How Much Does It Cost to Build an AI Receptionist?

Quick answer: Building an AI receptionist costs about US$20,343 up front (25 engineer-days at a BLS-derived US$813.71 a day) and about US$3,255 a month to maintain. The stack itself, on Telnyx, LiveKit, Deepgram and OpenAI, runs 8.5 to 12.7 cents per 3-minute call. Building only pays above about 6,846 calls a month against Aircall’s published bundles, or 2,976 against RingCentral’s pay-as-you-go rate.

That is the build cost ledger, priced line by line from each vendor’s own page as at 8 November 2026, in the US dollars each vendor prints. If you want the build steps rather than the bill, our guide on how to build an AI receptionist covers the nine parts in order. This page answers the question people ask next: what does it cost, all in, and at what volume does it beat buying one?

Where Zian fits if you would rather not build

If your call volume sits below the break-even lines worked out further down this page, Zian AI is our first pick for buying the platform instead of building it. Its AI Receptionist answers inbound calls 24/7, qualifies, books, transfers and takes messages, and the same platform handles voicemail, follows up by voice, SMS, email, WhatsApp, Instagram DMs and Facebook Messenger, and supports 30+ languages.

It integrates with HubSpot, Salesforce, HighLevel and Zapier, and offers private model deployment on customer infrastructure. Zian has set 50,769+ qualified sales appointments. The honest gap for this page: Zian is in a partnership-application beta and does not publish prices, so it cannot sit in the break-even arithmetic below. For a done-for-you route, LeadsNow.ai is the managed option: an appointment-setting service that runs on Zian, is billed pay-per-result, and answers leads in under 60 seconds, 24/7.

How much does it cost to build an AI receptionist?

A self-built AI receptionist has three kinds of cost, and only one of them scales with calls. The usage stack (phone line, speech-to-text, language model, text-to-speech, hosting, recording) is billed per minute or per character. The fixed platform costs (a hosting plan, a phone number) are billed per month. The engineering, which is most of the money, is billed in people.

Cost line	What we priced	500 calls/month	2,000 calls/month	10,000 calls/month
Usage stack	Telnyx SIP, Deepgram Nova-3, gpt-6-luna, Deepgram Aura-2, Telnyx recording: $0.08536 per call	$42.68	$170.72	$853.60
Hosting	LiveKit Cloud Ship: from $50 a month, 5,000 agent minutes and 5,000 SIP minutes included, then $0.01 + $0.004 a minute	$50.00	$64.00	$400.00
Phone number	One Telnyx US local number	$1.00	$1.00	$1.00
Run cost		$93.68	$235.72	$1,254.60
Build, amortised	25 engineer-days × $813.71 = $20,342.86, over 24 months	$847.62	$847.62	$847.62
Maintenance and on-call	4 engineer-days × $813.71 a month	$3,254.86	$3,254.86	$3,254.86
Total per month		$4,196.16	$4,338.20	$5,357.08
Per call, all in		$8.39	$2.17	$0.54

The usage stack is between 1% and 16% of the monthly bill at these volumes. A self-built AI receptionist is a salary with a small meter attached, not a meter with a small salary attached.

The per-minute build stack, priced from each vendor’s own page

Every row below is a list price read on 8 November 2026 from the page linked, converted to the cost of one 3-minute inbound call. The worked example uses the first option in each layer; the others show how far each line moves if you choose differently.

