The demo shows how the agent sounds. This page is about what stands behind the call: what the operator is built from, the ways it can be delivered, what has to be decided before a price, and what each part costs.
There is no single "operator model" that can be handed over as a file. It is a chain of parts, and each of them works in real time.
Takes the call and carries the audio both ways: a SIP trunk, Asterisk or your PBX.
A call to your number arrives through the carrier and lands in Asterisk, which streams the audio to the agent.
Turns the caller's voice into text while they are still talking.
"Hi, I have bedbugs in a two-bed flat" is text before the caller has finished the sentence.
A language model decides what to say, from the script, your prices and your rules.
It knows the price list, asks one question at a time, never offers a discount that does not exist, and steers back to the booking.
Speaks the answer in the operator's voice, with pauses, breaths and intonation.
The first words start playing while the model is still writing the end of the sentence.
Writes the outcome where your team works with it: CRM, calendar, spreadsheet, messenger.
After the call a record appears in the CRM: name, phone, request, visit date, price quoted, outcome.
On top of the chain sits conversation control: knowing the caller has finished, not stopping at an "uh-huh", going quiet when interrupted, filling the pause when an answer is slow. The demo has all of it tuned, and that is what separates a conversation from an answering machine.
People expect an answer within about a second of pausing. This is how that second split on a real call to our demo.
Measured on a demo call on 9 October 2026, cloud models. For an on-premise install this split has to be measured again on your models and your hardware: it is what decides whether the conversation feels alive.
The question "cloud or our own" settles almost everything else: timeline, budget, hardware, and how close the voice will be to the demo.
As in the demo: models at the providers, you pay per minute of conversation.
When launch speed and voice quality matter, and keeping recordings at a provider is acceptable.
Telephony, conversation control, integrations and data stay with you; models are called over an API.
When data has to stay with you, but calling external models from inside the perimeter is allowed.
The whole chain on your servers, with nothing going out.
When nothing may leave, and there is budget for the hardware and time for the testing.
Price and timeline come after the specification. Without one, a thousand and one wishes surface along the way. These are the questions that need answers, with examples.
Where does it run?
Decides the delivery option, the hardware and half the budget.
"Everything on our servers, nothing goes out" or "with us, but cloud models over an API are fine".
Voice only, or a full conversation?
Speaking a given text and an operator that listens, understands and leads a conversation are systems of very different size.
"Read out an order status by its number" or "take a request and answer questions".
What does the operator do?
Every scenario is a prompt, knowledge, checks and tests. Without examples there is no way to size it.
"Take a pest-control request, quote from the price list, book a time, and hand over to a person if the caller argues".
How many simultaneous calls?
Peak load decides the hardware, or the cloud bill.
"Up to 10 on a normal day, 30 at peak, 1-2 at night".
What do we connect to?
The boundary of the work: what you connect, and what we build.
"We run Asterisk and give you a SIP trunk; we do the CRM integration ourselves against your API".
What hardware is there?
For on-premise: we either fit models to your hardware, or hardware to the quality you need.
"A rack in a data centre, no GPUs, a purchase budget to be agreed".
How is the result accepted?
Without acceptance criteria every change becomes an argument.
"Ten agreed scenarios pass on our real line, median response no longer than 1.5 seconds, interruption works".
The word "operator" hides very different amounts of work. Two examples on the same line.
Finds out who is calling and why, answers routine questions, books or hands over to a person with the context already collected.
AgentBrightHome Pest Control, how can I help?
CallerWe have cockroaches, how much is it?
AgentGot it. Is it a flat? How many bedrooms?
CallerTwo.
AgentFor a two-bed it is a hundred and fifty pounds. Shall I book you in for tomorrow?
One or two scenarios, the price list, handover rules. Integration: a record in the CRM.
Handles objections, offers alternatives, checks stock and slots, places the order in your systems and confirms it.
CallerThat is expensive. Others charge less.
AgentI understand. We guarantee it for a year and the follow-up visit is free. If it is a light infestation, there is a cheaper option.
CallerFine, Saturday then.
AgentRight, let me check the technicians... Saturday, eleven in the morning is free. Booking it now.
Objection scripts, live access to schedules and stock, order creation over an API, checks before confirming. Noticeably more testing.
If the whole chain has to run with you, it needs compute. Options depend on the load and the model. The baseline for 20-30 simultaneous conversations is option 2, the one we would begin testing from.
