AI Agents Calling Businesses: The 2026 Phone Handoff

AI agents calling businesses is no longer a demo. Google Search can place automated calls to local businesses on a customer's behalf across most of the United States, and the businesses picking up increasingly answer with voice agents. Two models now negotiate a haircut over a phone line.

TL;DR

At I/O on May 19, 2026, Google said that "for select categories like home repair, beauty or pet care, you can ask Google to call businesses on your behalf," and that "these capabilities will roll out to everyone in the U.S. this summer."¹ Google's Business Profile documentation describes the feature as available "only in the United States, except in these states:" — Indiana, Louisiana, Minnesota, Montana and Nebraska — with calls "monitored and recorded for quality assurance," and warns that it "may not be available for all users."²

Meanwhile the answering end is automating just as fast. xAI released a no-code Grok Voice Agent Builder in beta on July 1, 2026, billed at $0.05 per minute of audio plus $0.01 per minute for telephony on a provisioned number, joining ElevenLabs, Synthflow, Retell AI, PolyAI and Vapi in a crowded field.³⁴ The result is a machine-to-machine transaction running over the one interface both sides already had: a phone number. That is an interoperability gap wearing a product launch costume, and it costs real money per minute.

What You'll Learn

• What Google actually shipped, and how it differs in stage from "Ask for Me"
• The five US states where Google says these automated calls are not available
• The line in Google's own documentation that shows structured data beats a phone call
• Why the business on the other end is increasingly a voice agent too
• What a minute of synthetic voice actually costs, and what the sticker price leaves out
• Why agents phone each other when an agent-to-agent protocol already exists
• The TCPA layer that local-SEO coverage tends to skip
• What a business owner can actually do about it this week

What Google shipped, and what stage it is at

Two Google efforts get blended together, and the distinction worth drawing is one of launch stage rather than product identity — they have different dates, different scopes, and different levels of permanence.

The first was Ask for Me, a Search Labs experiment launched on January 30, 2025, initially limited to nail salons and auto repair shops.⁵ Labs experiments are opt-in, reversible, and deliberately narrow.

The second is agentic calling, the feature in Search itself. Google's how-to page for it was published on November 25, 2025 and carries an "Updated July 24, 2026" line.⁶ That page describes a retail flow: search for a product with "near me" or "nearby," get an option to have Google confirm it is in stock nearby, answer a short list of questions tailored to the product, indicate whether you want the summary "via text, email or both," and let Google place the calls.⁶

At I/O on May 19, 2026, Search VP Elizabeth Reid extended it beyond retail. The post announcing the change says Google is "expanding agentic booking capabilities in Search to a wide range of new tasks, including local experiences and services," and adds that "for select categories like home repair, beauty or pet care, you can ask Google to call businesses on your behalf. These capabilities will roll out to everyone in the U.S. this summer."¹ The same post puts Gemini 3.5 Flash in as the default model in AI Mode globally, and claims AI Mode "has surpassed one billion monthly users" with queries "more than doubling every quarter" since launch — Google's own figures, not independently audited.¹

Summer 2026 is the window Google gave itself, and as of early August that rollout is still running. Google's help page carries its own hedge: the feature "is currently available for select regions and languages and may not be available for all users."² So the interesting question this month is not whether the capability exists. It is what the plumbing looks like where it does.

The five states where Google's agents do not call

Google's Business Profile help page is the most operationally specific public document Google has published on the feature, and it gets far less attention than the launch coverage. It lists three customer actions that can trigger a call: "Book appointments," "Check restaurant wait times," and "Confirm price and availability of services and products."²

Then it draws a geographic line. Appointment bookings, restaurant wait-time checks and product and service confirmations via Google automated calls, the page says, "are currently available only in the United States, except in these states:" — and lists five: Indiana, Louisiana, Minnesota, Montana, Nebraska.²

Google does not explain the carve-out. It would be easy to assume a legal cause and write it up as fact — several states do restrict automatic dialing-announcing devices in ways that predate the internet — but Google publishes no rationale, so this piece does not assert one. What can be said is narrower and more useful: the feature's availability map has state-level seams, which suggests outbound automated calling is being routed around something jurisdiction-specific rather than deployed as one uniform national service.

The same page also states that "the calls are monitored and recorded for quality assurance," that Google "limits the number of calls for the same info and avoids calling during late night or early morning hours," and that the feature is free to both businesses and customers.² It also lists five conditions under which Google will call at all: the three customer-triggered ones above, plus checking business hours and checking "the status of in-demand inventory."² Two of those five are Google's own data hygiene, not a customer request.

The line Google buried: structured data beats a phone call

One sentence on that help page is not about calling at all. It is about not calling, and it explains the architecture better than the launch posts do:

"If you've already shared your in-store product details via Google Merchant Center, Google will use that information and will not call your business to ask about inventory."²

Read that as an engineering statement rather than a support note. Google is saying that when a machine-readable feed exists, the agent reads the feed. The phone call is the fallback path for businesses that have not published structured data. The same page describes the booking flow in the same shape: "If available, Google can book through an online booking partner or by calling your business directly with an automated call."² API first, voice second.

