Place / receive the call
Hold the conversation
Classify the outcome
Extract the order
Log, route & sync to CRM
Context & problem
Two phone problems, same shop. Outbound: thousands of dormant customers who once bought and never came back — a list nobody has the hours to work by hand. Inbound: order and support calls that have to be caught, understood, and written down, or they never reach the CRM.
A human team can do either well and neither at scale. Reactivation is repetitive and easy to skip; inbound is bursty and easy to miss. The expensive part isn't the talking — it's that every unlogged call is context the business quietly loses.
So I built a voice agent for a German e-commerce shop that does both: a warm, unhurried outbound reactivation call, and an inbound line that classifies, captures, and logs. It speaks natural German, and when someone is genuinely interested it hands the call to a person instead of trying to close alone.
How it works
The diagram above shows the flow; here is what each step does. Deterministic stays deterministic, and an agent only shows up where judgement, language, or synthesis is actually needed.
- 01Deterministic
Place / receive the call
Outbound, a dialer works a campaign's contact list on a schedule; inbound, the shop's line routes to the agent. Which list, which number, which campaign is configuration — no model decides who gets called.
- 02Agent
Hold the conversation
The ElevenLabs voice agent talks with the customer in natural German — a warm reactivation call for a dormant customer, or an inbound order and support call. Real-time language and judgement, one product, no pressure, the customer always has the last word.
- 03Agent
Classify the outcome
After the call, a model reads the transcript and classifies it — order placed, purchase intent, support request, no interest. Deciding what a conversation actually was is judgement in language, so a model does it.
- 04Agent
Extract the order
On a completed order, the model pulls products, customer details, and opt-in out of the transcript into structured fields — the messy last step of turning a spoken conversation into a clean record.
- 05Deterministic
Log, route & sync to CRM
Every call is logged; orders, contacts, and notes with the full transcript and audio link are written into LeadDesk. On genuine interest, an outbound call warm-transfers to a human agent with a pre-qualification note attached. Sending and syncing are rules, not a model.
Checkpoints & logging
The human checkpoint is the warm transfer, and it sits exactly where it should: the moment a customer is genuinely interested, the AI hands the call to a person — with a pre-qualification note and the customer's number already on the agent's screen — instead of trying to close on its own. The agent's job is to open the door, not to sign the deal.
The agent is honest and safe by design. Asked whether it's an AI, it says so plainly. It never leaves a message on an answering machine, hangs up on voicemail or long silence, respects an opt-out immediately, and follows explicit rules for vulnerable people and for reaching a household where the contact has passed away. These aren't afterthoughts — they're written into how it's allowed to talk.
Every call is logged whether or not it converts, with the transcript and an audio reference stored next to the contact, so the CRM learns from every conversation instead of only the ones that end in an order. When a step fails, the run logs it and skips rather than writing a broken record.
Stack, and why
ElevenLabs Conversational AI carries the voice because it holds a natural, low-latency German conversation — the part a customer actually experiences. The conversation, the outcome classification, and the order extraction use models because they're language and judgement in real time; everything after — logging, order creation, contact updates, the CRM sync — stays deterministic so the record is auditable. n8n orchestrates the post-call webhook into the CRM so every step is inspectable, and LeadDesk is the client's own CRM, so calls, orders, and contacts land where their human team already works. Each campaign is an isolated, cloneable unit — a new one is a config change, not a rebuild.
Results
Running live · metrics published here as run data accumulates, not quoted before they can be verified.
It's running live. The production agent handles warm reactivation calls for the shop's dormant customers, and the same system is cloned per campaign across their reactivation lists — a diet product, a music bundle, a beauty line — from one master template.
The honest status: it's live and cloned across several client campaigns, and I publish hard call metrics here once the run data is mine to stand behind rather than quoting numbers I can't verify for your line. What I can say is structural — every call is logged and synced, and a person owns every conversation that turns into real interest.
It's also the direction I'm building further: DSGVO-by-design AI phone assistants for German medical and aesthetic practices — no audio storage, an AVV in place, and a spoken transparency notice — so the same voice-agent pattern works where privacy is non-negotiable.
Want an AI voice agent working your phone line?
Outbound reactivation, an inbound line that never drops a call, or both — pointed at your list and wired into your CRM. Not sure it's the right first move? Start with the audit and I'll map where a voice agent would earn its keep.