An AI receptionist for a cloud phone company's own support line
A European cloud phone company wanted to sell AI voice agents to its customers. First it wanted one on its own support line, to learn what really happens on a call. We built it: an agent that knows who is calling, answers from the Help Center, works with tickets during the call and hands off to the right person.
All case studies- Client
- a European provider of cloud phone systems for businesses
- Service
- AI voice agent, n8n automation, Zendesk integration
- Built on
- Retell Conversation Flow, n8n, SIP, Zendesk, the company's Help Center
- Status
- built for the company's own support line first
The situation
A phone company that sells to businesses cannot afford a clumsy AI on its own line. Its customers know telephony.
So the brief was strict. Answer on the company's own SIP trunk. Say up front that it is an AI. Know the caller. Never share account details without verification. Hand off cleanly when a person is needed.
What we built
A front-line receptionist on Retell and n8n.
- 01
It answers over the company's SIP trunk and looks up the number during the greeting. A known number gets no questions.
- 02
An unknown caller asking about an account is verified by voice on three factors: account ID, company and contact name. They are matched by spelling and by sound. Only the failed factor is asked again, three tries at most.
- 03
General questions are answered from 200+ Help Center articles, no verification needed. If the answer is not there, the agent says so and offers a person.
- 04
During the call it reads the caller's open tickets, gives the status, adds a note, reopens a ticket or creates a new one in Zendesk.
- 05
For a hand-off it checks office hours, public holidays and daylight saving, then transfers to the daytime or after-hours line.
What broke, and how we fixed it
This is the part that does not show up in a demo.
The contacts API silently ignored phone filters. A nightly n8n sync of thousands of contacts brought lookups down to about a second.
Calls would not transfer. A full SIP URI and caller ID passthrough fixed it, proven on a live call.
Zendesk moved a ticket to another organisation when the requester was not a member. Now the contact is created first.
The agent announced a save before the tool had run. Each result now has its own step and fixed wording.
Retries counted in the prompt kept looping. The limit moved into the flow.
Tool calls left up to 13 seconds of silence. The agent now keeps talking while a system responds.
From a single prompt to a flow
The first version was a single prompt, and it skipped tool calls. We rebuilt it as a flow-based agent. As the scope grew, the flow went from 8 to 18 to 39 steps.
Timeline
- Julycall and proposal, first agent, the full chain on a real phone call
- Augustticket writes, transfer to a person, test calls and a user guide, client testing and fixes
Want an AI agent on your support line, but only one that holds up on real calls?
We start with a fixed-price pilot from $1,000: one call type on your line.
Book a call about your processWant an AI agent on your support line, but only one that holds up on real calls?
Tell us in two lines. We'll reply with how an agent would take it over and what the first step would cost.
- A fixed-price first step, from $1,000
- Half at the start, half after we check it on your data
- NDA on request
- [email protected]Vitaliy, CEO
- Vesivärava tn 50-201, 10152 Tallinn, EstoniaHead and development office