That ordering tells you what the phone call actually is. It is not a feature. It is a compatibility shim for the long tail of businesses whose inventory, pricing and availability exist only in someone's head or on a whiteboard behind the counter. Every customer-triggered agentic call is a small tax paid because a structured endpoint did not exist.

This is the same pattern as browser agents driving human interfaces with a mouse cursor because the underlying service exposes no API — a pattern we covered when AI browser agents from Claude, ChatGPT and Gemini started competing for control of the screen. Voice is the audio-band version of the same compromise.

The other end of the line is an agent too

Here is what makes 2026 different from the era of the 2018 Duplex demo: the business being called now has cheap, credible automation of its own.

On July 1, 2026, xAI announced Grok Voice Agent Builder in beta, a browser-based, no-code platform that xAI describes as being for "operators and developers who want high-volume production voice agents without building the surrounding stack from scratch."³⁴ Users define call handling, connect knowledge bases and external tools, deploy over a phone number or a SIP connection, monitor conversations, and pick synthetic or cloned voices. It supports more than 25 languages.³

Architecturally, the pitch is a single speech-to-speech model rather than the usual cascade. Most voice AI applications stitch together separate speech recognition, language model and speech synthesis components, with each hop adding cost, latency and new failure modes; routing through one model is meant to cut all three.³⁴ Twilio's developer guidance on core latency in AI voice agents covers the same baseline delays from the telephony side.⁷

xAI is late rather than early here, and says so implicitly: its voice work began with a Grok Voice Agent API in December 2025, followed by speech-to-text and text-to-speech APIs, conversational voice models and custom voices before the builder appeared.³ The competitive set named alongside it includes ElevenLabs, Synthflow, Retell AI, PolyAI and Vapi.³ The competition is on price, latency and how little glue code an operator has to write.

Stack the two halves together and you get the setup this article is named for. A platform agent dials. A tenant agent answers. Neither one is a person, and the shared protocol is a voice channel.

To be precise about what is and is not known: there is no public measurement of how often a Google agentic call is answered by another agent rather than a human. This is a structural condition, not a measured phenomenon. What is measurable is the cost of the channel.

What a voice leg actually costs

Published list prices give a defensible floor. ElevenLabs' ElevenAgents pricing page, read on August 4, 2026, lists an additional call minute at $0.080 across every tier from the free plan through Business, a burst rate of $0.160 per minute, and text messages at $0.003 each.⁸

⚠ Prices change frequently. The values above are for illustration only and may be out of date. Always verify current pricing directly with the provider before making cost decisions: Anthropic · OpenAI · Google Gemini · Google Vertex AI · AWS Bedrock · Azure OpenAI · Mistral · Cohere · Together AI · DeepSeek · Groq · Fireworks AI · Perplexity · xAI · Cursor · GitHub Copilot · Windsurf.

Source: ElevenLabs ElevenAgents pricing page, read August 4, 2026. Prices exclude taxes, levies and duties.⁸

The number that matters is not in the headline row. The same pricing table lists external providers separately: the LLM is billed "based on usage and varies by model," and telephony is billed "at cost."⁸ So $0.080 per minute is the platform's cut, not the all-in cost of a conversation.

xAI publishes the arithmetic ElevenLabs leaves to the reader. Its launch post states that agents "are billed at our API rate (currently $0.05 / min of audio), with voices included and no separate platform fee," and that "telephony on a free provisioned number is an additional $0.01 / min."⁴ That is a real all-in number: $0.06 per minute on an xAI-provisioned line, before any tool calls. It is a useful benchmark precisely because it is complete.

Some illustrative arithmetic, with the assumptions stated so you can substitute your own. Take a 90-second call, which is a plausible length for "what do you charge for a fade and are you open Saturday." At xAI's rates that is 1.5 minutes at $0.06, or $0.09 per business called, and a comparison shop across five businesses is $0.45 on the answering side. At the ElevenAgents additional-minute rate the same call is 1.5 minutes at $0.080, or $0.12 per business and $0.60 across five — but those figures exclude the LLM and telephony that ElevenLabs bills separately, so treat them as a floor rather than a total.⁸⁴

These are list-price illustrations, not a claim about what anything costs Google, which runs its own stack and publishes no per-call figure. The point is the shape rather than the decimal: a structured inventory lookup costs effectively nothing per query, and the audio fallback costs cents per participant per attempt, on both ends, every time. Multiply by a query volume Google describes as more than doubling quarterly¹ and the case for publishing a feed writes itself.

Why the phone, when an agent-to-agent protocol already exists

The awkward part is that the industry already built the thing that would make these calls unnecessary.

