Live at a metropolitan municipality's 153 hotline

Every call scored.Every agent improving.

AI that listens to every contact-center call, scores it on the customer's own quality form, and ties each answer to the moment in the call.

audico.ai

Summary · manager briefing

Manager briefing: 910 calls, team score 77.4, up 11.1 points; focus on the closing announcement, missed in 47% of calls

The blind spot

A hotline taking 1,400 calls a day records 50 hours of audio.
One QA specialist can listen to seven.

Heard

7 h

Never heard

43 h · no score, no coaching, no record

Coaching runs on a sample

Supervisors coach from the few calls they happened to hear.

Feedback without proof

Agents dispute a score that comes without the recording.

Failures stay invisible

A dropped transfer or a wrong answer rarely reaches a report.

Volume from the hotline's signed technical specification (~1,400 calls/day). Assumes ~7 h of net listening per specialist per day.

The solution

One system that hears every call
and turns it into action.

153 → archive · daily pull

CONNECT

The archive flows in

Recordings arrive from the phone system by daily pull, ZIP, CSV or API. The extension identifies the agent.

✓ Opening announcement 0:04

SCORE

AI listens and scores

Every call is transcribed and scored on the customer's own form. Each answer is tied to its moment in the call.

Coaching note · listen from 01:07

ACT

Improvement becomes visible

Coaching notes for every agent, a manager briefing, and AIRA, an analyst managers can simply message.

Watch it work

Press play.
Watch a call get scored.

A simulated citizen call. Each criterion lights up at the second it is met, with the timestamp as evidence. The score lands when the call ends.

Simulated recording with AI voices. The same flow runs on every real call at the hotline.

Sample 153 call

SIMULATED · AI VOICES

0:00 / 0:35

Opening announcement
Active listening
Address verification
Accurate information
Request logged
Closing announcement

Call score · whole formSix of its criteria shown here—

Our edge

One call, two readings.
The difference: 77 silent seconds.

The left lane is this recording's real transcript. It shows the words, never the 77 seconds the caller spent waiting.

Sample call · billing dispute

SIMULATED · AI VOICES · 1:55

77 s silent

0:00 / 1:55

Transcript-based tool text

0:00Customer service, this is Emma. How can I help you today?
0:04Hi. My bill this month is twice the usual amount.
0:10Let's take a look. Could I have your name and customer number?
0:19Thank you, Mr. Carter. I'm opening a case for you. Could you hold for a second, please?
0:25Sure.
0:30Hello?
1:00Hello? Are you still there?
1:30Sorry to keep you waiting. Your case is open.

? UnknownThe text shows a hold, not how long it lasted

Audico audio

77s

silent hold
from 0:25 to 1:42

0:25Hold started
0:58Caller: “Hello?”
1:23Caller: “Hello, are you still there?”
1:42Agent returned and apologised

✗ FailHold rule

The left lane is this recording transcribed by Whisper large-v3 turbo, uncorrected (shortened). It heard both “hello?”s but stamped them at 0:30 and 1:00, spread across the silence, and put the agent's return at 1:30. Red: a wrong time.

Scoring

Scored on your form.
With proof.

The customer's own QA form, criterion by criterion, not a generic sentiment score.
Evidence for every answer: the transcript line and timestamp a QA team can check in seconds.
Blank when unsure. No guessed scores; the criterion is left out of the total.
Humans stay in charge. Every AI answer can be corrected, and the score follows.

Call analysis

Product screen · demo tenant, synthetic calls

How good is it

We grade ourselves against
the customer's own QA specialist.

93%

Agreement with the human QA specialist

Five audio criteria · 35 held-out calls

0.84–0.98

Same call, scored twice

Krippendorff's α across criteria on repeated runs

0.75%

False factual claims in published analyses

1,742 claims audited in 974 calls

51–60%

Tone and informal speech

Too weak to publish, so left blank when unsure. Main R&D target.

The 93% sample is small; a blind re-score by the QA specialist on a larger set is planned. We publish the weak criteria next to the strong ones on purpose.

