The short answer
What is actually required, and which commonly listed items are conditional.
AI calls do not need a PC, headset or MicroSIP per call
The AI session runs on the gateway server. Buy operator stations only for supervisors, transfers, callbacks or a human queue.
Numbers, extensions and SIP channels are different
A DID is the public number. An extension identifies an internal endpoint. A SIP channel is one simultaneous external call.
No GPU is required
Gemini performs inference in the cloud. Spend on reliable CPU, RAM, SSD, networking, UPS, backup and operations—not a gaming GPU.
For the audited traffic level, the measured Gemini bill was about 3.84 million UZS/month. A realistic greenfield 10-channel, 5-human-seat installation is roughly 77–162 million UZS in hardware and commissioning, plus 80–250 million UZS if the AI gateway, CRM tools, QA and production hardening must be engineered from scratch. Existing servers, PCs and the existing YangiAsrDialer code can reduce the initial cost substantially.
Reference architecture
One controlled path for telephony; one auditable path for data, leads and recordings.
Complete requirements list
A practical bill of materials for ten concurrent calls and up to five human seats.
| Priority | Component | Reference quantity | Minimum / target specification | Budget allowance | Decision note |
|---|---|---|---|---|---|
| Required | UZTELECOM business SIP trunk | 10 channels | PJSIP/SIP peer; G.711 A-law; DTMF RFC4733; defined RTP range; inbound/outbound CLI; CDR and provider SLA. | Carrier quote | Channel count sets the simultaneous public-call ceiling. |
| Required | Public DID/IP phone numbers | 1–3+ | One main number; optional campaign/branch DIDs; documented number format and inbound routing. | Carrier quote | A number does not automatically add capacity. |
| Required | Primary business connectivity | 1 | 100 Mbps+; stable routing to carrier and Google; static/public or private routed IP as contracted; low jitter and packet loss. | 0.5–1.5m UZS/mo | Confirm whether the SIP trunk arrives over a private interface or public Internet. |
| Recommended | Backup Internet / carrier path | 1 | Different provider and physical route; 50 Mbps+; tested failover that does not break SIP source-IP rules. | 0.3–1.0m UZS/mo | Internet failover is not SIP failover unless the carrier supports the alternate IP. |
| Required | Router / firewall | 1 | Dual-WAN, VLAN, QoS, VPN, stateful firewall, logs; RB5009-class or equivalent. | 3–6m UZS | Never expose FreePBX GUI, AMI, ARI or SSH broadly to the Internet. |
| Recommended | Managed switch + cabling | 1 | 16/24 Gigabit ports; voice/server VLANs; PoE only if buying IP phones; labeled Cat6 runs. | 2–7m UZS | MicroSIP on PCs does not require a PoE switch. |
| Required | PBX / AI gateway server | 1 | 8 CPU cores, 32 GB RAM, 2 × enterprise SSD in mirror, dual NIC, hardware health reporting, 3-year warranty. | 18–35m UZS | No GPU. Separate PBX and gateway VMs or hosts are preferable for failure isolation. |
| Recommended | Backup / recording storage | 1 | 4 TB usable with snapshots; encrypted backup; restore test; retention automation. | 5–15m UZS | Mixed mono WAV is roughly 58 MB per connected hour; confirm the actual codec and recording format. |
| Required | UPS | 1–2 | At least 1500 VA for the small stack; online UPS preferred for server + network; graceful shutdown testing. | 1.2–6m UZS | Local examples ↗ |
| Required | FreePBX + Asterisk | 1 system | Greenfield: current supported FreePBX 17/Debian 12 and supported Asterisk release; backups, firewall and update policy. | Core: free | Official install path ↗ |
| Required | AI voice gateway | 1 service pool | Audio bridge, resampling, session lifecycle, reconnect deadline, tool allowlist, concurrency cap, stable call ID, metering and fallback. | Custom engineering | The existing YangiAsrDialer gateway is relevant, but production drift must be resolved before reuse. |
| Required | Gemini paid API project | 1+ | Billing-enabled project, production quota, spend cap, alerts, separate test/prod keys, restricted secret access. | Usage based | Official pricing ↗ |
| Required | CRM + consent/opt-out registry | 1 | Lead ownership, source, consent evidence, do-not-call state, idempotent updates, retry/reconciliation queue. | Existing or custom | No outbound promotional call should bypass the consent check. |
| Recommended | Monitoring and alerting | 1 stack | SIP registration/OPTIONS, call completion, silent-call detector, worker health, disk, queue depth, Gemini spend, CRM failures. | 0.2–1.0m UZS/mo | Alerts need a named on-call owner and documented action. |
| Conditional | Operator PC | Per human seat | Core i5/Ryzen 5 class, 16 GB RAM, 512 GB SSD, wired Ethernet. Existing office PCs often suffice. | 5.5–9m UZS | 7.03m example ↗ |
| Conditional | USB headset | Per human seat | Wired USB, boom microphone, replaceable ear pads, acoustic shock protection; noise rejection preferred. | 0.5–1.8m UZS | 771k example ↗ |
| Conditional | MicroSIP softphone | Per human PC | Individual PJSIP extension, strong random secret, correct codec/DTMF, auto-answer disabled unless operationally required. | Free / GPLv2 | Licence source ↗ |
| Alternative | IP desk phone | Per human seat | Use instead of MicroSIP + USB headset where a dedicated appliance is preferred. | 1–3m UZS | Do not buy both for every seat unless there is a stated fallback requirement. |
What else is needed beyond the shopping list?