Hardware options to buy
| Configuration | Simultaneous calls* | Whole server | Fits when |
|---|---|---|---|
| 1. Pilot: 1 × L40S 48 GB or 1 × RTX PRO 6000 96 GB | 5-10 / 10-20 | $20-35k | A pilot, one department, a proof of concept |
| 2. Baseline: 2 × L40S, 32 cores, 256 GB, 2 TB NVMe | 20-30 | $35-50k | Production with a compact 7-14B parameter model |
| 3. Baseline with headroom: 2 × RTX PRO 6000 96 GB | 30-50 | $45-60k | A larger 30-70B model, or room to grow in the same box |
| 4. Large: 4 × L40S or 4 × H200 NVL | 50-100+ | $55-75k / $160-200k | A large model, 50+ calls, several sites on one cluster |
| 5. Budget, used: A100 80 GB | 15-25 / 2 GPU | $8-12k / GPU | A tight budget; no FP8, short or no warranty |
Rental: 2 × L40S by provider
| Provider | Per GPU-hour | 2 GPUs a month | Note |
|---|---|---|---|
| Vast.ai | ≥ $0.54 | ≈ $790 | Marketplace of community hosts, availability varies |
| Runpod | $0.79 / $1.09 | ≈ $1,150 / $1,590 | Community / Secure Cloud |
| Scaleway | €1.47 | ≈ €2,150 | A ready 2-GPU configuration, Paris only |
| Nebius | ≥ $1.55 | ≈ $2,260 | Configurations sized by vCPU |
| OVHcloud | $1.80 | ≈ $2,630 | A ready 2-GPU configuration, Europe |
| AWS (g6e) | $1.86 | ≈ $2,720 | No 2-GPU size: two 1-GPU instances, or one 4-GPU (≈ $7,660) |
| CoreWeave | $2.25 | ≈ $13,140 | Only 8-GPU nodes |
| Immers.cloud | — | ≈ 321 000 ₽ (2 × A100) | Data in Russia. No L40S; the nearest with a public price is A100 80 GB |
| Selectel, Yandex Cloud, Cloud.ru, Timeweb, MTS Cloud | — | — | Data in Russia. No public L40S; RTX 6000 Ada, A6000, A100, H100 on request |
| Lambda, Google Cloud, Azure, Hetzner | — | — | No L40S; Google, Azure and Hetzner offer RTX PRO 6000 instead |
* Call counts are assumptions, not measurements: they depend on the model size, recognition and synthesis, and are settled by a load test. Prices checked on 9 October 2026, excluding VAT; hardware and memory got dearer all through 2026. In Russia GPU prices run 1.5-2× higher and are usually made to order, 40-60 working days.
A voice that sounds clean in a browser crosses a mobile network with compression. If a call goes through two mobile legs, the audio degrades twice. We tested this on a model of the line: AMR-NB on each leg and G.711 between them.
A line model is an approximation. Quality is settled by calls through your real chain, and that is part of acceptance.
Hear it in the demo on "Two legs" →The criteria are fixed before work starts, so that "done" means the same thing to both sides.
If we take on the whole virtual call centre: Asterisk, telephony, the AI operator, integration and launch.
An order of magnitude, not a price list. The exact figure comes once the scope is fixed: an operator that qualifies and hands over and one that sells and places orders cost differently.
The questions that come up when people look at the table above and want to save. In short: one or two calls in a test setting run on a lot of things. Twenty or thirty simultaneous conversations answered within a second do not.
It has 24-32 GB of memory. One card has to hold the response model, recognition and synthesis, plus memory for every conversation in progress. For 20-30 parallel calls that is not enough, and answers start queuing. NVIDIA's driver licence also forbids GeForce cards in data centres, they have no error-correcting (ECC) memory, and their cooling is not built for 24/7 in a server rack. Fine for a one- or two-call prototype, not for production.
A language model on a CPU takes seconds per sentence. Real-time recognition and synthesis for dozens of streams hit the CPU too. The response budget is about a second, half of it for the model. CPUs miss it even with a single call.
For a pilot with a few calls, possibly - it is option 1 from the table, only weaker. The trouble is peaks: when five people finish talking at once, their answers queue, and a second becomes three to five. The T4 also lacks modern compute formats and is noticeably slower. Only worth it if the peak is a couple of calls.
It runs a single model well, and is handy for development. But it is built for one stream: it serves dozens of parallel conversations poorly, because it cannot batch requests the way server GPUs do. And there is no server management, redundancy or ordinary rack operation.
A small model fits, but holds the script worse: it mixes up prices, drifts off topic, agrees to discounts that do not exist, and handles languages worse. We saw this in tests even between cloud models of one tier: one answered briefly and by the rules, another of the same speed slid into monologues. Model size is a trade-off between conversation quality and hardware, and it has to be tested on your scenarios.
Calls arrive in bursts: in the morning, after a mailing, at lunch. Latency does not rise gently; it jumps as soon as requests outnumber what the card can process. A server sized for the average answers in three seconds at the peak, which is exactly when most people call.
Yes, and that is the right path. Rent and load-test first, then option 1 for a pilot. The architecture lets recognition, synthesis and the response model run on separate servers, so growing means adding hardware, not rebuilding the system.
Hear the agent first. If this order of budget works for you, the next step is to describe the first delivery and fix what has to work at acceptance.
What it costs: from $99 a month · Which outbound calls does the agent make?
Comparing it with a live answering service for your trade? Answering service by industry.
