Having an AI call local businesses for you, in your own voice

You asked for the best tools for AI outbound voice calling where the conversation actually sounds human, drivable from Claude Code or Codex, ideally speaking in your own voice. Here is what the market really looks like in July 2026, what it costs, what will break, and what I would build.

Researched 27 July 2026 across four parallel threads: the platform landscape, voice cloning quality, driving calls from code, and the legal rules. Prices verified against vendor pricing pages the same day.

Start here: the two things only you can do

Everything else on this page is mine to build. These two steps need your card and your voice, so they are yours.

What happens next, and a correction

I originally described this as a $6, half-hour test that ends with the agent calling your mobile. That was not quite right, and I want to correct it rather than let you discover it mid-setup. A real outbound call needs a phone number, which means a Twilio account on top: free trial credit, calls allowed to verified numbers, so your own mobile works fine. That is another twenty minutes of signup and identity verification.

But the question that actually decides your money is narrower than that: does your voice still sound like you once the phone codec has mangled it? I can answer that with no telephony at all. I have already confirmed your ElevenLabs key can synthesize speech and emit raw ulaw_8000, which is bit-for-bit what a phone line does to audio.

So once your clone exists, I will generate the same few sentences twice, once at full studio quality and once through the phone codec, and put them side by side on a hub page. You listen. Better still, have Bob listen without telling him which is which.

• If your voice survives it, the clone is worth using and we do the Twilio setup and build the real thing.
• If it does not, we have saved the effort and I would steer you to a well-chosen stock voice instead, and the "my own voice" premise quietly goes away.

Either way, the Professional clone still is not worth buying, because everything it improves sits in the band the codec deletes.

The short answer

Build it on ElevenLabs Agents, on their Starter plan at $6 a month, using an Instant Voice Clone. It is the only platform where your cloned voice is native rather than bolted on, and it ships the two unglamorous features that decide whether this works at all: voicemail detection and phone-tree navigation. It also has an official MCP server, which means Claude Code can place the call and read back the answer without me writing a dialer.

Do not pay for the Professional Voice Clone. This is the single biggest money-and-time saving in the research and I nearly got it wrong. The Professional clone costs $22/mo instead of $6 and wants one to three hours of clean studio audio instead of a minute or two. What it buys you is sibilant detail, breathiness and harmonic richness, and all of that lives above 3.4 kHz, which is exactly the band a phone line throws away. More on this below, because the physics is genuinely convincing.

The runner-up is Vapi, and it is close. Pick it instead if you want the agent to survive an ElevenLabs outage, because Vapi lets you swap any component. I would move you there if this turns into something bigger than a personal errand.

What it costs

Assuming 30 calls a month at about 3 minutes each, so roughly 90 minutes:

If you later want the Professional clone anyway, or you start making enough calls that 75 minutes is tight, the Creator plan at $22/mo gives you 275 minutes and one Professional Voice Clone slot. At your stated volume that is paying for a bucket you will not empty.

The build, in three steps

None of this is work you do. It is roughly an afternoon for me.

1. Clone your voice

An Instant Voice Clone needs one to two minutes of clean audio and completes immediately. You have plenty of usable audio already from your video and podcast work, so this is selecting a clip, not recording a session.

There is one step only you can do. ElevenLabs requires a voice verification for every clone, instant or professional: you read a short line off the screen, live through your microphone, once. Reading it more than once can fail the check, and a full failure means waiting 24 hours or opening a support ticket. It takes about a minute, but it needs your voice and your mic, so I cannot do it for you.

Related housekeeping: your ELEVENLABS_API_KEY is live but scoped too narrowly to read your account or list voices, so I could not confirm your current plan tier programmatically. I will need either a broader key or one dashboard step from you when we start.

2. Build the agent

The agent gets a system prompt describing the errand, your cloned voice, a disclosure opening line, voicemail detection turned on, and a data collection schema, which is the feature that turns a phone call into usable data. Instead of handing you a transcript to read, the call returns something like:

{
  "price": 149,
  "currency": "USD",
  "price_includes": "exam and cleaning, x-rays extra",
  "accepts_your_insurance": true,
  "soonest_appointment": "August 11",
  "reached_a_human": true,
  "notes": "Quoted range 149 to 189 depending on x-rays."
}

3. Wire it to Claude Code

ElevenLabs publishes an official MCP server (1,488 stars, updated four days ago) whose make_outbound_call tool does exactly what it sounds like. Installed once, you can hand me a list of businesses and a question, and I place the calls and come back with a table. For repeatable runs I would also write a small script so a batch of calls costs no Claude tokens at all.