- Carrier technical sheet: peer IPs, SIP/RTP ports, codecs, DTMF, number format, CLI policy, failover, rate limits and support SLA.
- Approved knowledge source with owner, validity dates and facts the AI is prohibited from guessing.
- Human-transfer route, queue hours, fallback greeting and after-call follow-up ownership.
- Call-ID correlation across Asterisk, gateway, CRM, recording, QA and billing.
- Staging environment, test DIDs, synthetic audio set, native-Uzbek evaluation and canary rollout.
- Retention schedule, access controls, backup/restore procedure and incident response.
What is usually unnecessary?
- A GPU for Gemini Live inference.
- A public IP for every MicroSIP or IP phone.
- FXO/FXS or E1 cards when the carrier supplies a pure SIP trunk.
- A computer, headset or softphone for every simultaneous AI call.
- Commercial FreePBX modules unless a named feature and support need justifies them.
- Exposing PBX management or SIP endpoints directly to the public Internet.
Capacity and performance sizing
Size from simultaneous calls and connected minutes—not from the number of leads in the CRM.
| Scale | SIP channels | AI concurrency cap | Server target | Internet target | Human seats | Design note |
|---|---|---|---|---|---|---|
| Pilot | 2–4 | 2–3 | 4 cores / 16 GB / mirrored SSD | 50 Mbps | 1–2 | Staff and opted-in canary traffic only. |
| Small production | 10 | 8–10 | 8 cores / 32 GB / mirrored SSD | 100 Mbps + backup | 0–5 | Reference design in this document. |
| Growth | 20–30 | 18–28 | Separate PBX/gateway; 12–16 cores total; 64 GB | 200 Mbps + backup | 5–15 | Introduce HA, second carrier path and load/failure testing. |
| Large / regulated | 50+ | Validated | Redundant nodes, replicated state, SBC, dedicated storage | Dual diverse links | As required | Carrier, quota, privacy and DR design become formal workstreams. |
Cost model
Measured AI usage is separated from carrier quotes, hardware allowances and operational staffing.
Gemini 3.1 Flash Live lists audio input at $0.005/min and audio output at $0.018/min. That is a simple audio-only ceiling of $0.023 when both directions each consume a full modality minute, but Live also bills the active text/audio context again on later turns. The audited production average was materially higher: $0.1135 per 63.3-second event, or about $0.1076 per connected minute.
Interactive monthly operating-cost estimator
* Planning placeholder: replace with the signed UZTELECOM outbound tariff. ** Historical reference only: 16,840 UZS/channel/month is from UZTELECOM’s public price sheet effective 2018 for a 10-channel group; it is not confirmed as the 2026 contract rate.
