Full coverage is less invasive than a sample, not more
What makes monitoring feel punitive is not that somebody listened. It is that somebody chose you.
The objection to reading every call is a good one, and I want to put it more strongly than a vendor normally would: more listening is more listening, and no amount of framing makes that untrue.
Why it matters: accept that premise as the whole story and you will roll this out as a surveillance programme, be technically correct, and lose your best advisers doing it.
But it is not the whole story. In a sampled regime, selection carries information. Being reviewed means a name came up: a complaint landed, a manager had a hunch, somebody is on a list. Advisers read that correctly, because it is usually true, and the review has already said something about them before anyone has listened to a second of it.
I lead product at Metricsense. Over the past four months I have worked through 9,218 calls from one anonymised Australian life-insurance business, 16 April to 10 August 2026, read end to end rather than sampled. By the end of this you should have four rules for introducing full coverage without wrecking your floor, and the words for the first briefing.
Two boundaries first, because they bound everything below. Every count here is what our scoring recorded, which is not the same as what an adviser did or did not say. And this is one business in one window, so it tells you what to go and check on your own floor, not what you will find there.
Being chosen is the accusation. Being heard is not.
Ask an adviser what they felt the last time a call of theirs was pulled for review. Almost nobody says "observed". They say "picked".
That word is the whole problem. A sampled review is an event with a cause, and everyone on the floor spends the first ten minutes trying to work out what the cause was. The coaching conversation that follows may be generous, well prepared and entirely fair. It arrives after the message has already landed.
Remove the choosing and the accusation goes with it. Nobody was singled out, because nobody was selected.
On 7,656 of 8,460 scored calls we found nothing
This is the part of full coverage people do not expect, and it is the part that does the political work for you. Most of what you read is fine.
scored calls came back with no escalation risk recorded at all. Reading everything mostly produces evidence that your floor is doing its job.
A sample cannot give you that sentence. A sample of 250 calls that finds three problems tells you there were three problems in 250 calls, and nothing whatsoever about the other 8,968. Full coverage is the only way to say "we looked at all of it" and have the sentence mean anything.
There is a second effect, and it is the one advisers notice within a fortnight: the review load stops being lumpy.
Your advisers are already monitored. What they are not is monitored evenly.
Under sampling
Four of twelve advisers had no call reviewed at all. Selection follows availability and suspicion, not risk.
Reading every call
Same checks, same window, everyone. Nobody is singled out, because nobody is selected.
Illustrative distribution across twelve advisers, not measured data. The point is the shape, not the counts.
A sample turns every review into an implied allegation
Think about what a QA sample actually is from the floor's point of view. It is a small number of calls, chosen by someone with a reason, and the reason is not published.
You cannot make being singled out feel supportive. You can stop singling people out.
Every attempt I have seen to fix this works on the wrong half of the problem. Softer language in the coaching template, a friendlier name for the scorecard, a promise that this is developmental and not disciplinary. All of it addresses the review. None of it addresses the selection, which is where the information actually is.
Selection-free monitoring: every call read against the same checks, in the same window, for everybody, so that being reviewed carries no information about you. It is a fairness property, not a coverage number, and the two are easy to confuse.
The fear of a floor leaderboard is completely justified
Here is where I have to concede more than the argument is comfortable with. Full coverage does not make monitoring fair. It removes one specific unfairness and hands you a much more powerful instrument, and a powerful instrument in the wrong hands is worse than a weak one.
If you read every call and then rank your advisers publicly by score, you have not built selection-free monitoring. You have built a leaderboard with better data, and your floor is right to be afraid of it. We were asked for exactly that feature and said no, which cost us a conversation I would rather have kept.
The fear is not really about being heard. It is about what happens to the number afterwards, and that part is entirely a decision you make.
Four rules that decide whether this lands or blows up
These are not soft-launch tactics. Each one is a commitment you either make in the first week or never credibly make at all.
- Announce it before it runs, to everybody at once. The thing that reads as surveillance is discovering it happened. Nothing else in this list survives getting this wrong.
- Publish the checks. Every adviser should be able to read the plain-English sentence that defines each check, before their first call is scored against it. A check nobody can read is a rule nobody can follow.
- Separate the flag from the consequence, out loud. A flag is a candidate issue with evidence attached. It becomes a finding when a person reviews it and agrees. Say who that person is, by name, in the briefing.
- Give the adviser the evidence too. If a score links to the exact moment in the transcript, the adviser gets that link as well as the manager. A score you can argue with is a score people will accept.
What to say in the first briefing, in full
Managers ask me for the wording more often than they ask for anything else, so here it is. This is a constructed script, not a transcript of a real briefing, and you should put it in your own voice before you use it.
Four questions before you switch anything on
- Who sees an adviser's score, and can that list get longer without anyone deciding it should?
- What happens between a flag being raised and a person looking at it? If the answer is "nothing, it goes on the report", you have automated an accusation.
- Can an adviser see the evidence behind their own score on the same day the manager can?
- If you stopped tomorrow, would the floor be relieved? Ask two advisers you trust, and take the answer seriously.
Coverage is the easy half. Fairness is a design decision.
Reading every call is now a procurement question. Whether the result feels like coaching or like a camera is not, and no vendor can sell you the answer.
One thing I still cannot tell you: whether the fairness effect holds after the first quarter, once the novelty wears off and full coverage is just how QA works. Four months is not long enough to know that, and I would rather say so than tell you it is settled.
Head of Product, Metricsense
Gaurav Soni leads product for Metricsense. He has spent 13 years in product at the intersection of data, analytics and AI, starting in engineering, building APIs, data pipelines and analytics products, and moving steadily closer to the decisions leaders actually make. The thread through all of it is the same: turning messy, unstructured conversation into evidence a team can act on. At Metricsense that means reading 100% of a company's calls, reviews and tickets, and linking every finding to the exact quote.
Run it where your advisers can see it
We will run a pilot on your own calls, in your own cloud and region, with PII stripped before any transcript reaches a model. Worth doing if you take enough advice calls that sampling has stopped feeling like a choice. If you run under about 200 calls a week, a sample probably is adequate and I would rather tell you that now than sell you a pilot.
Keep reading
Avesta Labs is an OpenAI Select Partner. Your calls stay put.
The company that builds Metricsense has joined the OpenAI Partner Network. Here is the short list of what that changes for the recordings you have trusted us with, and the longer list of what it does not.
A confidence score is not evidence
0.94 tells you how sure a model is. It tells you nothing about what your adviser said. Only one of those survives a regulator.
An ask was recorded on 34 of 8,479 scored calls
Read end to end rather than sampled, the clearest revenue pattern in this corpus was not a bad pitch or a hard objection. It was a conversation that simply stopped.
