Doch: Brief it in English, it calls in German
A live voice agent that phones German businesses for English speakers, and checks with you when something isn't in the plan.
Problem
Phone calls in German are the errand English speakers in Germany put off longest. An agent that speaks for you, to a stranger, in a language you cannot follow, has to earn trust fast.
What I did
Designed and built a voice agent in four weeks. You brief it in English, it calls the business in German from your own number, and it asks you on screen when something is not in the plan. It is built around six moments where trust breaks.
Result
Across 35 real calls the tech held, and the hold-and-ask flow worked for the few who stayed on the line. Trust broke at the one line the law requires: of the 26 people who picked up, 14 hung up during the AI disclosure.
How it works
You approve the plan, Doch makes the call
Live app at doch.akberahmed.com
You brief Doch in English, by voice. It turns your answers into a plan you check before it dials: who it's calling, why, which times it can agree to, and what it will and won't share. Then it calls the business in German from your own mobile number, with live English captions as it goes.
If something comes up that isn't in the plan, Doch says “Einen Moment bitte” (one moment, please), holds the line and asks you on screen. Afterwards you get a report of what was agreed and what was shared about you.
The steps show the shape of a call: German on the line, English captions, and a Needs you sheet when Doch has to ask.
Trust moments
Where trust breaks
I designed Doch around the moments where people stop trusting an agent, and built each one in. On real calls, trust broke somewhere else: at the moment Doch said it was an AI.
Each tag is one of the six moments, explained below.
Real calls
35 calls and the line the law requires
I pointed Doch at 35 real German businesses: pharmacies, bike shops, hotels, nail salons, a florist. The tech held. Trust broke at the one line the law makes Doch say.
Of 35 calls, 26 had a person pick up. Four stayed on the line long enough to answer Doch's question.
The biggest drop is at the AI line. Most hang-ups came faster than the line itself, so I suspect people react to the voice or the unknown number before the words land.
From 35 calls to 4 conversations
What I found
The four who stayed
They were small, local places with a one-sentence answer to give, like the Berlin bike shop on the phone screens on the right. And when a business needed the caller to make a decision, that showed up on the call like the Berlin hotel. Recreated from the real call, business names changed. The friction sits on the human side, at the moment someone learns they’re talking to a machine.
The working app
A call you can follow without German
A week later I tested the German agent in a browser and hung up on my own test call on 8 October: I couldn’t follow it, and neither will anyone Doch is built for. So I added live English captions before Doch rang a real phone.
When something isn’t in the plan, Doch says “Einen Moment bitte” and holds the line. A Needs you sheet slides up with the question in English and answers to tap. The 20 simulated users worried about it: 5 said they could not answer in time, 4 said the business would not wait.
I cut the “Take over the call” button: if you could follow the German, you wouldn’t need Doch.
When something isn’t in the plan, Doch holds the line in German and asks you in English.
The phone line
A call from your own number
Twilio only gives German phone numbers to registered businesses, so Doch can't have a number of its own. The call goes out from yours instead. The business sees a real person it can ring back, and the callback reaches you, not me.
When Doch DE needs you, it asks through the app (ask_user) and carries on with your answer.
Keeping the bill and the reach small
The account runs on a $20 balance with auto-recharge off, and voice calls only go to Germany, so the worst case has a ceiling. The Vercel functions run in Frankfurt, so no caption or question crosses the Atlantic.
Proving the number is yours
Twilio only lets you show a number as caller ID after you prove it’s yours: a voice call with a code typed on the keypad. The 20 simulated users were uneasy here: 6 saw nothing on the page while they waited, 5 got no warning that it would ring.
Two agents, two voices
Doch is two agents on ElevenLabs, both running Claude Sonnet 5: one briefs you in English, one calls in German and switches to English if the other person does.
The one that briefs you
The one that makes the call
Choosing the voices
English. The brief was a friend who says "I'll handle it": warm, calm, unhurried. When friends tried the briefing, it sounded annoyed or tired: too much empathy came out as heavy breathing and sighs. I turned it down and it sounded much better.
German. I shortlisted Johanna, Ramona and Ben and picked on a real phone call. Johanna was the friendliest.
Expressive mode
On for both agents: with it off, the German voice sounded robotic and harsh.
Why Claude, and why Sonnet 5
I chose Claude for rule-following and German quality, and Sonnet 5 over 5.5 at the same price because 5.5's worst-case reply took 17 seconds against 6.8.
Two Voices: who's speaking, at a glance
A Doch call has three participants and only two voices on the line: the business in German, and Doch speaking for you. You're watching from your phone, usually without following the German, so the screen has to tell you who's speaking without a label. I built the system in one morning with Claude, on the components Doch already had.
Choosing the pair
I tested three token sets on the same three screens, each with a colour per speaker. The first, Warm sand, paired terracotta and olive, which nearly merge for red-green colour blindness. The second, Head office, separated best but felt like the institutions you're calling. The third, teal and coral, won: warm for your side, cool for theirs, 180° apart, so where the call screen's glows meet they fade to off-white, not brown.
Doch | Design Tokens
Tokens, then components
Each colour became a ten-step primitive scale. Components never touch the primitives; they point at semantic tokens named by role, like the business's bubble and Doch's bubble, which live as CSS custom properties in one stylesheet. Badges take their tone from their label, and the error red holds about 5:1 on the app's backgrounds.
Doch | Component Library
From a colour in Figma to a line of CSS
Two components, followed from Figma to the shipped CSS. The Needs you badge takes its colour from a status and Doch's bubble from a role. In both, the value set in the design system is the token the CSS renders from.
The Needs you badge. Its colours live in the design system, not on the component.
Each colour is one named token, the only place the raw hex is written.
The badge picks its tone from its label, so a new status can't end up the wrong colour.
Looking back
Key takeaways
The tech works. Trust is the unsolved part
Replies came in under a second (median 0.79 s over 31 replies), the hold worked and the plan held. None of that matters if the person hangs up in the first six seconds, and 14 of the 26 people who picked up did, at the AI line.
Disclosure is the front door
Every trust moment I designed sits behind one sentence the law requires. That sentence is now the design problem, not the screens after it. I changed the opening over two rounds of calls (23 and 30 October) and people still left 2 to 6 seconds in.
A safeguard can become a gate
The check question was there to protect the call. It caught one wrong number, and 3 of the 7 calls that opened with it hung up at the question itself. Every step added to the opening has to earn its seconds.
A report has to be honest about failure
“Call finished” over a call where nothing was agreed is worse than no report. The audit trail is only useful if it says when things went wrong. 6 of the 20 simulated users asked for the same thing: tell me what really happened. Doch once promised a business that I would get back to them, and nothing backed it up.
Build log
Follow the build log
I wrote each week up as I built it. Four posts, from the rules and the first prototype to the phone line, the design system and what 35 real calls taught me.
Designing what an AI agent says no to
The rules, the first prototype and the call that broke three of them.
Read Part 1 →From prototype to working app
Twilio, live captions, two voices and the first hang-up.
Read Part 2 →var(--yours-soft)A design system that goes straight to code
Tokens, components and how they reached the app.
Read Part 3 →What 35 real calls taught me
Real businesses, real people, and what I changed because of them.
Read Part 4 →