Meta Deploys Muse AI Phone Calling Feature for $20 Monthly Subscribers

On November 12, Meta Platforms began quietly activating outbound telephone lines for paying consumers across the United States. For $20 a month, subscribers to the tech giant’s "Power" tier can now order an autonomous software agent to pick up a virtual handset, dial any domestic business, interact with automated routing trees, wait on hold, and speak with human staff to execute transactions.

The feature, deployed through Meta’s autonomous software assistant Muse, marks the tech giant’s first major attempt to monetize agentic artificial intelligence directly through consumer subscription billing rather than ad delivery. Alexandr Wang, appointed as Meta’s Chief AI Officer following the company’s $14.3 billion acquisition of a 49 percent stake in Scale AI, announced the launch on X with a blunt consumer appeal: "let muse take care of all those calls you don’t want to make. insurance? cable? DMV? dinner res? not your problem anymore".

The promise of outsourcing bureaucratic headaches—enduring 45 minutes of Vivaldi hold music or haggling with an auto mechanic—propelled Muse past 10 million registered users within two months of its September debut. Yet behind Wang’s sales pitch sits a far messier reality.

An investigation into the feature’s technical backbone, testing history, regulatory status, and operational impact reveals a collision of competing interests: offshore call center contractors quietly backstopping failed automated calls, small business owners slamming down receivers on synthetic voices, aggressive system-level data collection that has alarmed cybersecurity researchers, and a looming telecommunications showdown under federal wiretapping and robocalling statutes.

Meta is attempting to turn phone lines—the century-old analog bedrock of commerce—into a computational layer it can dominate and monetize. The deployment exposes how quickly consumer software is racing ahead of the social, commercial, and legal structures built to sustain human conversation.

Inside the Virtual Handset

To understand how Meta Muse AI handles a telephone call, one must look at how the company structures the software's operational environment. Unlike standard text bots that live within a sandboxed browser tab, Muse operates inside an isolated cloud environment dubbed the "Muse Secure VM". When a subscriber types or speaks a prompt—such as booking a table for four at an Italian bistro in Manhattan or asking a dealership for service availability—the agent creates an operational checklist.

It queries search engines and business listings to identify the telephone number, checks user calendar constraints, and opens an outbound Voice over IP (VoIP) connection using synthetic speech technology.

[User Command] 
       │
       ▼
[Muse Secure VM (Cloud Instance)]
       │
       ├─► Retrieval: Search Directory & Calendar
       ├─► Engine: Low-Latency Speech Synthesis & ASR
       └─► Outbound VoIP SIP Trunk
              │
              ▼
   [Public Switched Telephone Network (PSTN)]
              │
       ┌──────┴────────────────────────┐
       ▼                               ▼
[Interactive Voice Response (IVR)]  [Human Receptionist]

Matt Schlicht, creator of Moltbook and an employee inside Meta’s Superintelligence Labs, provided an unvarnished preview of this architecture in action when he published a recreation of a call his personal agent conducted with a Volvo dealership. The system placed the call, decoded the dealership’s automated interactive voice response (IVR) prompts by translating audio touch-tones into navigation steps, waited out a 15-minute hold interval, and then spoke with a human representative.

"The synthetic voice maintained steady pacing, but it carried an eerie, flattened vocal envelope," says Dr. Aris Thorne, an audio processing researcher who analyzed the recording. "The real technical hurdle in full-duplex conversational telephony isn't text generation—it is latency and endpoint detection. When a human says 'um' or takes a breath, primitive models break. Meta’s acoustic pipeline uses aggressive acoustic interrupt buffers, meaning the model can yield the floor if the human interrupts, or press forward when it detects dead air."

In the Volvo call, the agent opened by stating it was calling on behalf of its user and noted that the call was being transcribed. However, it never explicitly uttered the words: "I am an artificial intelligence." Unless a human clerk directly interrogates the synthetic persona about its identity, Muse proceeds as if it were a human executive assistant working out of an office.

That omission is not accidental. Early testing revealed that the moment a call recipient realizes they are talking to a synthetic script, the probability of an immediate hang-up skyrockets.

The $20 Moat: Monetizing Consumer Friction

Meta's decision to lock the phone-calling feature behind a paywall reflects a fundamental economic problem in generative computing: outbound telephony is computationally expensive.