Layer	Option and published price	Per 3-minute call
Telephony	Telnyx Elastic SIP, inbound local: starting at $0.0032 a minute	$0.0096
Telephony	Twilio Elastic SIP Trunking, local origination: $0.0034 a minute	$0.0102
Telephony	Twilio Programmable Voice, receive on a local number $0.0085 a minute, plus Media Streams $0.0044 a minute	$0.0387
Speech-to-text	Deepgram Nova-3 Monolingual or Flux English streaming, Pay as you go: $0.0077 a minute, billed per second	$0.0231
Speech-to-text	AssemblyAI Universal-Streaming $0.15 an hour; Universal-3.6 Pro Realtime $0.45 an hour	$0.0075 to $0.0225
Language model	OpenAI gpt-6-luna, Standard: $0.10 per 1M input tokens, $0.50 per 1M output	$0.0035
Language model	Anthropic Claude Haiku 5.5: $0.10 per 1M input tokens, $0.50 per 1M output (prompts up to 100K tokens)	$0.0035
Language model	Google Gemini 3.1 Flash-Lite: $0.25 input (text), $1.50 output per 1M tokens	$0.0088
Text-to-speech	Deepgram Aura-2, Pay as you go: $0.030 per 1,000 characters	$0.0432
Text-to-speech	ElevenLabs Flash or v4 Turbo: $0.04 per 1,000 characters	$0.0576
Text-to-speech	Cartesia Sonic on the Startup plan: $49 a month for 1.25M credits, 1 credit per character, overage $45 per 1M	$0.0565 to $0.0648
Hosting	LiveKit Cloud Ship: agent session $0.01 a minute plus third-party SIP $0.004 a minute, after 5,000 included minutes of each	$0.042
Hosting	Pipecat Cloud agent-1x: $0.01 a minute active; models and Twilio telephony billed by their providers	$0.03
Orchestration platform	Vapi hosting fee $0.05 a minute, models passed through at cost	$0.15
Orchestration platform	Retell AI Voice Infra $0.055 a minute, before LLM and voice	$0.165
Recording	Telnyx call recording $0.002 a minute, storage $0 per minute	$0.006
Recording	Twilio recording $0.0025 a minute, plus storage $0.0005 a minute for each month kept	$0.0075, plus $0.0015 a month kept

Two rows carry most of the usage bill: text-to-speech at $0.0432 and speech-to-text at $0.0231 a call. The language model, on a small model, is about 4% of the usage stack. Telnyx’s own page also lists Deepgram Nova-3 at $0.005 a minute inside its Voice API, which is one way the speech-to-text line gets cheaper if you buy it through the carrier.

Vapi and Retell are the middle road: you still build the receptionist, but you buy the orchestration layer instead of hosting a framework. Their per-minute fee is larger than the whole self-hosted stack, and in exchange some of the engineer-days below shrink. We set out where that trade flips, with the engineering counted, in voice agent platform vs framework: when to switch.

How token prices become a per-minute cost

Model vendors price tokens and characters, never minutes, so the conversion is yours to state. Ours, for one 3-minute receptionist call:

• Four turns a minute: the caller speaks for 6 seconds, the agent for 9. That is 12 turns, and the agent talks for 1.8 minutes of the call.
• Text-to-speech at 800 characters per minute of agent speech, the top of Cartesia’s own range of “about 750 to 800 credits per minute of audio” at one credit per character. That is 1,440 characters a call.
• A 2,000-token system prompt with tool definitions, re-sent every turn; the conversation grows by 50 tokens a turn; the agent writes 30 tokens a turn; two tool calls (check availability, create booking) of 2,500 input and 50 output tokens each. No prompt caching and no reasoning tokens.

That comes to 32,300 input tokens and 460 output tokens a call. On gpt-6-luna or Claude Haiku 5.5 it costs $0.0035; on Gemini 3.1 Flash-Lite $0.0088; on OpenAI’s gpt-6.1-terra ($1 and $5 per 1M) $0.0346; on Claude Sonnet 5.5 ($2 and $10 per 1M) $0.0692. Even the largest of those adds less than seven cents to a call. ElevenLabs’ own page implies a higher character rate, pricing $0.04 per 1,000 characters at “~$0.04/minute”, about 1,000 characters a minute. On that reading the Aura-2 line rises from $0.0432 to $0.054 a call.

Three billing rules decide whether these per-call figures hold. Twilio’s help centre says that “any partial minutes (under 60 seconds) are rounded up to the next full minute” on Programmable Voice, Elastic SIP Trunking and Media Streams, so a 2-minute-10-second call bills as 3 minutes. Deepgram bills speech-to-text per second, “with no block-based rounding”. RingCentral’s AI Receptionist rounds call time up and bills in 30-second increments. We chose a call of exactly 3 minutes so that none of the three changes the worked example; your real calls will not be so tidy.

