overlay item triggers an immediate supervisor alert. For the regulatory background see CFG Medicare services.

Debt collection (FDCPA, Reg F)

Overlays: FDCPA Section 805 first-party contact verification, validation-notice scripting accuracy, Reg F frequency cap honoring (statements about prior-contact count), scripted limited-content message rules, deflection of settlement statements from non-licensed fronters. See CFG debt collection services.

TCPA outbound (any vertical)

Overlays: caller identification, prior-express-written-consent capture or verification, do-not-call honor, time-of-day compliance, offshore-origination disclosure where the call originates outside the United States. See TCPA compliance call center outsourcing and the FCC CG Docket 02-278 compliance checklist for the rulebook.

HIPAA healthcare

Overlays: PHI handling within business-associate-agreement scope, no clinical-advice statements from a non-licensed fronter, scheduling and demographic capture only, encrypted channel adherence. See CFG customer support outsourcing for the operational pattern.

Insurance (NAIC)

Overlays: state Department of Insurance scripted-language adherence, no binding statements from a non-licensed fronter, no rate quoting or product recommendation, FNOL intake scope only, NAIC Producer Licensing Model Act scope deflection. See CFG insurance services.

How to evaluate an AI QA vendor: 7-question checklist

Procurement teams should walk these seven questions with every vendor claiming AI QA. The honest answers will separate genuine 100 percent AI QA from sample QA with a marketing rewrite.

1. What share of calls are scored? Answer should be 100 percent of completed calls. If the answer is "we score more than the industry average" or "we use AI to pick a smarter sample," the vendor is running sample QA.
2. What is the median latency from call disconnect to supervisor dashboard? Answer should be under 5 minutes (90 seconds to 3 minutes is typical on a modern stack). If the answer is "end of day" or "next business day," the coaching loop does not run on real-time data.
3. What ASR engine and what scoring model? Answer should name the ASR vendor and the model class. If the vendor cannot name either, the stack is likely a transcription-only layer without true scoring.
4. How is the rubric configured per program? Answer should describe a base rubric plus vertical overlays editable per client. If the rubric is "the same one we run for every client," compliance overlays are not being applied.
5. What happens when the AI scoring flags a breach? Answer should describe an alert to the supervisor dashboard and a documented coaching workflow. If the answer is "we email a weekly report," the coaching loop is not closing.
6. Can the buyer access the raw scored data? Answer should be yes, via API or dashboard export. If the buyer has no access to the raw scoring, the vendor is gatekeeping the QA layer.
7. How is the AI QA priced? Answer should be either bundled into the seat rate or transparent per-minute pricing. If the AI QA is a hidden add-on or a vague "premium service" tier, the economics will not hold at scale.

Cost of AI QA vs traditional sample QA

The cost math is what convinced most operators to flip. A traditional QA team scoring 1 to 3 percent of calls runs roughly 4 to 8 percent of agent labor cost in QA-analyst headcount, supervisor time, and tooling. The labor scales linearly with call volume (more calls means more analysts), with no economic incentive to scale coverage higher.

AI QA scales with call seconds, not analyst hours. The cost line is ASR cost plus LLM inference cost plus dashboard hosting. On 2026 cloud pricing, the loaded cost runs roughly 1 to 3 percent of agent labor for 100 percent coverage. The buyer pays less and gets more. At CFG the AI QA layer is included in the seat rate. There is no separate billing line. For the underlying seat economics see how pricing works.

The compounding effect: agents improve faster on real-data coaching

The economic case for AI QA is the cost line. The strategic case is the coaching velocity. An agent coached on a real breach from yesterday's call carries the correction into today's calls. An agent coached on a hand-picked sample from two weeks ago has already produced hundreds of calls with the same uncorrected pattern. The first agent improves on a one-day learning loop. The second improves on a two-week learning loop.

Compounded across a 90-day onboarding window, the gap is large. A new agent entering the floor with daily real-data coaching reaches full-floor scorecard parity in roughly 21 to 28 days. The same agent entering the floor with weekly sample coaching takes 60 to 90 days. The faster ramp shows up in conversion rate, compliance posture, and attrition. For the broader CFG operational pattern see customer support outsourcing.

Frequently Asked Questions