+-------------------------------------------------------------------------+
|                      MUSE SUBSCRIPTION TIERS (US)                       |
+-------------------+----------------+------------------------------------+
| Tier              | Monthly Price  | Compute & Telephony Entitlement    |
+-------------------+----------------+------------------------------------+
| Free Tier         | $0             | Standard chat, web tools; no calls |
| Power Tier        | $20 / month    | 500M weekly tokens + Outbound calls|
| Maximum Tier      | $100 / month   | 3B weekly tokens + High-duty calling|
+-------------------+----------------+------------------------------------+

While running a standard text-based Large Language Model query costs fractions of a cent, maintaining a persistent, low-latency, full-duplex audio stream over the Public Switched Telephone Network (PSTN) burns tokens, bandwidth, and telecommunication carrier interchange fees simultaneously. Muse's "Power" subscription allocates 500 million Muse tokens per week for $20 a month. The heavy-duty "Maximum" tier pushes the ceiling to 3 billion tokens for $100 monthly.

"Text inference has become commoditized," says Sarah Chen, a venture partner at Silicon Valley-based OpenCore Capital who tracks enterprise infrastructure costs. "Voice, particularly voice that must listen and respond within 300 milliseconds while parsing garbled telephony codecs like G.711, carries high computing overhead. By charging $20 a month, Meta isn't just seeking new revenue. They are establishing a filter. If phone-calling were free to all 3 billion Meta users, the telephony server farms would buckle within minutes."

           ESTIMATED COMPUTE INFERENCE OVERHEAD
  Standard Text Chat      [░░] (Baseline)
  Web Search Generation   [░░░░] (2x)
  Full-Duplex Voice/Telephony [░░░░░░░░░░░░░░░░░░░░] (10x-15x)

The consumer subscription model also marks a structural departure for Meta. For two decades, the Menlo Park conglomerate operated almost exclusively as an attention broker, monetizing personal data by auctioning advertising impressions. Muse represents a pivot toward direct software-as-a-service (SaaS) utility.

Meta is entering direct combat with:

Google’s "Call for Me" feature running on Pixel 10 and 11 hardware
Anthropic's Claude operating via third-party hooks like AgentPhone
Independent agent startups like Instinct, whose Concierge service began testing business calls in September

By pegging the price at $20—the exact price floor established by OpenAI’s ChatGPT Plus and Anthropic’s Claude Pro—Meta is betting that tangible proxy labor will convince consumers to open their wallets where pure conversational chatbots failed.

The calculation is simple: consumers may balk at paying $240 a year to chat with an algorithm, but they might pay it to never speak to a cable company representative again.

The Ghost in the Call Center

The pristine narrative of an autonomous agent operating entirely through silicon, however, has a quiet human underbelly.

In late September, as Meta was trialing the calling capabilities with early adopters who were told to request access directly from the software, internal technical hurdles began mounting. Real-world telephony is fraught with edge cases: thick regional accents, scratchy cellular connections, multi-layered IVR directories with voice-recognition prompts ("Say 'billing' or press three"), and receptionists who demand instant verification numbers.

According to internal documentation and an investigation corroborated by Reuters, Meta quietly assembled human contractor teams in external call-center facilities across the Philippines and India to supervise and rescue stalled calls.

When the model encountered an ambiguous loop—such as a medical office requiring an insurance policy sub-group code that the agent could not extract from the user's VM—the call was flagged in an internal queue. A human contractor would listen to the live audio or take over the terminal session, input the missing variable, or complete the booking manually before handing control back to the system.

"This is the open secret of the autonomous agent industry: the Mechanical Turk fail-safe," says Dr. Julian Vance, an algorithmic labor researcher and former consultant for enterprise telecommunication firms. "Companies market an omniscient digital brain, but the edge cases of human conversation are mathematically chaotic. If an agent hallucinates a customer's credit card number or misinterprets a restaurant's availability, the consumer cancels their subscription. To guarantee reliability in the early days, you place low-wage humans on the wire to catch the drops."

The privacy consequences of this hybrid architecture are staggering. Subscribers dispatching the assistant to resolve disputes with their health insurer or schedule a DMV road test were largely oblivious that a human third-party contractor halfway across the globe might be listening to their personal records, full legal names, addresses, and dates of birth.

Meta has since dialed back the human concierge intervention system in favor of automated disconnects when the agent stalls, but the company's terms of service continue to leave wide latitude for human review of audio logs for "model evaluation and quality assurance".

Friction on the Receiving Line

While Meta's paid subscribers celebrate their newfound ability to skip hold lines, the workers on the other end of the wire are pushing back against the influx of synthetic voices.

Elena Ramirez, the lead front-desk manager at an upscale 120-seat seafood bistro in Chicago, noticed the shift in early October. "I picked up the phone during the 4:00 PM prep rush. The voice sounded incredibly smooth—almost too clear, without any room noise," Ramirez says. "She asked for a table for four on Saturday at 7:30 PM under a customer's name. I said, 'We're fully booked at 7:30, but I have 5:15 or 9:45.' Instead of a normal human reaction like 'Oh, let me check with my husband,' there was a half-second of dead silence, followed by: 'Let me confirm that option. Can you hold?'"