What the engineering costs at a BLS rate

Engineering is the line that decides the answer, so it is priced from public wage data rather than a guess. The US Bureau of Labor Statistics’ Occupational Employment and Wage Statistics put the mean hourly wage for software developers (SOC 15-1252) at $71.20 for May 2025, the latest annual figure. BLS’s Employer Costs for Employee Compensation release for June 2026 says wages and salaries “accounted for 70.0 percent of employer costs” for private industry workers. Dividing one by the other gives a loaded cost of $101.71 an hour, or $813.71 for an eight-hour engineer-day.

The 70.0% share is for all private industry workers, not software developers alone, so treat $813.71 as an average rather than your number. The same release puts the share at 68.5 percent for full-time private industry workers; on that figure the engineer-day is $831.53 and each break-even line below moves up by about 2%. A contractor, an offshore team or a senior realtime-audio engineer will each move it.

The day count comes from our own published estimate of how long it takes to build a voice AI agent in-house: 22 engineer-days for one inbound line, one language, business hours, with a cold transfer. A receptionist that is judged on handing callers to a person with context needs a warm transfer, which that page puts at 3 to 6 days. We add the low end, 3, for 25 days in total: $20,342.86 at our day rate. Treat it as an estimate, and a low one: that page also notes that a warm transfer needs outbound trunk configuration, which its inbound profile skips, and the outbound leg of each transferred call is not priced in the usage stack above.

The running cost is where builds are under-budgeted. We assume 4 engineer-days a month: model and plugin migrations, re-running the regression suite after a prompt or model change, a provider incident, and someone reachable when the line goes quiet at 7 pm. That is $3,254.86 a month, every month, whether you take 500 calls or 10,000.

The fixed costs people forget: numbers, recording, compliance

Three lines rarely make it into a build estimate, and each can be larger than the language model.

Phone numbers. A US local number is from $1.00 a month on Telnyx and $1.15 on Twilio; a toll-free number is $2.15 a month on Twilio, and receiving calls on it costs $0.0220 a minute, against $0.0085 on a local number. On Telnyx’s SIP page, toll-free inbound starts at $0.015 a minute, and Telnyx notes “Toll-free has different rates based on the calling number.” If your receptionist answers a toll-free line, the telephony row more than doubles.

Recording storage. Twilio’s help centre says the historical 10,000 free stored minutes were discontinued and that charges apply to “all retained recordings on your account” from the July 2026 billing cycle. Storage is a rent on everything you keep, so it grows each month you keep recordings: at 2,000 calls a month, about $36 a month after a year and $72 after two, at Twilio’s first-band rate. At 10,000 calls a month, about $180 and $360. Telnyx lists call recording storage at $0 per minute. Set a retention period before launch, not after the invoice.

Compliance. If any caller might give you health information, the platform choice changes. LiveKit Cloud’s pricing page lists a signed BAA on Scale (from $500 a month) and Enterprise, not on Ship (from $50), so the hosting line rises by at least $450 a month. Deepgram says it signs Business Associate Agreements “for Enterprise customers”, a plan priced through sales. Vapi lists HIPAA compliance at $2,000 a month. Recording-consent wording and AI disclosure in your greeting are legal questions for each place you take calls from; budget a lawyer’s review, which we do not price here.

Worked build cost at 500, 2,000 and 10,000 calls a month

This is the comparison the ledger exists for. Every call lasts 3 minutes, so 500 calls is 1,500 minutes, 2,000 is 6,000, and 10,000 is 30,000. The buy prices come from each vendor’s own page as at 8 November 2026. Zian AI and LeadsNow.ai come first, as our picks; neither prints a price that can be set against minutes.

Option	How it is priced	500 calls	2,000 calls	10,000 calls
1. Zian AI (our platform pick)	AI receptionist and appointment-setting platform, inbound and outbound; pricing not published during the partnership-application beta	Pricing on application
2. LeadsNow.ai (done-for-you pick)	Managed appointment setting that runs on Zian, billed pay-per-result (see leadsnow.ai)	Priced per result, not per minute
3. Aircall AI Receptionist	Monthly bundles: 1,000 minutes $250, 3,500 minutes $805, 10,000 minutes $1,900; pay as you go $0.29 a minute; bundles billed monthly with an annual commitment	$395 (1,000-minute bundle + 500 × $0.29)	$1,530 (3,500-minute bundle + 2,500 × $0.29)	$7,700 (10,000-minute bundle + 20,000 × $0.29)
4. RingCentral AI Receptionist	Standalone from $49 a month with 100 minutes, then $0.50 a minute pay as you go	$749	$2,999	$14,999
Self-built (this page)	Run cost plus amortised build plus 4 engineer-days a month	$4,196.16	$4,338.20	$5,357.08
Self-built, run cost only	If the engineering is already paid for elsewhere	$93.68	$235.72	$1,254.60