How the platforms actually compare

A few things that table does not capture:

• Retell is genuinely cheaper and has the best backchannel, which is the feature that makes an agent murmur "mhm" while you talk. It is disqualified for you on two counts: it requires government ID verification before it will let you make a single outbound call, and it charges more for long prompts, which penalizes exactly the detailed, natural-sounding persona prompt you would want.
• Vapi looks cheapest at $0.05/min but that is the orchestration fee only. Independent operators who normalized all platforms to the same stack put Vapi's real all-in cost at $0.135 to $0.177 per minute. Its genuine advantage is that every component is swappable, which matters because, in one builder's words, ElevenLabs goes down, OpenAI goes down, Gemini goes down, and it is not rare.
• Bland AI is the one I would actively avoid. In June 2024 WIRED documented its demo bots lying about being human, including telling a tester "I can assure you that I am not an AI." I found no retraction, no enforcement, and no follow-up. It is also the most expensive per minute. Calling your own neighborhood businesses is the wrong context to be standing behind that brand.
• Google and OpenAI are both out, for the same reason. Gemini's Live API and OpenAI's Realtime API are excellent conversation engines, but both lock you to their own fixed voice sets. Gemini gives you 30 stock voices and no way to bring your own; OpenAI is the same. Since your own voice is the point, that disqualifies both. Twilio's ConversationRelay is out on the same grounds: it has no mechanism to supply your own credentials, and custom cloned voices are not documented as addressable.
• Two platforms are effectively dead. Layercode pivoted to a different product entirely and its domain now serves something else. Vogent's site still carries a 2024 copyright. Do not let anyone build you anything on either. Separately, Air.ai, which appears on a lot of older shortlists, was sued by the FTC and settled in March 2026 for $18M with a permanent ban on marketing business opportunities.

What actually makes a call sound human

This was the most useful thing the research turned up, and it contradicts every vendor homepage.

Latency is the tell, not voice quality. Every platform advertises sub-500ms response times. Those are component measurements, usually time-to-first-byte on one stage of the pipeline. Real full-turn latency on a live phone call in 2026 measures 700ms to 1.2 seconds, and worse under load. Twilio's own engineering team publishes a target of 1,115ms, which is the most honest number any vendor has put in writing. Vapi's docs quietly admit their dashboard latency figure excludes two of the five stages.

People on the receiving end describe the giveaway consistently, and it is never the voice:

One giveaway was a long pause after that as it processed. Another was the flatness, no mistaking the lack of normal human randomness in tone. No taking breaths, that sort of organic stuff.

The one real quality measurement I found comes from someone running production outbound calls for 18 months, who tracks what they call a first-five-second detection rate, meaning how often the person realizes they are talking to AI within five seconds. With ElevenLabs they measured 15 to 20%. With Azure it was closer to 40%. That gap is the single strongest argument for building on ElevenLabs, and it is one anecdote, so hold it loosely.

What breaks a call is not the voice, it is going off script. From someone running these at volume: voice AI will crush a demo and then implode the second someone interrupts, asks a weird question, or gets annoyed. Which is precisely what a busy receptionist does.

The rules, and the opening line

I am not a lawyer and none of this is legal advice. It is what the statutes and agency documents actually say, separated from what is merely common practice.

The good news: this is much lighter than the "AI robocall" headlines suggest. You are an individual asking a business a factual question, not a company selling to consumers. Nearly every state AI-disclosure law regulates businesses talking to consumers and does not bind you. Cloning your own voice is unambiguously legal, because every voice-cloning statute exists to protect a person against others using their voice, and you are the rightsholder.

Three things do matter, and two of them are invisible from the outside:

1. Screen out cell numbers before dialing. This is the single highest-value item in this document, so it has its own section below.
2. Do not point this at medical offices. Federal law specifically protects any emergency line of a hospital or physician's office, with no exemptions at all. Small practices routinely run one line that doubles as the after-hours emergency line, and you cannot tell which is which from outside. High risk, low reward, and front desks usually refuse to quote prices by phone anyway.
3. Disclose, every time. California has required a verbal AI disclosure at the opening of AI-voice calls since January 2025, at $500 per call. A federal rule requiring the same is proposed but not final. Separately, ElevenLabs' own terms contractually require you to disclose both the AI and the recording before any interaction, so on the recommended stack this is not optional regardless of what the law says.

The good news is that disclosure is also the effective choice, not just the safe one. The pattern across every recipient account I found is that disclosed AI is tolerated and undisclosed AI gets hung up on. Here is an opening that satisfies the AI disclosure, the federal caller-identification requirement, recording consent in all fifty states, and gives the person a graceful exit, in one breath:

Hi, I'm an AI assistant calling on behalf of Annette Thompson, and this call is recorded. She asked me to check what you charge for a new patient exam. Is now a good time, or would you rather she call you directly?