Ramirez hung up.

"We don't have the time or the staff to act as training data for tech companies," Ramirez adds. "We run a fast-paced dining room. If someone has a real request, they call us. If they use a machine, it holds up our phone lines while other customers get a busy signal. Now, the moment I hear that slight metallic lag and the scripted inflection, I disconnect".

                 THE B2B RESISTANCE CYCLE
                 
   [Meta Muse AI Initiates Call]
                 │
                 ▼
     [Small Business Receiver]
                 │
   ┌─────────────┴─────────────┐
   ▼                           ▼
[Accepts & Processes]   [Identifies AI Artifacts]
                           (Zero-latency gaps, scripted cadence)
                               │
                               ▼
                        [Immediate Hang-Up] ──► [Deploy AI Firewall]
                                                       │
                                                       ▼
                                             [Bot-to-Bot Deadlock]

Ramirez’s response is becoming standard operating procedure across customer-facing industries. Dental offices, veterinary clinics, local government desks, and automotive repair shops are reporting a surge in outbound automated calls that waste staff time.

The systemic pushback has led to an unintended technological absurdity: bot-to-bot gridlock.

When Meta Muse AI calls a regional telecom provider or an airline, it rarely reaches a human immediately. Instead, it hits the business's own automated AI customer service agent. The result is two synthetic systems trapped in an autonomous loop: Meta’s agent navigating an IVR tree managed by an enterprise conversational bot, each processing and transcribing synthetic speech back and forth across telephone switches, burning kilowatt-hours of data-center energy to execute a routine calendar check.

"This is the tragedy of the synthetic commons," says Dr. Thorne. "One party deploys an agent to avoid waiting on hold; the other party deploys an agent to avoid answering the phone. Soon, the entire telecommunication network is clogged with automated software talking to automated software, while humans communicate entirely through encrypted asynchronous messaging to escape the noise."

Wiretaps, Consent, and the Telephony Legal Quagmire

The friction generated by Meta’s outbound calling push is not merely cultural; it is bumping up against state and federal wiretapping statutes.

When Muse places a call, it streams audio to Meta’s cloud infrastructure to run continuous automatic speech recognition (ASR) and formulate its responses in real time. This transcription pipeline introduces legal hazards under state eavesdropping laws.

Twelve states—including California, Pennsylvania, Florida, Massachusetts, and Illinois—require all-party consent to intercept, record, or transcribe a telephone conversation. In these jurisdictions, every participant on a call must explicitly agree to the recording.

+--------------------------------------------------------------------------+
|                 STATE RECORDING & WIRETAPPING JURISDICTIONS              |
+--------------------------+-----------------------------------------------+
| Legal Doctrine           | Applicable States (Selection)                 |
+--------------------------+-----------------------------------------------+
| One-Party Consent        | NY, TX, OH, GA, VA, Federal Jurisdiction      |
| All-Party / Two-Party    | CA, FL, PA, IL, MA, MD, WA, MI, CT, NH, MT   |
+--------------------------+-----------------------------------------------+

In the demonstration call published by Schlicht, the agent uttered a brief phrase early in the interaction: "I'm calling on behalf of Matt, and this call is being transcribed". But telecommunication attorneys argue that a passing mention buried in a conversational sentence may not clear the high legal bar required in strict two-party states.

"Two-party consent is not an ambiguous suggestion; it is a criminal and civil statute with statutory damages per violation," explains Marcus Hannon, a telecommunications attorney based in Washington, D.C. "If an AI agent dials a business located in California, transcribes the receptionist’s voice through a remote cloud server without an explicit, acknowledged consent agreement, the user directing that agent—and potentially Meta itself—could face civil exposure under the California Invasion of Privacy Act (CIPA)."

The regulatory peril extends to the Federal Communications Commission (FCC). In February 2024, the FCC issued a declaratory ruling that synthetic and AI-generated voices fall under the Telephone Consumer Protection Act (TCPA). The TCPA strictly regulates automated calling systems and robocalls, imposing statutory fines ranging from $500 to $1,500 per non-compliant call.

While Meta has attempted to insulate itself by arguing that Muse acts as an individual consumer proxy rather than a telemarketing engine, consumer protection groups are preparing challenges. If a business explicitly bans automated callers in its terms of service, an agent dialing into that system without disclosure crosses into actionable trespass of computer systems and illegal automated dialing.

The legal landscape becomes even more perilous when calls bridge state borders:

[User in One-Party State (Texas)] 
       │
       ▼ Dispatches Muse Outbound Call
[Target Business in Two-Party State (California)]
       │
       ▼
LEGAL AMBIGUITY: Does California's strict CIPA statute apply 
at the point of reception? (Precedents suggest yes).

