AI Call Centre Quality Assurance: Scoring What You Cannot Sample

Research date: 23 September 2026. Last updated: 23 September 2026.

Traditional call centre QA reviews four to eight calls per agent per month. Every AI QA vendor now promises to score 100% of conversations instead. Both numbers are worth interrogating: the first is smaller than it sounds, and the second is a capability claim rather than a plan entitlement — where vendors publish an AI-scoring allowance at all, it is metered, and one of them includes 500 scores a month on its mid tier.

The harder question sits underneath both. Once an AI agent is handling part of your volume, you are no longer sampling people in order to coach them. You are testing a system. That is a different discipline, and almost nothing written about call centre QA addresses it.

Brilo AI publishes this page and sells an AI voice agent, which is one of the things described here as needing QA. Brilo does not sell QA software. What Brilo does not publish about transcripts and retention is stated below alongside everything else.

How we compiled this

This is documented research, not a hands-on evaluation. Every price, allowance and vendor claim below was read from the vendor's own page or a named industry source on 23 September 2026 and carries its source. No QA platform was trialled and no scorecard was run, so nothing here describes how any of these perform in practice.

Where a figure is arithmetic rather than a published number, the working is shown so you can substitute your own volumes.

What manual QA actually covers

There is no industry consensus on a sampling rate, which is the first thing most guides get wrong by asserting one. Call Centre Helper's practitioner survey puts the common range at four to eight calls per agent per month, with five to six described as the balance point, one contributor arguing for at least 8% of total inbound calls, and 36% of respondents monitoring more than ten. Source: callcentrehelper.com, "Determining the Right Number of Calls Analysed", published 12 April 2022, modified 22 July 2026, read 23 September 2026.

Convert that to coverage and the picture changes. An agent handling 500 calls a month, scored on six, is reviewed on 1.2% of their work. That is the basis on which performance is rated, coaching is assigned and, in some organisations, pay is decided.

It is not indefensible. Sampling is a legitimate statistical tool and six calls will surface a consistently poor performer. What it cannot do is catch a rare, expensive failure — the compliance breach, the mishandled vulnerable caller, the one conversation that becomes a complaint. Those live in the other 98.8%, and no sampling rate a human team can sustain will find them reliably.

What the vendors promise

Every AI QA vendor has converged on the same answer, in near-identical language.

Level AI publishes: "Score 100 percent of calls, chats, emails, and bot conversations", alongside "100% coverage, 100% automation, 100% trusted", and states that traditional QA reviews only 1–3% of interactions. Source: thelevel.ai/quality-assurance-contact-center, read 23 September 2026.

Zendesk describes its quality assurance capability as "automatic human and AI agent scoring" and sells it inside a Workforce Engagement bundle. Source: zendesk.com/pricing, read 23 September 2026.

The claim is real in the sense that the technology can do it. Transcription plus a scoring model can in principle evaluate every conversation you have, and that is a genuine change in what is possible.

What they actually sell

Here the published detail diverges from the pitch, and it is worth checking before you plan around 100%.

Level AI publishes no pricing at all and states no quota, credit allowance or limit on AI scoring. Source: thelevel.ai, read 23 September 2026. You cannot tell from the public page what 100% coverage costs at your volume, or whether it is bounded.

Scorebuddy publishes no prices either, but it does publish allowances — and they are metered. Its Accelerate tier includes 500 AI auto-scores a month; its Elite tier includes 1,000, with additional AI credits sold as an add-on. Source: scorebuddycx.com/pricing, read 23 September 2026.

Run the arithmetic against a real workload. A contact centre handling 2,000 calls a month on Accelerate's 500 included scores is covered on 25% of its volume, not 100%. At 10,000 calls a month it is 5% — better than the 1.2% manual baseline, and nothing like the promise. Buying to 100% means buying credits, and the credit price is not published.

This is not an accusation of bad faith. "Score 100% of conversations" describes what the product can do; the allowance describes what your plan includes. But the two are quoted in different places, and only one of them appears in the headline.

Ask every QA vendor the same question in writing: how many AI scores are included on the quoted plan, what an additional score costs, and whether voice transcription is counted separately from scoring. Scorebuddy, for instance, lists AI voice transcription only on its top tier.

What QA tooling costs, where it is published

Sources: zendesk.com/pricing, thelevel.ai, scorebuddycx.com/pricing — all read 23 September 2026.

