TL;DR — The control unit AI is not a virtual power station with a more sophisticated voice menu. He is an agent who listens, understands, responds in natural Italian, qualifies the caller and updates the CRM — without a person in the middle. It makes sense for SMEs B2B with at least 15-20 inbound calls per day and standardized flows (booking, qualification, sorting). With less volume or too uneven processes, the cost of setup exceeds the benefit: better wait or stay on the traditional cloud control unit.
Every week we talk to entrepreneurs who have the same problem: incoming calls arrive, but they are not handled well. The secretary's on a break. She's late. She takes notes as they disappear. Or the automatic answer is an IVR menu of 2009 — "Premi 1 for the commercial office, press 2 for assistance" — that 60% of the callers abandon before arriving at the right choice (source: Aircall State of the Call 2024).
The control unit AI is the answer to this problem. But like every technology, it is only worth it if applied in the right context. This guide explains exactly what it is, how it works, how much it costs and — above all — when it does not suit.
What is an AI control unit (and what is not)
The confusion arises because the term "AI central" is used for very different things. Let's try to make clarity.
A traditional virtual control unit (or cloud control unit) manages the telephone infrastructure: geographic numbers, call routing, voicemail, recordings. Companies like Ringcentral, Vonage, Wildix operate in this space. They are solid solutions, but the "brain" is not there: the system does not understand what the caller says, it just passes the call where you tell him.
A Advanced IVR (Interactive Voice Response) adds voice menus and key words recognition: "commercial sayings" for the commercial department". Better than the key menu, but still rigid — comes out from the tracks as soon as the caller says something unexpected.
A control unit AI is something different: a voice agent AI connected to the company number. When a call comes, the agent:
- Listen in real time (speech-to-text)
- Understands the caller's intent with a linguistic model (LLM)
- Responds in natural Italian with synthesized voice (text-to-speech)
- Performs actions: book an appointment, create a record in CRM, passes the call to a commercial, sends a confirmation SMS
It is not a text chatbot with the voice above. It is a system that manages end-to-end telephone conversations, with the same flexibility of a person — but available 24/7, at variable cost, without training and without diseases.
How it works: the 3 technological layers
To understand the limits (and costs) of an AI control unit, you have to understand how it is done. There are three technological layers that work in sequence on each word pronounced by the caller:
| Layer | What's he doing? | Typical technologies | Added latitude |
|---|---|---|---|
| Speech-to-Text (STT) | Transcribes spoken in real-time text | OpenAI Whisper, Google STT v2, Deepgram | 200–400 ms |
| LLM (recreation) | Understand the intent, decides the answer, executes actions | GPT-4o, Claude 3.5, Gemini 1.5 | 300–700 ms |
| Text-to-Speech (TTS) | Converts the answer to natural voice | ElevenLabs, OpenAI TTS, Google WaveNet | 150–300 ms |
Total latency — the time between the end of a caller's sentence and the beginning of the agent's response — is typically 700–1.400 ms. Quick enough to look natural, slower than a conversation between people. This latency improves every month with the latest models, but it is still the main focus for very fast conversations.
To these three layers the layer is added integration: the AI control unit must be able to read and write on calendar, CRM, ticket system, business calendar. It is here that you play the true value — and the true work of configuration.
How much an AI control unit costs: the honest comparison
The costs depend on three voices: telephone infrastructure, AI model API, and setup/maintenance. Here is a comparison with the main alternatives, based on market benchmark 2025-2026 (sources: Retell AI pricing, Vapi.ai, ElevenLabs, public prices of the main providers):
| Solution | Fixed/month cost | Variable cost | Initial setup | Availability |
|---|---|---|---|---|
| Receptionist part-time | €800–1.500 | €0 (included) | Selection + training | Working time |
| Traditional cloud control unit | €30–80/user | €0 (flat) | €0–500 configuration | 24/7 (voicemail) |
| AI control unit (voicebot) | €100–400 (platform) | €0.05–0.15/minute | €1,500–5,000 setup | 24/7 (conversational) |
With 500 minutes of inbound calls per month (about 25 calls/day from ~20 minutes each), the operating cost of the AI control unit is €125–175/month of API, plus the platform. Difficult that exceeds the cost of a part-time receptionist already over 200 minutes/month of operated traffic.
The setup cost (€1.500–5.000) shocked over 12 months adds €125–416/month. This is the critical point: the AI control unit is only economically convenient if the volume of calls is sufficient to cover the amortization of the setup — and if the flows are stable enough not to require continuous rewriting.
When it makes sense (and when it doesn't suit)
This is the part that AI control providers tend to omit. It's not a solution for everyone.
Makes sense when:
- Volume ≥ 15-20 inbound/day calls. Under this threshold, the cost of setup is not damaging in reasonable times.
- Flows are standardizable. Booking appointments, qualifying lead with 4-6 fixed questions, sorting by department, answer to FAQ recurring. If every call is a case in itself, Agent AI will do it.
- You already have a CRM in use. The value of the voice agent multiplies when the data of each call ends automatically in your Pipedrive or HubSpot — qualified lead, notes, next step. Without built-in CRM, you lose half the advantage.
- Off-clock calls are a real problem. If you lose leads because they call at 19:00 p.m. and find voicemail, an AI control unit solves the problem at the root.