Measured production baseline
Read-only 30-day ledger audited 11 August 2026
Gemini volume scenarios
Scaled at the historical $0.1135 average/event
Indicative one-time investment
| One-time item | Low | High | Included assumption | Evidence quality |
|---|---|---|---|---|
| PBX / AI server | 18m | 35m | 8 cores, 32 GB, mirrored enterprise SSD, dual NIC, warranty | Budget quote |
| Backup / recording storage | 5m | 15m | 4 TB usable, snapshots and backup media | Budget quote |
| Firewall, switch and cabling | 5m | 13m | Dual-WAN router, managed switch, Cat6, patching | Budget quote |
| UPS / power protection | 1.2m | 6m | Line-interactive floor to small online-UPS configuration | Market range |
| Five human seats | 32.5m | 52.5m | Five PCs and five USB headsets; omit if AI-only or reuse existing PCs | Market range |
| Installation, testing and documentation | 15m | 40m | Network/PBX build, carrier testing, backup, monitoring and handover | Service quote |
| Hardware + commissioning subtotal | 76.7m | 161.5m | Greenfield 10-channel / 5-seat setup | Indicative |
| Custom AI/CRM/QA engineering | 80m | 250m | Only if building the gateway, tools, dashboards, QA and hardening from scratch | Scope quote |
| Greenfield total | 156.7m | 411.5m | Before carrier deposits, VAT differences and ongoing operations | Indicative |
All equipment and engineering values are planning allowances, not offers. Obtain at least two written supplier quotes with warranty, VAT, stock status and delivery included.
UZTELECOM order checklist
Ask for one signed commercial annex and one technical annex. Do not accept an informal channel count.
Commercial questions
- One-time trunk connection charge and deposit.
- Monthly price per channel at 10, 20 and 30 channels.
- Monthly DID/number fee and cost of additional numbers.
- Outbound local, mobile, intercity and international per-minute rates.
- Billing increment, minimum charge and whether unanswered calls are billed.
- Caller-ID presentation rights for each DID and campaign.
- SLA, fault escalation contacts, service hours and restoration target.
Technical questions
- Carrier signaling and media IPs; customer peer IP(s); static routes and interface.
- SIP transport/port, RTP port range, G.711 A-law, DTMF and fax policy.
- Authentication method: IP peer, registration or mutual TLS.
- Inbound number format and outbound presentation format (+998/E.164 or national).
- OPTIONS/keepalive interval, session timers and maximum CPS.
- Failover to a second customer IP and carrier SBC redundancy.
- Test plan: 100 inbound attempts, parallel calls, cause codes and packet capture window.
The only detailed public SIP-trunk price sheet found on the official UZTELECOM domain is effective from 1 February 2018. It lists a 93,600 UZS one-time access charge and 16,840 UZS per channel/month for a ten-channel group, with lower unit rates at larger groups. Treat this only as a historical reference. The signed 2026 carrier quote controls the budget.
Implementation sequence
Build a testable system in layers. Carrier, PBX, AI and CRM each get their own acceptance gate.
Contract & policy
Approve purpose, consent evidence, recording notice, retention, DIDs, channel count, call windows and named owners. Obtain carrier annexes and Gemini paid project.
Network & power
Install VLANs, firewall, static routes, QoS, UPS, NTP and management VPN. Prove primary and backup Internet behavior before telephony.
PBX foundation
Install supported FreePBX/Asterisk, PJSIP trunk, DIDs, inbound routes, queues, recordings, CDR, backups and least-privilege management.
Human endpoints
Create individual extensions, configure MicroSIP or IP phones, test transfers/queues, and prohibit shared secrets. Confirm wired audio and headset levels.
AI gateway
Bridge call audio to Gemini, enforce concurrency, deadlines and cleanup, meter every session, correlate one stable call ID and add a safe fallback route.
CRM & facts
Connect consent/opt-out checks, lead writes and reconciliation. Generate scenario knowledge from an owner-approved, versioned fact manifest.
Failure testing
Test silence, packet loss, Gemini disconnect, carrier timeout, CRM failure, disk full, worker crash, long calls, repeated calls and human transfer.
Canary & scale
Run staff calls, then an opted-in 5% canary. Expand only when quality, cost, privacy, carrier delivery and rollback gates remain green.
FreePBX currently directs new installations to FreePBX 17 on Debian 12, and its March 2026 guidance calls Asterisk 22 the preferred current long-running choice. The audited production system used Asterisk 18. Migrate in parallel with rollback; do not in-place upgrade the active PBX simply to match this document.
Security, privacy and legal gates
These are launch requirements, not documentation to add after go-live.
Uzbekistan’s Advertising Law Article 30 allows advertising through telecommunications services only with the subscriber’s prior consent. Store evidence and enforce opt-out centrally.
Open with the organization, the AI identity and a clear recording/data notice approved by counsel. Provide a human route where the use case requires it.
Document processing purpose, fields, owner/operator, access, retention, deletion, subject requests and any cross-border transfer basis.
The Personal Data Law was amended in 2026. Certain categories must stay in Uzbekistan; other data may be stored abroad only under stated conditions. Obtain a written legal classification for call audio/transcripts sent to Gemini.
