The real audio chain
Telephony, speech input, the model, speech output and conversation flow are tested together.
Operator-run audits · First live audits passed
We call your voice agent. And we measure latency.
We test your voice agent with authorized calls in German. You receive measurements for reachability, response times and call flow, with a recording, transcript and traceable evidence.
Request an audit
No self-service: an n3tz operator prepares every audit and supervises it personally.
Illustrative audit report
A demo is not a regression test
One small change can reshape the whole call.
A new prompt, another model or a changed tool schema can silently break bookings, transfers and safety rules. Text evals cannot see what happens on the phone line.
Repeatable scenarios
The same approved scenario version runs before and after a change.
Sealed evidence
The audit report and evidence package carry SHA-256 checksums. You can recompute every one yourself.
How an audit runs
From first contact to the sealed evidence package.
- 01
Request and scenarios
Tell us about your setup. Together we define which scenarios run against your number.
- 02
Authorize the target
You authorize the target number and time window in writing. No call starts without it.
- 03
Supervised test calls
An operator runs the German test calls on a dedicated host, separate from the production n3tz suite.
- 04
Report and evidence
You receive the audit report and the sealed evidence package: recording, transcript, event log, each file with a checksum.
From our own tests with Griffi
What we found and tested again.
Quality Lab now also checks whether answers meet the agreed expectation in meaning. These examples come from internal tests with synthetic data. Measurements, content assessment and backend evidence stay separate; results apply to the scenarios tested.
1
A sentence fragment is not evidence
An earlier call captured only an incomplete answer. The current content check treats such fragments as insufficient evidence. They cannot establish a correct answer or a booking.
2
No availability, no booking
A booking test failed because the test tenant had no bookable calendar slot configured. The run remained a failure. We then added an availability window only to the synthetic test tenant.
3
Tested again, verified separately
After a booking-confirmation fix and with availability configured, the repeat test call passed all three content checks. A separate database query confirmed exactly one newly booked appointment. The spoken promise alone would not prove that.
Controlled testing
Test calls need hard boundaries.
Every audit runs only with written authorization, inside the agreed time window and under personal supervision by an operator. Execution lives on a dedicated host, separate from the production n3tz suite.
Authorized targets
No arbitrary production numbers and no unlimited dialing.
Isolated audit host
Test runs execute in a separate test environment on a dedicated host, isolated from the rest of n3tz operations.
Human supervision
Every run stays deliberate. An operator prepares it, watches it and reviews the report.
For teams shipping voice AI
Built for the gap between “it works” and “it can go live.”
01
Voice AI teams
Catch regressions before releases and reproduce the same evidence after a fix.
02
Agencies
Test client agents against agreed scenarios and hand over results that can be traced.
03
Product and QA leads
Review the outcome, phone path and control boundaries together, not just a transcript.
Personally supervised
Put your voice agent in the Quality Lab.
Tell us briefly what you run. A person from n3tz replies personally, clarifies scenarios, process and price in a call and tells you honestly if an audit makes no sense for you yet.
Frequently asked questions
What to know before requesting.
Which voice agents can Quality Lab audit?
Any agent that is reachable through a real phone number and that you authorize us to call in writing. Test calls run in German. Agents without a phone number, for example web-widget or chat only, cannot be tested today.
Why real phone calls instead of text simulations alone?
Text tests are faster and remain important. They cannot reveal SIP failures, speech recognition, speech output, interruptions or phone latency. Quality Lab adds a controlled end-to-end path.
Does Quality Lab record the test calls?
Yes. The evidence package for each call contains the recording, the transcript and the event log, every file with its own SHA-256 checksum. Only the authorized test call is recorded, inside the agreed time window.
Does Quality Lab also check what my agent says?
Yes. We assess observed answers in their conversation context against content expectations agreed in advance. The result is met, not met or insufficient evidence. Missing answers and evaluation errors never count as success. Whether a booking was actually saved requires separate backend evidence, which we agree with you.
You build voice agents yourselves. How independent is an audit?
Full disclosure: n3tz runs its own voice agent, Griffi, for German trade businesses, not a competing product to voice-AI platforms. Audits run on a dedicated host, separate from production, and you can verify every checksum yourself. For test scripts and internals we offer a mutual NDA.
What does an audit cost?
There is no fixed pricing model yet. We clarify scope and price in the first call, honestly and without obligation.
Is the request a newsletter signup?
No. Your request triggers exactly one personal reply. You receive n3tz updates only if you enable the separate option and then confirm your email address.
Your voice agent deserves more than one good demo call.
Request a supervised audit. We will get back to you with a concrete proposal.
Request an audit