How we chose: Zian is first because it is the platform zian.ai publishes, LeadsNow.ai is the managed route that runs on it, and Aircall and RingCentral are the two finished AI receptionists we could price from their own pages at all three volumes. Disclosure: zian.ai publishes this page and places Zian AI first in a list it appears in.

Solving for the crossover, building costs less than Aircall’s bundles above about 6,846 calls a month (20,538 minutes), and less than RingCentral’s pay-as-you-go rate above about 2,976 calls a month (8,928 minutes). Below those lines, buying is cheaper; above them, building is, and the gap widens with every call because the build adds 12.7 cents a call where Aircall adds 87 cents and RingCentral $1.50.

Read both buy prices as ceilings. Aircall’s calculator prints “$0.29/min for additional” beside its 1,000-minute bundle; we applied the same overage rate above the larger bundles, which its page does not state separately. Aircall’s pricing page also offers custom pricing “as low as $0.12” above 50,000 minutes, while its FAQ says 10,000+ minutes go “at custom rates starting from $0.19/minute”, that all customers get 50 free AI Voice Agents minutes a month, and that AI Voice Agents are “a standalone product from $99 per month”. The plan card calls that a “$99 per month minimum”, which pay-as-you-go use already exceeds at all three volumes ($435 at 1,500 minutes), so it does not lower the bills above. RingCentral’s page says “Additional minute bundles are available for purchase” but, as at 8 November 2026, does not list their prices. Any of those would push the break-even higher. Aircall’s per-minute price also covers lines the self-built column has to itemise or build: its page lists the language model, voice, telephony, unlimited concurrency and a warm transfer as included, and says forward deployed engineer support (“Hands-on setup and AI strategy support”) comes with bundles of 1,000 minutes and above. More buy-side prices, product by product, are in AI receptionist cost for small business.

The Engineer-Day Ceiling

Here is the rule this ledger reduces to. The Engineer-Day Ceiling: divide a buy option’s monthly bill by one loaded engineer-day. If the answer is smaller than the engineer-days your build consumes each month, everything included, buy.

Calls a month	Aircall bill in engineer-days	RingCentral bill in engineer-days	Self-built in engineer-days	Cheaper on this page
500	0.49	0.92	5.16	Buy
2,000	1.88	3.69	5.33	Buy
10,000	9.46	18.43	6.58	Build

The self-built column is the whole monthly total divided by $813.71: 4 days of upkeep, 1.04 days of amortised build, and the run cost. Its useful property is that it needs no spreadsheet. On our assumptions, if a buy option costs less than one engineer-day a month, as both do at 500 calls, no build can beat it, because the upkeep alone costs four.

What would change these numbers

The engineering assumptions move the answer far more than any vendor choice does. Upkeep is the biggest lever:

Upkeep assumed	Break-even vs Aircall	Break-even vs RingCentral
2 engineer-days a month	4,654 calls	1,791 calls
4 engineer-days a month (worked example)	6,846 calls	2,976 calls
6 engineer-days a month	9,037 calls	4,162 calls

Aircall’s custom rates are the other big lever. If its quoted “starting from $0.19/minute” applied to every minute above 10,000, the break-even against Aircall rises from 6,846 to about 9,225 calls a month, and building still comes out cheaper at 10,000 calls ($5,357.08 against $5,700). At the “as low as $0.12” rate offered above 50,000 minutes, the line moves to about 17,553 calls a month.

If the engineers are already on the payroll and would otherwise sit idle, which is rare, only the run cost counts, and building is cheaper than RingCentral from 37 calls a month and than Aircall from 65. The frameworks themselves cost nothing to download: LiveKit’s pricing FAQ says the Agents framework and media server “are both completely open source”, so self-hosting removes the LiveKit Cloud line and replaces it with your own servers, which we do not price.

