I made 35 phone calls in German. Four people stayed on the line.

Week four of building Doch.

30 Oct 2026

Doch has now phoned 35 real German businesses. Twenty-six had a person pick up, and only four of them stayed on the line. This post is about the first few seconds: where an AI voice agent loses people, whether changing the opening helps, and what I can’t say yet.

35 businesses called, 26 had a person pick up, 4 stayed on the line, 9 never reached a person, 2 of them voice assistants.

“Stayed” means they heard the question and answered it. Two more calls on 23 October rang out and aren’t counted.

Funnel of the 35 calls: 35 called, 26 had a person answer, 24 were still on the line when Doch spoke, 20 heard Doch say it’s an AI, 6 stayed after hearing it, 4 answered Doch’s question.

From 35 calls to 4 people who stayed. The biggest drop is at the AI line.


How I ran the calls

Until now Doch had only called friends playing receptionists. A doctor’s practice on 15 October was the first real errand. The real run came on 23 and 30 October: pharmacies, bike shops, hotels, nail salons, a florist. I picked each business by hand, and Claude read every ElevenLabs transcript and logged the call in a sheet, tagged by where it ended.

Two of the nine weren’t people at all. A hotel and a restaurant answered with their own voice assistants, and both talked back: one explained its room rates and breakfast price, the other offered a table at 18:00 or 20:15. That’s 2 of 2, so a lead, not a rule, since both businesses chose to run one. The friction sits on the human side, at the moment someone learns they’re talking to a machine.


Most people leave at the AI line

Doch has to say it’s an AI (EU AI Act, Article 50), so every call hits the same gate. After the hello, it says it’s a digital voice assistant and asks if it may ask a quick question. That sentence is where the funnel narrows. Of the 26 people who picked up, 14 hung up during it, 3 more left at the question before it, and 4 stayed.


Changing the opening barely moved the drop-off

The opening is a branch in my server, not in the prompt. If the greeting already names the business, Doch skips ahead and says the AI line with the name in it. If not, it first asks a check question, “Spreche ich mit …?”, to catch wrong numbers, and the AI line comes after the yes.

Across 15 and 23 October, 1 of 8 people stayed. After the change, 3 of 15 did. Neither opening changed the drop-off much: most people still left within seconds of the AI line.

Stacked columns of people by opening. 15 and 23 October, 8 people: 1 stayed, 6 left at the AI line, 1 confused by the recording ask. 30 October named greeting, 9 people: 3 stayed, 6 left at the AI line. 30 October check first, 6 people: 2 left at the AI line, 3 at the check question, 1 said no thanks.

People by opening: 8 on 15 and 23 October, and 15 on 30 October.

Within 30 October, the named greeting did better than check first: 3 of 9 stayed, against 0 of 6. That isn’t proof, since a gap that size turns up about 2 times in 10 by chance and I picked the businesses. The check question also caught only one wrong number, while 3 of 6 check-first people hung up on it before Doch reached the AI line. A step meant to protect the call became a second gate.

Most hang-ups land within six seconds

On 30 October, eleven people hung up on Doch. Nine of them did it between two and six seconds in, often mid-sentence.

Column chart of when 11 people hung up on 30 October, in 2 second bins. 1 to 2 seconds: 1. 3 to 4 seconds: 6. 5 to 6 seconds: 2. 11 to 12 seconds: 1. 15 to 16 seconds: 1. 9 of 11 hung up within 6 seconds.

Hang-ups on 30 October, in two-second bins.


That’s shorter than the AI line, so reply speed never gets a chance to matter. I suspect people react to the voice, or to an unknown mobile number, before the words land. Only the audio can tell.

The four who stayed

The people who stayed were small, local places: a doctor’s practice, a pharmacy, a hotel and a bike shop. Each question had a one-sentence answer, whether a price, a yes or no, or an email address. Here is the bike shop, Fahrradhaus Meyer, start to finish. It’s recreated from the real call, and business names are changed.

Briefing, call and report for the Fahrradhaus Meyer call. Recreated from the real call, business names changed.


A routine question, answered by name, in one sentence: that was the pattern. The type of business alone didn’t predict who stayed.

When Doch has to ask me

Needs you is the sheet that slides up on my phone when Doch hits something that isn’t in the plan. At Sommerhof Hotel the person said reception was closed and asked for the request by email. Doch said “Einen Moment bitte” and held the line. The sheet showed the question, why Doch was asking, and what they’d said in German. I answered about nine seconds into a 25-second countdown, and Doch read the address back.

The Needs you sheet during the Sommerhof Hotel call, through to the report. Recreated from the real call, business names changed.




It held again at the restaurant for about 10 seconds, but there the other side was a voice assistant. One person waited, and I answered fast, so how long a receptionist will hold is still untested.

Latency: about 0.8 seconds a reply

From ElevenLabs’ own timings on ten German calls, across 31 replies, the model took a median of 0.79 seconds, between 0.55 and 0.96. The voice needed about 0.09 seconds to start and speech recognition about 0.07. That isn’t the silence a person hears, which also includes the wait for them to finish speaking. I haven’t measured that from the audio.

What I tested, and what I haven’t

Disclosure: who stays after “I’m an AI”?

Tested

4 of 26 stayed.

Confidence and escalation: asking me when it has to

Tested

Needs you worked twice.

Hold tolerance

Partly

One person and one voice assistant waited.

Latency

Partly

About 0.8 s per model reply. The silence a person hears isn’t measured.

Intent preview, action audit and undo, and interruption didn’t get a proper test. No call reached a booking, and nobody has been asked how the voice sounds.

What I dropped

The opening used to ask permission to record the call. Calls ended at that question, so on 30 October I dropped it. Doch no longer asks to record, audio is kept for a day and deleted by hand after I’ve listened, and the transcript stays for about 30 days.

Key takeaways

  1. The AI line is where people leave. 14 of 26 left during it, 3 more at the check question, and 4 stayed.

  2. Changing the opening barely moved it. 1 of 8 (13%) stayed before the change and 3 of 15 (20%) after.

  3. The check question costs more than it saves. It caught one wrong number, and 3 of 6 check-first people hung up on it.

  4. Hang-ups are fast. 9 of the 11 on 30 October came 2 to 6 seconds in.

  5. The people who stayed had a cheap question to answer.

  6. Voice assistants didn’t mind. 2 of 2 talked back.

  7. Replies took about 0.8 seconds. That’s model time only.

  8. Escalation held the call. Needs you worked twice. Hold tolerance is untested.

What came next

Calls end today, because Doch only phones on weekdays and the hard stop is 1 November. Before then I’m fixing what the calls showed: digits in names (“Phone Profi 24” came out as “zwanzig vier”), the end_call tool that fails on some goodbyes, and the line that promises “Akber will get back to you”, which nothing backs up. Next is the case study, which has to be honest about a sample this small.

If you run a small business in Germany and an AI has ever called you, I’d like to know what you’d have wanted it to say first.

Doch build log
01Designing AI agents: what they say no toWeek one · 7 Oct02AI voice agent prototype to working app in a weekWeek two · 14 Oct03Two Voices: a design system for an AI appWeek three · 21 Oct
04I made 35 phone calls in German. Four people stayed on the line.Week four · 30 OctYou're hereYou're here