Manual reviews cover 1–2% of calls. What are you missing in the other 98%?

In most call centers, quality control looks the same: a coach or quality manager listens to a few random recordings per agent per month, fills in a scorecard and discusses the results in a feedback session. The problem isn’t the quality team’s competence — it’s arithmetic. Below we do that arithmetic in full: coverage, the probability of catching a critical error, and cost in euros — then put manual review and AI analysis of 100% of calls side by side, criterion by criterion.

Where does the 1–2% come from?

Let’s count with a simple example we’ll reuse throughout this article. A team of 30 agents makes about 60 calls a day each — that’s 1,800 calls daily and, over 20 working days, roughly 36,000 a month. An experienced quality coach can reliably review and score a dozen or so calls a day: the call itself takes several minutes, then comes the scorecard and notes.

Even doing nothing else, they’ll evaluate 300–400 calls a month. That’s the famous 1–2% — assuming no training sessions, calibrations, meetings or holidays. Note that this isn’t a policy anyone chose; it’s simply the ceiling of what one person can physically listen to.

What a random sample can’t show

A 1–2% sample might suffice if errors were spread evenly. But the most dangerous events — skipped identity verification, misleading a customer about contract terms, a missing mandatory clause — are rare. Run the numbers: if a critical error occurs in 0.5% of calls, that’s about 180 affected calls out of our 36,000. A reviewer sampling 350 calls a month will, on average, stumble on fewer than two of them — and any single bad call has a roughly 99% chance of never being heard by anyone.

In practice this means the organization learns about serious errors not from monitoring, but from a customer complaint or a regulator’s letter — weeks after the fact.

Sample-based monitoring doesn’t answer the question “how is my team performing”. It answers “how did a few randomly picked calls go”.

Randomness hurts agents too

From the agent’s perspective, sample-based evaluation can be simply unfair. One weaker call drawn for review can weigh down a monthly score, while dozens of good ones go unnoticed. It’s hard to build a feedback culture on that — the result depends on the draw, not the work. Agents know it, and it undermines trust in the whole quality program.

What manual review actually costs — a worked example in euros

Coverage is one side of the equation; cost is the other. Three explicit assumptions — swap in your own numbers, the arithmetic stays the same:

• Volume: 30 agents × 60 calls a day × 20 working days = 36,000 calls a month.
• Throughput: one quality specialist reliably scores about 350 calls a month (the 300–400 range from above).
• Cost: a fully loaded quality specialist — salary, employer contributions, tooling — costs €3,000 a month. In Western Europe this figure is often higher, in Central Europe lower; adjust to your market.

Now the math. One scored call costs €3,000 ÷ 350 ≈ €8.60. One full-time reviewer covers 350 ÷ 36,000 ≈ 1% of traffic. Want the “industry standard” 2%? That’s two specialists — €72,000 a year to hear 700 of 36,000 monthly calls. And full manual coverage of 100%? You’d need 36,000 ÷ 350 ≈ 103 reviewers, about €309,000 a month — at the same €3,000 fully loaded cost, more than three times the payroll of the 30-agent team they’d be checking. Nobody does that, of course. Which is exactly why 98% of calls go unheard.

Manual review vs AI analysis of 100% of calls

Here’s the same comparison laid out criterion by criterion, using the example above:

One honest caveat on the right-hand column: CallSea doesn’t publish a price list — pricing is discussed at a demo. But the structural difference holds regardless of the number: software doesn’t need one additional reviewer for every additional 350 calls.

What analyzing 100% of calls changes

Automatic AI call analysis flips the sampling logic. Every call is transcribed and scored against the same criteria — checklists, point scales, critical-error rules, plus an overall 0–100 AI score with a written rationale — within minutes of hanging up.

• Critical errors surface the same day, with the specific call and a transcript quote you can bring to the feedback session — see how critical-error detection works.
• Agent evaluation rests on complete data — per-agent trend dashboards across hundreds of calls instead of an impression from three sampled recordings.
• The quality team stops listening and starts managing quality: calibrating criteria, working with results, coaching where the data points.

Where to start?

The best first step costs nothing: write down your current scorecard and the rules you evaluate calls by today — they become the criteria for AI scoring. The second step is technical and smaller than it sounds: getting recordings into the system, for example via automated SFTP ingestion. How to design criteria so automatic scoring is unambiguous is a topic for a separate guide, and if you read Polish, our post on designing scoring criteria covers it in detail.