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.
When a vendor announces a model partnership, the only question that matters to you is whether anything about your own deployment just changed.
Why it matters: Avesta Labs, the company that builds Metricsense, has been named an OpenAI Select Partner within the OpenAI Partner Network. That is our news rather than yours, and the gap between those two things is the whole subject of this post.
So take the concession before the argument. A partner tier is a commercial relationship between two companies. It is not a security control, it is not an audit, and it is not evidence about how your call recordings are handled. If this announcement reached you as a reason to trust us more, that is not a good reason, and we would rather say so on the day we announce it than let you work it out later.
We are the team that reads the calls at Metricsense. The corpus behind our published work is 9,218 calls from one anonymised Australian life-insurance business, 16 April to 10 August 2026. What you should have by the end of this is five questions to put to any vendor whose model provider changes, including us.
The first question this news should raise is where your calls go
Compliance leads have a reliable instinct about this, and it is the right one. The words that matter in an AI announcement are never the tier or the logo. They are the ones describing the path your data takes: what leaves your systems, what is stripped before it leaves, which region it is processed in, and who can read it on the way.
A partner announcement answers none of those. It is a fact about two companies, published by both of them, and it can be entirely true while telling you nothing you could take to a risk meeting.
Nothing about your deployment changes, and that is the point
Three things hold today exactly as they held last week. They are the ones we would expect you to check rather than take on trust.
- Metricsense deploys in your own cloud and your own region. The partnership does not move a deployment, create a new copy of your recordings, or add a destination for them.
- PII is stripped before any transcript reaches a model. That step sits in front of the model, so it does not care which model is behind it.
- Every score still links to the transcript moment that produced it. A flag you cannot open and read is not something we would ship, whoever built the model underneath.
If any of those had changed, that would be the announcement, and it would not be phrased as good news.
A partner tier is a commercial status, not a security control
Here is where we will argue against our own press release. The villain worth naming is not a company, it is a habit: judging an AI vendor by the logo of its model provider. It is an easy habit to fall into, because the logo is the part that is legible from outside, and the data path is the part that takes an afternoon and an uncomfortable meeting to establish.
The habit fails in both directions. A vendor with an impressive model partner can still be shipping your customers' health disclosures to a region you never approved. A vendor with no partnership at all can be running a clean, regional, PII-stripped pipeline that would survive an audit. The partnership and the data path are simply different facts, and only one of them is yours.
Our own site has a check that fails the build if a customer's company name, their principals, their suburb or their workspace identifier appears anywhere in it. It runs before every deploy and again before every commit, so a slip cannot ship quietly. A build gate is a blunt instrument and it is a poor substitute for consent, which is why we ask for that separately. It is also the only kind of promise we know how to keep mechanically.
The finding still has to survive without the model that produced it
The reason we are relaxed about model changes underneath is that we do not ask the model to be the evidence. A better model produces a better candidate. It does not get to produce a conclusion, because the conclusion has to be readable in the transcript by a person who does not trust either of us.
Scored calls that came back with no escalation risk recorded, in one anonymised Australian life-insurance corpus, over the window 16 April to 10 August 2026. Restated here word for word from work already published on this site rather than re-queried, and the same rule applies to it as to any finding: it is defensible because you can open the calls behind it, not because of what produced it.
That is the test a model partnership does not affect either way. Reassurance you cannot open is worth about as much as an accusation you cannot open.
What we expect from this, and what we will not claim yet
The honest version of the upside is narrow: closer access to the people who build the models we already use, earlier sight of what is changing, and a shorter route to an answer when something behaves oddly on real transcripts. That is worth having. It is also the sort of benefit that shows up in our week rather than in your quarterly numbers, and we would be overclaiming to describe it any other way.
For us, this is a continuation, not a beginning. We have already helped organisations adopt and scale AI in the real conditions of their business: responsibly, reliably, and with human teams in control the whole way.Vivek Satasiya, Chief Product Officer, Avesta Labs
What we will not do is tell you the analysis got better today. We do not yet know which of the newer model capabilities will move the numbers that matter here, and the only way to find out is to measure them against transcripts we have already scored. When we have run that, we will publish the direction of travel including the parts that go the wrong way. We have published our own denominators moving underneath live copy before, so this is a habit rather than a promise.
Five questions for any vendor whose model provider changes
These work on us. They work better on a vendor who was hoping the announcement would do the reassuring for them.
- Where is my audio processed, and in which region is it stored? Name the region, not the provider.
- What is removed from a transcript before a model sees it, and can I see that step run on one of my own calls?
- Does anything about that answer change if you change model provider next quarter?
- When your system flags a call, can I open the exact moment that produced the flag, without asking you?
- What did you get wrong in the last six months, and where is it written down?
The fifth one is the one that separates vendors, and it is the one we would rather you asked us in a meeting than skipped.
The announcement worth reading is the one about your data path
We are glad about the partnership, and we are aware that being glad about something is not an argument. If this post is useful at all, it is because it prompted you to send five questions to a vendor, and we would count that as a good outcome even if you never send them to us.
Evidence Intelligence, by Avesta Labs
The Metricsense team reads 100% of customer conversations, calls, reviews, tickets and surveys, and links every finding to the exact source quote. We write about what the full record shows once you stop sampling: compliance, quality, and the calls that never make it into a QA report.
Ask us the five questions on your own calls
A pilot runs Metricsense on your recordings, in your cloud and your region, so the deployment answers are things you watch rather than things we assert. If you are running fewer than a few hundred calls a week, a sample probably is adequate for you and we would rather say so now than sell you a pilot you do not need.
Keep reading
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.
When everyone makes the same mistake, it is the script
One adviser getting it wrong is a coaching conversation. The same gap across the whole floor is a wording problem, and coaching will not touch it.