The pattern worth noticing: QA is an add-on to an add-on. On Zendesk, a seat that handles voice and is covered by quality assurance is a base licence plus $83 plus $50 — before any AI consumption. Two of the three vendors here publish no price at all, which signals a configured deal rather than a higher one, but it does mean you cannot budget this from public information.

What changes when the AI is the thing being scored

This is the part the category has not caught up with, and it is the reason a QA programme built for humans does not transfer.

Sampling logic inverts. You sample humans because agents vary and you need a representative view of each person. An AI agent does not vary that way: given the same intent and the same context it behaves the same way every time. Reviewing 30 random AI calls mostly tells you 30 times that the common path works. The failures cluster by intent and edge case, not by operator — the caller with an unusual accent, the request the configuration has no branch for, the interruption mid-sentence. Random sampling is close to the worst way to find those. Stratify by intent, and oversample the rare ones.

Coaching has no target. The output of human QA is a conversation with a person who then does better. There is nobody to coach here. The output of AI QA is a configuration change — a new branch, a reworded prompt, a changed escalation trigger — which means your QA process has to terminate in a change-management step or it produces nothing. A scorecard with no owner and no deployment path is theatre.

One fix is global, and so is one mistake. Retraining a human improves one agent. Changing the AI's handling of a billing question changes every billing call from that moment. That cuts both ways, and it is the strongest argument for keeping a human review loop: the blast radius of a bad change is your entire volume, immediately.

A regression can appear without anyone touching anything. Underlying models get updated. Behaviour that passed in March can fail in September with no change on your side. Human QA assumes the thing being measured only changes when someone trains it. That assumption does not hold, which makes periodic re-testing of scenarios that already passed a genuine requirement rather than a nicety.

Escalation quality matters more than resolution. The dangerous AI failure is not "could not help" — it is "handled it confidently and wrongly", or "should have escalated and did not". Score the handoffs specifically: did it escalate when it should, did the human receive context, did the caller have to repeat themselves.

A scorecard that fits an AI agent

Adapt rather than reuse. These carry over from a human scorecard: accuracy of information given, compliance with required disclosures, resolution of the caller's actual request, and tone.

These are new, and none of them appear on a standard QA form:

• Escalation correctness — both directions. Escalating everything is as much a failure as escalating nothing.

• Containment honesty — was the call genuinely resolved, or merely ended? A call that "contained" because the caller gave up is a failure recorded as a success, and it is the single most common way containment rates flatter a deployment.

• Context transfer — what the human agent received on handoff.

• Fabrication — did it state anything it had no basis for. This has no analogue in human QA and it is the failure mode with the most downside.

• Latency and interruption handling — dead air and talking over the caller are quality failures that no human scorecard measures because humans rarely do them in the same way.

When you do not need QA software

A page that says every buyer should buy is not worth reading.

Your volume is low. Under a few hundred calls a month you can read every transcript. A spreadsheet and an hour a week beats a platform, and you will learn more.

You have just deployed. In the first weeks the useful activity is reading transcripts yourself, not scoring them against a rubric you have not yet earned the right to write. Scorecards encode what you already know matters. Deploy first, find out what breaks, then write the scorecard.

You only run an AI agent, with no human team. Most QA platforms are priced per agent seat and built around coaching workflows, calibration sessions and agent dashboards. If you have no agents to coach, you are buying a coaching product to use 20% of it.

Nobody owns the output. QA that produces findings nobody acts on is cost with no return. This is the same disqualification that applies to the deployment itself, and it fails for the same reason.

Brilo's position, and what it does not give you

Brilo publishes a free tier at 10 minutes a month, Pro at $149 for 600 minutes with overage at $0.16, Growth at $499 for 2,500 minutes at $0.14, and a custom tier above 5,000 minutes. Source: brilo.ai/pricing, read 23 September 2026.

What it does not do: Brilo is a voice agent, not a QA platform. It has no scorecard builder, no calibration workflow and no agent-coaching module, and if you need to run QA across a human team it is not the product for that.

And a gap worth naming: Brilo's pricing page does not publish terms for transcripts, call recordings, retention periods or data export. Those are exactly the inputs a QA programme needs — you cannot score what you cannot read, and you cannot demonstrate compliance against records you cannot retain. If you intend to QA a Brilo deployment, get the retention period and the export route in writing before you sign, the same as you should with any vendor on this page. It is not published, so do not assume it.

Related reading

For measuring the deployment itself rather than the conversations inside it, see AI call centre costs, pricing and ROI on cost per resolved call, and how to implement AI in your call centre on setting targets before launch rather than after. For the adjacent question of raising quality rather than measuring it, see how to improve customer support quality with voice AI.