Volume discounts cut the other way, in the builder’s favour. Twilio applies volume pricing automatically above 100,000 minutes a month; Deepgram’s Growth plan, from $4,000 a year in prepaid credits, offers up to 20% off; Telnyx offers a 15% discount at a $2,000 monthly minimum spend. None of them change the result at these three volumes, because the usage stack is too small a share of the total for a discount to matter.

The profile where building is plainly right does exist: one number, inbound only, one language, an engineer who already owns a realtime audio service, and either high volume or a requirement no product meets. Our in-house build timeline sets that profile out row by row.

Frequently asked questions

How much does it cost to build an AI receptionist?

On our worked example, about US$20,343 of engineering up front (25 engineer-days at US$813.71 a day, derived from BLS wage data) plus about US$3,255 a month to maintain it. The per-call stack on top is small: US$0.085 for a 3-minute call inside LiveKit Cloud’s included minutes and US$0.127 above them. At 2,000 calls a month that is US$4,338 a month all-in, with the build spread over 24 months.

What does a self-built AI receptionist cost per minute to run?

About 2.8 cents a minute for telephony, speech-to-text, the language model, text-to-speech and recording, on Telnyx, Deepgram Nova-3, OpenAI gpt-6-luna and Deepgram Aura-2 at the prices on their own pages as at 8 November 2026. Hosting on LiveKit Cloud’s Ship plan adds 1.4 cents a minute once its 5,000 included minutes are used. Text-to-speech is the largest single line, not the language model.

At what call volume is building an AI receptionist cheaper than buying one?

Against Aircall’s published bundles, at about 6,846 calls of 3 minutes a month; against RingCentral’s pay-as-you-go rate, at about 2,976. Both assume 4 engineer-days a month of upkeep. Cut that to 2 days and the lines move to 4,654 and 1,791 calls; raise it to 6 and they move to 9,037 and 4,162. Aircall lists its bundles on its AI Receptionist page. Its pricing page also quotes custom rates from US$0.19 a minute above 10,000 minutes and as low as US$0.12 above 50,000, either of which would push the Aircall line higher.

How many engineer-days does it take to build an AI receptionist?

Our own estimate is 22 engineer-days for one inbound line in one language with a cold transfer, and 25 once you add a warm transfer at the low end of a 3 to 6 day range. That is an estimate, not a measurement; the stage-by-stage table is on our in-house build timeline page. Add booking edge cases, more languages or outbound callbacks and the number rises quickly.

Does call recording add much to the cost of an AI receptionist?

Capture is cheap; storage is the line that grows. Twilio charges US$0.0025 a minute to record and US$0.0005 a minute for every month a recording is kept, and its help centre says the 10,000 free stored minutes were discontinued from the July 2026 billing cycle. At 10,000 calls a month, two years of kept recordings come to roughly US$360 a month in storage. Telnyx lists call recording storage at US$0 per minute.

What does HIPAA compliance add to an AI receptionist build?

It changes vendors and plans, not just paperwork. LiveKit Cloud lists a signed BAA on its Scale plan, from US$500 a month, not on the US$50 Ship plan. Deepgram says it signs BAAs for Enterprise customers, a plan it prices through sales. Vapi lists HIPAA compliance as a US$2,000 a month add-on. Check every model vendor in your stack for its own BAA terms before you take a single patient call.

Does Zian AI publish its pricing?

No. Zian AI is in a partnership-application beta and does not publish prices. Each of the three plans zian.ai lists, including Intel + Widgets (Free Forever), starts with a partnership application. LeadsNow.ai, the done-for-you service that runs on Zian, is billed pay-per-result.