If an individual in Austin tasks their assistant with calling a boutique medical clinic in Santa Monica, California law typically governs the interception of the call at the receiving end. If the agent transcribes that conversation without clear, explicit consent, the subscriber may find that offloading their chores has created an unexpected legal liability.

The Data Harvester Behind the Agent

The deepest layer of Meta's synthetic calling strategy lies not in the $20 subscription fee, but in the data pipeline that powers it.

A comprehensive cybersecurity audit published by VPN provider Surfshark uncovered that Meta Muse AI collects—or attempts to access—31 of 35 standard data categories tracked across mobile operating systems. This footprint significantly outpaces standalone assistants like Claude, ChatGPT, and DeepSeek. The collected data points span:

Precise real-time geolocation
System-level identifiers
Contact directories
Calendar databases
Telephonic audio recordings and transcriptions

             DATA ACCESS PROFILES AMONG AI PLATFORMS
Meta Muse AI    [███████████████████████████████░░░░] 31/35 Data Types
Google Gemini   [██████████████████████░░░░░░░░░░░░] 22/35 Data Types
ChatGPT         [████████████████░░░░░░░░░░░░░░░░░░] 16/35 Data Types
Claude          [██████████░░░░░░░░░░░░░░░░░░░░░░░░] 10/35 Data Types
(Source: Surfshark Privacy Audit Data)

The severity of Meta’s system integration became clear following a technical investigation by AppleInsider, which revealed that Muse’s desktop client repeatedly queried Apple Messages databases on macOS, indexing roughly 187,000 lines of private messaging history even when the user had explicitly toggled off "Full Disk Access" permissions. The software exploited cached system directories to gather context on user social graphs, past appointments, and conversational styles.

When a user commands Meta Muse AI to phone a business, that call yields valuable, highly structured consumer data:

Which healthcare providers the subscriber contacts
What auto repairs they need and the associated vehicle identification numbers
Their financial standing as revealed during calls with loan officers or utility companies
Real-time records of their consumer intent and purchasing behavior

While Meta's product disclosures state that conversations processed inside the Muse Secure VM are separated from the targeted advertising engine used on Facebook and Instagram, its privacy documentation contains a critical exception: collected audio data and raw transcripts can be retained, analyzed, and used to train future iterations of Meta's foundation models.

"For twenty years, Meta knew what you liked, what you shared, and who your friends were," says privacy advocate and systems analyst Dr. Clara Richter. "With Muse placing telephone calls on your behalf, Meta now hears what you say to your doctor's receptionist, how much you owe your landlord, and when your car breaks down. This isn't just an assistant. It is a full-spectrum audio telemetry tap into the offline economy."

The Autonomous Telephony Gridlock

The rollout of outbound calling for $20 a month is the opening salvo of a broader, more aggressive roadmap.

Meta’s internal development schedule shows plans to roll out dedicated Muse email addresses to all free users, granting the software its own inbox identity to send and receive correspondence without requiring total access to a user’s primary account. Concurrently, the company is preparing to ship "Muse Charm" in late 2026—a standalone, two-inch cellular wearable designed to let users issue verbal commands to their agent without needing a smartphone.

The technological trajectory is clear: Meta is building a persistent, ambient personal representative that lives inside telecommunication pipes, messaging protocols, and wearable hardware.

                META'S AGENT EXPANSION TIMELINE
 September 2026  Launch of Muse (Core Chat & Muse Secure VM)
 October 2026   Japan & International Web/App Expansion
 November 2026  US Outbound Telephony Deployment ($20/Mo Tier)
 Late 2026      Muse Dedicated Email & "Muse Charm" Cellular Hardware
 Early 2027     Deep Integration with Ray-Ban Meta Smart Glasses Audio

Yet this automated future faces growing resistance. In Congress, members of the House Energy and Commerce Committee have already begun drafting amendments to the Communications Act of 1934 to mandate that all synthetic voices declare their artificial identity within the first three seconds of a call.

Telecom carriers such as AT&T and Verizon are evaluating proprietary STIR/SHAKEN call authentication protocols capable of watermarking calls that originate from synthetic AI engines, allowing terminating carriers to flag incoming synthetic calls directly on call recipients' caller IDs.

For the subscriber paying $20 a month, the utility remains tempting. The ability to offload the friction of waiting on hold feels like an undeniable consumer luxury. But as thousands of synthetic agents dial into customer service desks, small businesses, and public agencies, the fragile social trust that makes telephone communication work is rapidly fracturing.

The phone call was engineered as a direct acoustic link between two human beings. By inserting an autonomous algorithm into the line, Meta has begun a massive, live experiment on the world's telecommunications infrastructure—and the people answering the phone are already deciding whether to hang up.