Use a billing-enabled paid project. Google says paid prompts/responses are not used to improve products; ZDR requires approval and avoiding session resumption, which creates a reliability trade-off.
Allow SIP/RTP only from carrier peers, manage via VPN, restrict API keys by project/service account, remove phone numbers from filenames and use per-service filesystem permissions.
Set separate periods for recordings, transcripts, CDR, CRM leads, debug logs and QA files. Backups must inherit expiry and access restrictions.
Keep secrets out of code and operator PCs; rotate them; log access and changes; maintain reproducible deployment source, version hashes and rollback packages.
Prices, dates, admission facts and promises need an owner, source, validity window and “do not quote” state. The model must say it does not know instead of inventing.
This section is an engineering checklist, not legal advice. Uzbekistan law and the exact controller/operator role must be reviewed for the organization’s actual flows.
Go-live acceptance criteria
No item below should be “assumed working.” Capture evidence and an owner for each result.
| Area | Acceptance test | Pass criterion | Evidence |
|---|---|---|---|
| Carrier delivery | 100 timed inbound attempts across each DID and busy-hour route. | No unexplained missing INVITEs; every failure has carrier/Asterisk cause evidence. | Caller log + pcap + CDR |
| Concurrency | 1, 5 and 10 simultaneous calls; hold for 60 minutes; include transfers. | No rejected call below the cap, silent audio, CPU runaway or leaked channel/session. | Asterisk channels + gateway metrics |
| Audio quality | Native-Uzbek callers, male/female voices, echo, noise, interruption and digits/names. | Intelligible both directions; no repeated missing audio; approved pronunciation/number accuracy. | Blind review scorecard |
| Latency | Measure answer-to-greeting, end-of-speech to first audio, and tool-call waits. | Targets approved by operations; p95 regressions block release. | Per-call timestamps |
| Facts | Adversarial questions for price, date, branch, discount and unknown data. | No invented value; verified fact hash recorded; unknown facts fail closed. | Golden test set |
| CRM / leads | Success, timeout, malformed response, duplicate callback and retry. | Deterministic final state; no “success” on error; reconciliation clears every pending item. | State ledger + CRM record |
| Human handoff | Available, busy, closed and abandoned transfer paths. | Caller always receives a clear outcome; context reaches the human without over-sharing. | Queue/CDR + agent review |
| Cost | Compare per-modality API usage with the local meter for a canary batch. | Difference within 5% or explained; project spend cap and alerts proven. | API bill + local ledger |
| Privacy | Consent denied, deletion request, unauthorized lookup and access review. | Call blocked/limited as designed; least-privilege access; request handled within approved procedure. | Audit log + ticket |
| Recovery | Gateway crash, Gemini disconnect, PBX restart in staging, WAN failure and restore. | Bounded fallback; no 35-hour wedge; tested restore and written rollback time. | Incident drill report |
Sources and verification notes
Primary sources are used for product pricing, software status, exchange rate and legal text. Local equipment ranges remain estimates.
- Gemini Developer API pricing3.1 Flash Live text/audio token and minute equivalents.
- Gemini Live best practicesCompounding active-context billing, compression, audio streaming and session guidance.
- Gemini 3.1 Flash Live modelModel ID, capabilities and migration notes.
- Gemini API rate limitsProject-level quotas; confirm active limits in AI Studio.
- Gemini zero-data-retention guidancePaid data use, Live session resumption and ZDR trade-offs.
- FreePBX official siteCore platform is free; commercial add-ons and support are optional.
- FreePBX getting startedCurrent greenfield install path and built-in feature scope.
- MicroSIP source and licenceGPLv2 softphone; no per-seat licence cost.
- UZTELECOM SIP-trunk public price sheetEffective 1 Feb 2018; historical reference only, not a current quote.
- UZTELECOM official siteCorporate support number 1155 and current service entry point.
- Central Bank USD rate1 USD = 11,889.95 UZS on 12 Aug 2026.
- Uzbekistan Advertising LawArticle 30: prior subscriber consent for advertising through telecom services.
- Uzbekistan Personal Data LawCurrent text including 2026 changes to storage and cross-border conditions.
- Internal production auditMeasured 30-day Gemini ledger and production risks; historical snapshot with known live/local drift.
Verified: official vendor/law/exchange source or internal measured ledger. Historical: an official source whose effective date is old. Estimate: planning allowance that needs a current written quote. Do not use this document as a purchase order.