Where every figure on this page comes from

Figure	Who published it	Link	Date read
Telnyx Elastic SIP inbound local from $0.0032/min; toll-free from $0.015/min	Telnyx	telnyx.com/pricing/elastic-sip	8 November 2026
Telnyx call recording $0.002/min, storage $0; Deepgram Nova-3 $0.005/min via Telnyx	Telnyx	telnyx.com/pricing/voice-api	8 November 2026
Telnyx US local number $1.00/month	Telnyx	telnyx.com/pricing/numbers	8 November 2026
Twilio Elastic SIP local origination $0.0034/min	Twilio	www.twilio.com/en-us/sip-trunking/pricing/us	8 November 2026
Twilio receive local $0.0085/min, toll-free $0.0220/min, numbers $1.15 and $2.15/month, Media Streams $0.0044/min, recording $0.0025/min, storage $0.0005/min per month, volume pricing above 100,000 minutes	Twilio	www.twilio.com/en-us/voice/pricing/us	8 November 2026
Twilio partial minutes rounded up to the next full minute	Twilio	help.twilio.com/articles/223132307	8 November 2026
Twilio 10,000 free stored minutes discontinued from July 2026 billing	Twilio	help.twilio.com/articles/56876301519003	8 November 2026
Deepgram Nova-3/Flux $0.0077/min billed per second; Aura-2 $0.030/1K characters; BAA for Enterprise; Growth up to 20% off from $4,000/year	Deepgram	deepgram.com/pricing	8 November 2026
AssemblyAI Universal-Streaming $0.15/hr; Universal-3.6 Pro Realtime $0.45/hr	AssemblyAI	www.assemblyai.com/pricing	8 November 2026
OpenAI gpt-6-luna $0.10/$0.50 and gpt-6.1-terra $1/$5 per 1M tokens	OpenAI	developers.openai.com/api/docs/pricing	8 November 2026
Claude Haiku 5.5 $0.10/$0.50 and Sonnet 5.5 $2/$10 per 1M tokens	Anthropic	claude.com/pricing	8 November 2026
Gemini 3.1 Flash-Lite $0.25 input (text), $1.50 output per 1M tokens	Google	ai.google.dev/gemini-api/docs/pricing	8 November 2026
ElevenLabs Flash / v4 Turbo $0.04 per 1K characters, “~$0.04/minute”	ElevenLabs	elevenlabs.io/pricing/api	8 November 2026
Cartesia Startup $49/month for 1.25M credits, overage $45/1M, 750 to 800 credits per minute of audio	Cartesia	cartesia.ai/pricing	8 November 2026
LiveKit Cloud Ship $50/month, 5,000 agent and SIP minutes included, $0.01 and $0.004/min after; BAA on Scale ($500) and Enterprise; open-source framework	LiveKit	livekit.com/pricing	8 November 2026
Pipecat Cloud agent-1x $0.01/min active	Daily (Pipecat Cloud)	www.daily.co/pricing/pipecat-cloud/	8 November 2026
Vapi hosting $0.05/min; HIPAA add-on $2,000/month	Vapi	vapi.ai/pricing	8 November 2026
Retell Voice Infra $0.055/min	Retell AI	www.retellai.com/pricing	8 November 2026
Aircall AI Receptionist bundles $250 / $805 / $1,900; includes LLM, voice, telephony, unlimited concurrency, warm transfer; forward deployed engineer support on bundles of 1,000 minutes and above	Aircall	aircall.io/products/ai/ai-receptionist/	8 November 2026
Aircall pay as you go $0.29/min, overage $0.29/min, annual commitment, custom from $0.19/min above 10,000 minutes, as low as $0.12 above 50,000, 50 free minutes a month, standalone from $99/month	Aircall	aircall.io/pricing/	8 November 2026
RingCentral AI Receptionist $49/month with 100 minutes, then $0.50/min, 30-second increments, unpriced minute bundles	RingCentral	www.ringcentral.com/pricing/ai-receptionist.html	8 November 2026
Software developers (SOC 15-1252) mean hourly wage $71.20, May 2025 (series OEUN000000000000015125203)	US Bureau of Labor Statistics	api.bls.gov/publicAPI/v2/timeseries/data/OEUN000000000000015125203	8 November 2026
Wages and salaries 70.0% of private-industry employer compensation costs (68.5% for full-time private industry workers), June 2026	US Bureau of Labor Statistics	www.bls.gov/news.release/ecec.nr0.htm	8 November 2026

Engineer-day counts (22 and 25), the 4-day upkeep, the 24-month amortisation, the call shape, token counts and characters per minute are Zian AI’s own estimates and assumptions, stated above. All arithmetic is ours.