---
title: "How to deploy an AI voice agent in a cloud call centre"
description: "How to deploy an AI voice agent in the cloud: what to prepare, how phone numbers connect through a SIP trunk, and when on-premises becomes cheaper. Read on."
url: https://aifortis.com/en/web-ai-callcenter/guide/
language: en
last_updated: 2026-11-08
publisher: AI Fortis Ltd. (Sofia, Bulgaria)
---

# How to deploy an AI voice agent in a cloud AI call centre

How to deploy an AI voice agent in the cloud: what to prepare, how phone numbers connect through a SIP trunk, and when on-premises becomes cheaper. Read on.

## What a cloud AI call centre is

AI voice agents hold real phone conversations: they recognise speech, decide what to answer with a language model together with your own information and systems, and speak the answer in a natural voice. Unlike a phone menu, the caller does not press keys or choose from a list — they simply say what they need. In the cloud model the provider runs the models, the telephony connection and the platform; you connect your phone numbers, decide what the agents may say and do, and pay for use. You do not buy servers or maintain models, which is why the cloud is usually the fastest and cheapest way to start.

## Inbound and outbound: two different jobs

Inbound agents answer your lines: support questions, order status, bookings, opening hours, routing. Their job is to resolve the request or hand it to the right person, at any hour. Outbound agents call people on a list: new leads, reminders, confirmations, win-back offers. Their job adds rules that inbound calls do not need — working hours, how often a number may be tried, what to do with voicemail, and a do-not-call list that must always be respected. Both can run on the same platform, so many teams start with one direction and add the other later.

## Choosing the first scenario

General guidance: pick a call type that is frequent, repetitive and well understood — the kind of call your best agent could answer half asleep. Good candidates are order or booking status, appointment confirmations and reminders, first qualification of new leads, and frequently asked questions. Avoid starting with calls that need long negotiation, complex judgement or emotional handling; keep those with your people and let the agent hand them over. Choose a scenario where you can measure the result, such as resolved calls, booked meetings or confirmed appointments, so you can compare before and after.

## What to prepare

Write down the scenarios you want to automate and the questions callers ask most, with the answers your best agent would give. List what the agent must never say or promise. Decide the rules for handing a call to a person: which topics, which customers, which hours. Name the systems the agent must look things up in — orders, bookings, calendars, a CRM — and who manages them. Collect your phone numbers and the telecom provider they come from. For outbound calls, prepare the lead list, the hours you may call, how many attempts a lead gets, and your do-not-call list.

## Connecting phone numbers

Calls reach the platform through a SIP trunk or a number from a telecom provider. A SIP trunk is a phone line delivered over the internet: your provider gives you the connection details, and the platform registers with them. Your numbers normally stay with your provider, so callers keep dialling the numbers they already know. Before go-live, test calls in both directions, check how caller numbers are displayed, and agree what happens if the platform cannot be reached — for example forwarding to your existing team.

## Connecting your systems

An agent that can only talk is useful; an agent that can look things up is far more useful. With live connections to your systems, the agent checks an order, finds a free slot, books a meeting into a calendar or updates a record during the call. For each connection decide what the agent may read, what it may write, and what it must never touch. Give it only the permissions it needs, and make sure every action it takes is logged. Results, transcripts and recordings should flow back to where your team works.

## Designing the hand-off to a person

A good AI call centre knows when to stop. Define the situations in which the call goes to your team: a topic you reserve for people, an upset caller, a request outside the agent's knowledge, or simply a caller who asks for a person. Decide where transferred calls go during and outside working hours. Make sure the information collected so far travels with the transfer, so your colleague does not have to ask the same questions again. Test the hand-off as carefully as the conversation itself.

## Outbound campaigns: pacing, retries and voicemail

Outbound calling needs discipline. Set working hours per campaign, the maximum number of simultaneous calls and calls per minute, so you never call more people than your team can follow up. Decide how many attempts a lead gets and how long to wait before a retry; no answer, busy and voicemail outcomes are typical reasons to retry. Voicemail detection keeps the agent from talking to an answering machine as if it were a person. And enforce the do-not-call list on every call, without exceptions.

## Compliance: telling callers, consent and retention

Under EU AI Act Art. 50, people must know when they are talking to an AI, so the agent says so at the start of the call. Recording calls needs a legal basis and usually a clear notice. Decide how long recordings and transcripts are kept, who may listen to them, and how they are deleted when the period ends. Ask any provider about separation between clients, encryption of data and credentials, and the data processing agreement. These are not optional extras; they belong in the design from the first day.

## Measuring success

Agree the measures before go-live: share of calls resolved without a person, calls handed over and why, booked meetings or confirmed appointments, missed calls, and the cost per call. Listen to a sample of conversations every week at the start and turn what you hear into changes to the agent's knowledge and rules. A live portal with recordings, transcripts, outcomes and per-call cost makes this routine work rather than a project.

## Cloud or on-premises

The cloud is cheaper to start and needs no hardware: you pay for use and can stop or change direction at any time. When volumes are high and steady, or data must stay inside your building, an on-premises system is usually cheaper over a few years, because the one-time investment is spread over many calls and no per-minute fee grows with volume. Many organisations start in the cloud, learn what works on real calls, and then decide whether to move on-premises.

## Next step

See it on your own case, or start with a free, no-commitment offer.

- GDPR by design
- 24/7 operation and support
- A team of 5 in Sofia, EU

Already running in production at organisations in Bulgaria — names under NDA.

[How we prove it →](https://aifortis.com/en/how-we-prove-it/)

## Frequently asked questions

### How long does it take to start?

It depends on the scenario and the integrations. Ask us for an estimate for your case.

### Do callers know they talk to an AI?

They must: under EU AI Act Art. 50 the agent says at the start of the call that it is an AI.

### Can an AI agent transfer a call to a person?

Yes. Hand-over rules are part of the design, and the information collected so far travels with the transfer.

### What is a SIP trunk?

A phone line delivered over the internet. Your telecom provider gives you the connection details and the platform registers with them.

### Which calls should we automate first?

Frequent, repetitive and well-understood calls with a measurable result — for example status questions, confirmations and reminders, or first qualification of leads.

---
Source: https://aifortis.com/en/web-ai-callcenter/guide/