Agent profile

Product screen · demo tenant, synthetic calls

Coaching

Scores don't sit in a dashboard.
They become practice.

Every criterion against the team, for every agent, every week.
A ready coaching note with the exact call and minute to listen to.
The team's habits, not one person: the briefing names the topic for the next training.
Measured week over week, so improvement is a number, not an anecdote.

AIRA · the AI analyst

Managers ask in plain words.
AIRA answers with the call.

On WhatsApp, where managers already are: no new tool to learn.
Names, numbers and the evidence call in every answer.
Briefings and alerts arrive before anyone asks.
Read-only access, scoped to the customer's own data.

Ask AIRA

AIRA1,114 call context

Which three agents need coaching most this month, and on what? Give me one example call for each.

The three agents who need coaching most this month are Sophie Turner, Olivia Bennett and David Park. The team averages 75.3; all three sit well below it.

Agent	Calls	Score
Sophie Turner	26	41.3
Olivia Bennett	143	67.3
David Park	26	67.5

Sophie Turner: opening and active listening met on only 53.8% of her calls (team 87.7%). Example: Sophie Turner · 10/3/2026, where the caller was sent to the website instead of the request being logged.

AIRA's answer, shortened · demo tenant, synthetic calls

Traction

From pilot to a paid public-sector contract
in six months.

MAR 2026

Pilot

First use at Trabzon Metropolitan Municipality's 153 hotline; calibration with its QA specialist.

AUG 2026

Specification

Written with the institution: 99.5% availability SLA, data-protection terms.

23 SEP 2026

Contract signed

Paid license, public-sector procurement.

30 SEP 2026

In production

Hosted in Istanbul, fed by an automated daily pull.

EARLY 2027

Annual renewal

The offer provides for 12-month renewal.

~1,400

calls a day, scored with no manual step

34

agents, each with a profile and score history

KVKK

Türkiye's GDPR counterpart · in-country hosting · no training on customer data · one tenant per institution

Competition

Built for evidence,
not keyword search.

Capability	Global platformsObserve.AI · CallMiner · NICE	Custom projectAgency or in-house build	Audico
Scores the form by listening to the call, not from a transcript	Transcript first	Varies	✓
Local-language phone audio (8 kHz)	Partial	Varies	✓
The customer's own QA form	Partial	Varies	✓
Evidence for every answer	—	—	✓
Blank when unsure	—	—	✓
AI analyst for managers	Emerging	—	✓
What they actually sell	Search & dashboards	Consulting hours	Scored evidence

Based on public information about each vendor category; individual products may differ.

Next market

Türkiye's 153 is Germany's 115.
Same problem, same buyer.

153 · TÜRKIYE · LIVE

1

metropolitan municipality in production

Scored on the institution's own form, hosted in Türkiye under KVKK.

115 · GERMANY · NEXT

14.2M

calls accepted in 2023

By the service centres of the 115 network, run by municipalities and states. Same need: a score for every call, not a sample.

BUILT FOR THE EU

AI Act

ready for high-risk rules

Systems that evaluate workers are high-risk (Annex III). Ours ships evidence per answer, human override and blank-when-unsure. GDPR: call data stays in the EU, never trains a model, one tenant per institution.

115 figures: Steuerungsbericht 2023 (115.de). A first German pilot is a goal for this round, not a signed agreement.

Seed round

The AI quality layer
for every public contact center.

40%

Build the team

A speech/ML engineer, customer success and public-sector sales.

35%

Sell repeatably

From one reference to five institutions on one codebase, plus certification and in-country infrastructure.

25%

Open Germany

Local presence, a GDPR review with a German DPO, a German phone-audio benchmark, and a first 115 service-centre pilot.

Next step: 20 minutes on real calls Kaan ÖzkanCo-founder · audico.ai

Contact-center QA software: $2.25B in 2025, about $4.1B by 2032 (Coherent Market Insights). Round size and terms in conversation.