It doesn't matter when:
- Calls require negotiation or personal relationship. An AI agent can qualify and sort out, can't handle a complex business negotiation or respond to an angry client who wants to "talk with a human being immediately".
- Your industry has stringent regulatory constraints. Health, legal, finance: Voice data processing and recordings have GDPR implications that require a specific analysis before implementing.
- The volume is low and the processes are irregular. Under the 10 calls/day with always different flows, the cost/benefit ratio does not hold. A traditional cloud control unit + part-time secretary is more efficient.
How to integrate the AI control unit with CRM and business systems
Integration is real work — and often what determines whether the project succeeds or remains a nice prototype that then no one uses.
The typical flow that we configure in the implementations we follow through our service of voice agents AI:
- Receiving the call → the company number (or a dedicated number) routed to Vapi, Retell or similar platform
- The agent responds → shows up, understands the reason for the call (first open question)
- Qualification → 3-5 questions defined in the prompt: sector, number of employees, main problem, urgency
- Action → book slot in the calendar (Calendly, Cal.com, Google Calendar) or create the CRM record with all data collected
- SMS/email confirmation → via Make.com or Zapier, automatically sends to the caller the recap and the appointment link
- Internal notification → the commercial receives a Slack/email notification with the summary of the call and lead profile
Integration with CRM as Pipedrive takes place via Make.com (webhook → CRM module), which does not require custom development. For environments with more specific requirements (SSO authentication, legacy CRM, ERP systems) you work via direct API — this is the case when implementation times are extended to 8-12 weeks.
Why "AI central" and "AI voice agent" are not the same thing
A useful clarification if you are looking for suppliers: terms overlap but do not coincide.
Voice Agent AI is the technical term for the system that manages the conversation — the software with STT + LLM + TTS described above. It can be used both inbound (some call the company) and outbound (the agent calls a contact list).
AI control unit is the application of the voice agent to the case of inbound use: respond to the calls arriving, sorting, qualifying, booking. It's a subset of voice agents' perimeter.
In our article on status of AI vocal agents in 2026 find a wider overview — including outbound usage cases and main platforms. This article focuses only on inbound: the control unit.
The factor no one tells you: the prompt is the product
The quality of an AI control unit depends almost entirely on the quality of the prompt — the instructions that define how the agent must behave, what he can say, what he cannot say, how he handles the limit cases.
A mediocre prompt produces an agent that responds mechanically, loses the thread of the conversation as soon as the caller says something unexpected, or — worse — invents answers on prices and timing that then do not correspond to reality.
A good prompt requires:
- Precise corporate documentation: who you are, what you sell, to whom, with what prices and times (public ones).
- Conversational handwritten flows for the main cases, with the most common variants.
- Explicit management of limit cases: the angry caller, the out-of-mouth question, the "I want to talk to a person."
- Real Traffic Test with iterative review in the first 2-4 weeks.
This is why the setup has a cost. And because companies that buy "the product" without the configuration service often get disappointing results. Leaving AI automation that really work are those thought starting from the process, not from the software.
FAQ: AI switchboard for companies
Does an AI control unit really understand Italian, including accents and informality?
Yes. Modern speech-to-text engines (OpenAI's Whisper, Google Speech-to-Text v2) recognize Italian with an error rate of less than 5% even with marked regional accents. The understanding of natural language is managed by an LLM (GPT-4o, Claude) that understands ambiguous phrases, subject changes and implicit responses. The interlocutor perceives a fluent conversation, not an IVR menu.
How long does it take to activate an AI control unit?
It depends on the complexity of flows. A standard AI control unit — inbound response, qualification with 3-5 questions, sorting to CRM — configures in 2-4 weeks: 1 week to define the streams and write the prompt, 1 week to integrate CRM and agenda, 1-2 weeks test with real traffic. More articulated flows (e.g. multilanguage, multiple departments, complex conditional logic) take 6-10 weeks.
How do I integrate the AI control unit with CRM?
Using API or webhook. After each call, the AI switchboard writes in CRM (Pipedrive, HubSpot, Salesforce, Zoho etc.) the contact, the qualification obtained, the summary transcription and the following action (e.g. appointment booked, scheduled callback, lead passed to a commercial). Integration can be done with Make.com or Zapier without writing code, or via API for more structured environments.
What is the difference between AI control unit and traditional virtual control unit?
The traditional virtual power station manages telephony (addition, voicemail, geographic number), but does not understand what the caller says: it is limited to pre-registered menus (IVR) and transfers. The AI switchboard adds a conversational layer: listen, understand, respond in natural language, qualify, book, update CRM — without a human in the middle. The virtual power station is telephone infrastructure; the AI control unit is smarter telephone infrastructure.
When should I adopt an AI control unit?
It doesn't matter when: the inbound call volume is under 10-15 per day (the setup cost is not shocked); calls require complex negotiation or consolidated personal relationships; processes are too irregular to be documented in a stream; your industry has regulatory constraints on recordings and speech data processing that require specific procedures. In those cases, a traditional cloud control unit with a part-time secretary is more efficient.
Read more: AI voice agents for companies: where we are in 2026 — the complete view of outbound and inbound usage cases. Or How to automate business with AI in 2026 for the operating framework on how to identify processes to be automated first.