Two smaller rules worth building in from the start: keep calls inside 8am to 9pm local time, and never call the same number more than three times in 30 days. Also, use a real number that traces back to you and never spoof caller ID. Spoofing is what actually gets people fined; the one $6M AI-voice penalty on record was a caller-ID case, not a cloned-voice case.

The one risk I cannot fully resolve for you

ElevenLabs' Prohibited Use Policy bans "unauthorized robocalling, meaning the use of automated dialing systems or artificial/pre-recorded voice messages to place phone calls to individuals or businesses without direct human intervention."

That clause is aimed at spam campaigns, and the word "unauthorized" is carrying a lot of weight. A person placing thirty disclosed calls a month to ask factual questions is plainly not what it was written to stop. But I am not going to pretend the language clearly exempts you, because it does not, and this is the platform I am recommending.

What this actually means in practice: the realistic downside is account suspension, not legal trouble. It is a terms risk, not a liability risk. Three things keep you well clear of the line, and I would build all of them in regardless: disclose the AI in the opening sentence every time, keep the volume low and human-paced rather than running a list, and never let the agent deny being an AI if someone asks. If you would rather not carry even that risk, Vapi with your own ElevenLabs key is the alternative, because Vapi's terms contain no robocalling clause at all. It costs a few dollars more a month and a bit more setup.

One practical footnote on that Twilio number. If you register a Business Profile with Twilio (24 hours) and a caller-verification profile (72 hours), your calls get signed at the highest attestation level, which is what keeps carriers from labeling you "Spam Likely". Skip it and repeated calls from one number will eventually get flagged, which degrades the number for your real use too. It is free and I would do it before the first call.

The cell number problem, in detail

This is the one rule that can actually cost you money, and it is worth understanding properly rather than taking on faith.

Why business landlines are fine and cell numbers are not

The TCPA's artificial-voice restrictions live in 47 U.S.C. 227(b)(1), and they are split across two separate subsections that behave completely differently.

Subsection (b)(1)(B) covers residential lines. That is the clause everyone thinks of as "the robocall law". It says nothing about business lines. Congress actually noticed this gap and, in 227(b)(2)(A), instructed the FCC that it "shall consider prescribing regulations to allow businesses to avoid receiving calls made using an artificial or prerecorded voice." The FCC has never adopted those rules. That un-filled gap is the entire reason calling a business landline with an AI voice is not prohibited.

Subsection (b)(1)(A)(iii) covers cell numbers, and it is written far more broadly. It reaches "any telephone number assigned to a paging service, cellular telephone service, specialized mobile radio service, or other radio common carrier service, or any service for which the called party is charged for the call."

Read that list again and notice what is not in it:

• No residential versus business distinction. A cell phone that a plumber uses purely as a business line is covered exactly the same as a personal phone.
• No commercial versus non-commercial distinction. It does not matter that you are asking a question rather than selling something.
• No "small number of calls" allowance. One call is one violation.

Why the exemption you would expect to save you does not

There is a non-commercial exemption, at 47 C.F.R. 64.1200(a)(3)(ii), and it fits your use case almost perfectly: a call "not made for a commercial purpose", capped at three calls per 30-day period to that line.

The problem is that it is an exemption from (b)(1)(B) only, the residential clause. It does not reach the cell-phone prohibition at all. The wireless exemptions are a separate, much narrower list at 64.1200(a)(9), and they cover things like bank fraud alerts, healthcare messages from covered entities, and package delivery notifications. "A person asking a shop what they charge" is not on that list, and there is no catch-all.

What it would actually cost you

The exposure is not really the FCC. It is 227(b)(3), the private right of action, which lets the person you called sue directly:

To be fair about the actual risk: the probability that any one Boulder business sues you over a single polite pricing question is very low. But it is the only path in this entire document that leads to real money changing hands, and unlike the other risks it does not require you to do anything careless. You would just have dialled a number that happened to be a mobile.

One nuance that cuts both ways: in McLaughlin Chiropractic v. McKesson (June 2025) the Supreme Court held that district courts are not bound by the FCC's interpretations in private TCPA suits. So a defendant could argue a real-time conversational agent is neither "artificial" nor "prerecorded". Nobody should plan around winning that argument.

The fix, and it is genuinely cheap

Run a carrier line-type lookup on every number before dialing and skip anything that comes back mobile. Twilio's Line Type Intelligence costs $0.008 per lookup, and returns the carrier and line type: mobile, landline, fixed VoIP, non-fixed VoIP, toll free.

At 30 calls a month that is 24 cents. It removes essentially all of the TCPA exposure, and it is one API call that runs before the dialer ever fires. I would build it as a hard gate, not a warning: if the lookup says mobile, the number never gets called, and it lands on a short list you can phone yourself if the answer matters enough.

