Quality assurance hears two percent of the calls. Audiqa hears all of them.
Audiqa is a review layer for contact centres. It scores every call in a campaign against the rubric you already use, builds an evidence file for the ones that need a supervisor, and waits for a human before any coaching goes out.
CALL 4-1187 · 07:42 · Collections, campaign “Mora temprana”
Apertura
PASSVerificación de identidad
PASSDisclosures
FAILResolución primera llamada
WARNTono
PASSFrases de cumplimiento
FAIL3 pass · 1 warn · 2 fail → Requires supervisor review
“el descuento lo autoriza mi supervisor”
source: transcript
audio clip, 11 s
source: recording
rule QA-DISC-03 not cited
source: rubric
calls a day on a sixty-seat collections floor
percent programme: the share of calls it hears
calls heard out of four thousand
go out unheard, scored by nothing
A sixty-seat collections floor takes around four thousand calls a day. A two percent programme hears eighty of them. The other three thousand nine hundred and twenty go out unheard, scored by nothing, and the failures that matter are not evenly distributed through the day: they cluster at the end of a shift, on the longest calls, and on the agents with the least seniority.
The sample is defended on cost, and the cost is real. Listening is the most expensive hour in a quality department. What changed is that listening is no longer the only way to find the call.
Audiqa does not remove the supervisor. It removes the search.
Three stages. The last one is yours.
Ingest.
Point Audiqa at the audio you already store, or drop it to S3. Per campaign you set a consent flag and a retention window. Nothing is scored before both exist.
Score.
Every call runs against the campaign rubric. The transcript is a by-product; the output is a score per dimension with the span that caused it.
Escalate.
Calls under threshold produce an evidence file and land in the supervisor’s queue. Coaching stays a human decision.
Six dimensions. Yours to edit.
The default rubric is the one collections and telco floors in Argentina already run on paper. Every dimension is a rule with an id, and every fail cites one.
The cost of listening fell through the floor.
Three years ago scoring every call cost more in compute than the agent cost in salary. Speech recognition was billed per minute by vendors who priced for the enterprise, and a classifier good enough to cite a rule did not exist at a price a mid-size floor could carry.
Both halves moved. Recognition on open models runs at a fraction of a cent per minute on commodity GPUs, and scoring a rubric dimension is a classification problem with a short context. The work did not become easier. It became affordable.
What a supervisor opens.
An evidence file is not a verdict. It is the smallest set of things a person needs to decide in under a minute.
Autonomy inside limits you set.
Thresholds are per campaign, not global.
Coaching is a proposal until a supervisor approves it.
Permissions are per role: read audio, write score, write file.
Retention defaults to 90 days and is configurable down to 7.
Audio is never used to train a third-party model without opt-in.
Security and data handling
Where the audio already is.
Read-only in every case. Audiqa never writes back into the telephony platform.
Floors where the sample has already failed an audit.
Collections, telecom, prepagas, utilities. Forty seats and up. Teams that already store audio and already run a written rubric.
Under ten seats, pure outbound sales (different rubric), teams with no stored audio.
By audio hours, monthly.
Three bands, a monthly minimum, invoiced in USD. No per-seat charge, because seats are not what costs us money.
What Audiqa does not do.
It does not score calls where the consent flag is off.
It does not handle overlapping speech well: two people talking at once is the known weak point.
It does not coach. It proposes, a person decides.
It does not run on your hardware yet.
Questions we get.
No. The transcript is a by-product of scoring and is kept only as long as the retention window.
No. It replaces the search for the call. The judgement stays.
The recognition model is tuned for it, and each campaign carries its own glossary of product names, plan names and legal phrasing.
Ninety days by default, configurable down to seven, deleted on request inside the window.
Not today. Cloud only, Toronto region.
Where this goes.
Rubric scoring and evidence files.
A coaching assistant that drafts the conversation from the file.
Trend detection per seat, shift and campaign, before the audit finds it.
We expand one stage at a time and only after the previous one holds.
Two people.
Malena Rossi
ProductTen years running quality on collections floors, the last four as head of QA on a three-hundred-seat BPO.
Iván Sequeira
Machine learningSpeech recognition and classification.
Audiqa SAS, Lavalle 1520, piso 6, CABA. Founded 2026.
TeamWe are taking a small number of design partners.
Tell us what you are trying to catch and what your floor looks like. We reply from a person, not a queue, and we will say no if the fit is wrong.