Two practical notes on it. Non-fixed VoIP numbers (Google Voice and similar) are a gray area, because they are not literally "assigned to cellular telephone service" but they often ring a mobile handset. I would treat them as mobile and skip them too, since the cost of skipping is nothing and the cost of being wrong is $500. And a landline result is not a permanent guarantee, because numbers get ported, so I would re-check anything older than a few months rather than caching results forever.

Does your voice survive a phone line?

This was the question I most wanted answered, because if the codec flattens your clone into a generic voice then the whole premise falls apart. The answer is yes, mostly, and the reason is worth understanding because it is also why you should not buy the expensive clone.

A phone line passes roughly 300 Hz to 3,400 Hz. What carries your identity to a listener is pitch, the first three formants, and glottal timing, and all of that sits inside that band. What gets destroyed is sibilance (most of the energy in an "s" lives from 4 kHz up), breathiness, and harmonic sparkle.

Measured effect: human identification of known voices drops from about 88% unprocessed to about 69% through narrowband telephone processing. Degraded, nowhere near destroyed. The uncomfortable real-world confirmation is that voice-clone fraud works precisely because identity survives the phone codec, and scammers explain away residual artifacts as a bad connection, which works because the codec's own degradation is already masking the clone's imperfections.

One technical detail I will build in: request the audio as ulaw_8000 directly rather than generating at 24kHz and letting the carrier downsample. Both ElevenLabs and Vapi publish engineering guidance that the extra resampling step adds artifacts and latency for no benefit.

What I would test first, before building anything

Two cheap tests, in this order.

Clone yourself and call your own mobile. $6 and about half an hour. Then have Bob or one of the kids listen without being told which it is. That settles the one gap in the evidence above, and it tells you whether the voice is convincing enough to be worth using at all.

Then make ten real calls and see whether businesses actually answer the question. This is the risk nobody can research away. A pile of "we don't quote prices over the phone" transcripts would mean the tool works perfectly and still saves you nothing. Better to learn that at $11 than after we have automated it.

What I could not verify

• Whether the FCC finalized anything in its AI-disclosure docket during 2026. I found no final order, but that deserves a direct docket check before you rely on it.
• Whether Retell accepts your own ElevenLabs API key. No documentation exists either way. Moot unless you reverse the recommendation.
• Whether an Instant clone and a Professional clone are actually distinguishable over a phone codec. The physics says barely, but nobody has run the controlled test. Yours would cost $6 and half an hour.
• How ElevenLabs actually enforces the robocalling clause in its use policy for low-volume disclosed calling. No published enforcement examples exist either way.
• Independent end-to-end latency benchmarks. None exist publicly that I could trace to a primary source. Every latency number any vendor publishes is self-measured, using different definitions.
• Real-world success rates for this specific use case. Genuinely nobody has published one.

Sources

Every price and capability above traces to one of these, checked 27 July 2026.

• ElevenLabs Agents pricing and main pricing (plan tiers, included minutes, cloning tiers)
• ElevenLabs batch calling, voicemail detection, data collection
• ElevenLabs TCPA guide, mandatory disclosure requirement, and the Prohibited Use Policy (the robocalling clause)
• ElevenLabs voice cloning guide (Instant needs Starter, Professional needs Creator) and the Twilio cookbook (why to request ulaw_8000 directly)
• Human speaker identification of known voices through different transmission channels (the 88% to 69% figure) and Interspeech 2024 on familiarity as the dominant factor
• ElevenLabs MCP server (the make_outbound_call tool)
• Vapi pricing, outbound calling, what their latency figure excludes, MCP server
• Retell pricing, mandatory KYC, backchannel config
• Bland pricing and WIRED, "This Viral AI Chatbot Will Lie and Say It's Human"
• Cartesia pricing and acceptable use policy
• Twilio engineering on real voice-agent latency (the 1,115ms target)
• FCC Declaratory Ruling 24-17 (AI voices are "artificial" under the TCPA) and 47 U.S.C. 227 (the cell-phone and emergency-line prohibitions)
• 47 C.F.R. 64.1200 (the non-commercial exemption at (a)(3)(ii), the narrow wireless exemptions at (a)(9), and the caller-identification rules at (b))
• McLaughlin Chiropractic v. McKesson, 606 U.S. (2025), on courts not being bound by FCC interpretations
• Twilio Lookup pricing (Line Type Intelligence at $0.008 per request)
• C.R.S. 18-9-303 (Colorado is one-party consent for recording)
• FTC settlement with Air.ai, March 2026
• Google "Ask for me", the closest thing to prior art for this exact use case