The Agent2Agent protocol, hosted by the Linux Foundation, marked its first year on April 9, 2026 with more than 150 supporting organizations, integration across the Google, Microsoft and AWS platforms, and production deployments spanning supply chain, financial services, insurance and IT operations.⁹ The specification reached a stable 1.0 with Signed Agent Cards for cryptographic identity and multi-tenancy support.⁹ We walked through the mechanics of this in our A2A protocol multi-agent TypeScript tutorial.

None of that reaches a two-chair barbershop. The adopters named in the Linux Foundation's announcement are enterprise software and cloud vendors, and the Foundation says nothing about small-business adoption in that release.⁹ Both parties to an A2A conversation need to be running agent infrastructure and exchanging an Agent Card. The neighborhood salon has a phone number, a Google Business Profile, and possibly a booking widget. It does not have an endpoint an external agent can discover and call.

So the long tail gets served by the lowest common denominator. A phone number is the interface a small business is most reliably addressable and answerable on, which makes it the de facto agent transport for the segment that most needs one. This is the same gap that agent payment standards are trying to close from the money side — a race we tracked in the x402 agent payments standards fight. Identity, payment and now conversation are each getting a protocol; the local long tail is on none of them.

There is a plausible read where this is fine, even good. Voice is a genuinely universal interface, it requires zero onboarding from the business, and a shim that works today beats a standard a barbershop adopts a decade from now. The counter-read is that every cent and every second spent transcoding structured intent into speech and back is pure overhead, and that overhead scales with adoption. Both can be true. What is hard to argue is that audio is the right long-run substrate for two programs exchanging four facts.

The TCPA layer under all of this

Much of the search coverage of this feature is local-SEO advice: be ready, answer fast, train your staff. Less of it engages with the statute that governs automated voice calls in the United States.

On February 2, 2024 the FCC adopted a Declaratory Ruling — released on February 8, 2024, and effective on release — confirming that the Telephone Consumer Protection Act's restrictions on "artificial or prerecorded voice" reach current AI technologies that generate human voices.¹⁰ The document is FCC 24-17 in CG Docket No. 23-362, and its caption block reads "Adopted: February 2, 2024 Released: February 8, 2024" — two distinct dates worth keeping straight, since secondary coverage tends to cite only the release date.¹⁰

The ruling matters here for two reasons beyond the headline.

First, the scope it quotes is narrower than "all calls." The FCC cites 47 U.S.C. §227(b)(1)(B), which prohibits initiating "any telephone call to any residential telephone line using an artificial or prerecorded voice to deliver a message without the prior express consent of the called party," and §227(b)(1)(A), which covers non-emergency automated or artificial-voice calls to "certain specified categories of telephone numbers including emergency lines and wireless numbers."¹⁰ A business landline is neither a residential line nor a wireless number. A small business that publishes a mobile number as its main contact — common in home repair, beauty and pet care, which are precisely the launch categories — is in a different position from one that publishes a landline. This piece is not legal advice, none of this should be read as a conclusion that calls to businesses are exempt, and how these provisions apply to a platform's agent calling a business on a consumer's behalf is not something the Commission has squarely addressed in the material cited here.

Second, the identification rule is already on the books. The ruling quotes 47 CFR §64.1200(b)(1): "All artificial or prerecorded voice telephone messages shall: (1) At the beginning of the message, state clearly the identity of the business, individual, or other entity that is responsible for initiating the call."¹⁰ The Commission also rejected the idea that a human in the loop launders an artificial-voice call, citing its earlier Soundboard Ruling that a live agent selecting prerecorded messages "does not negate the clear statutory prohibition against initiating a call using a prerecorded or artificial voice."¹⁰

What has not landed is the AI-specific disclosure regime. The Commission's Notice of Proposed Rulemaking in the same docket, FCC 24-84, published in the Federal Register on September 10, 2024, proposed defining "AI-generated call" and requiring disclosure both at the point of consent and at the start of the call.¹¹ As of August 4, 2026, searches of the Commission's public documents and of practitioner commentary dated April and May 2026 surface no final rule adopted in that docket; the proposal appears to still be a proposal. Treat that as a bounded, dated observation rather than a permanent state of affairs — dockets move, and this one has a two-year record behind it. Regulators elsewhere are moving on agent conduct on their own timelines, as we covered in China's AI agent regulations.

The practical asymmetry is worth stating plainly. The calling side of this exchange operates under a federal robocall regime built in 1991 and extended to AI voices in 2024. The answering side — a business deploying a voice agent to pick up its own inbound calls — falls outside the provisions quoted above, which are framed around initiating a call rather than answering one.¹⁰ That is an observation about the text the FCC cites, not a compliance opinion; other rules, including state ones, may still apply.

What a business on the receiving end can do this week

Four concrete moves. The mechanics of each come from Google's own documentation; where a consequence is inferred rather than documented, it is labelled as such.

Publish structured data so the call never happens. If in-store product details are in Google Merchant Center, Google says it will use that data and "will not call your business to ask about inventory."² For bookings, Google will route through an online booking partner "if available" before it dials.² Feed beats phone.

Decide deliberately whether to opt out — and read the fine print on what it does. Google's controls live under Business Profile settings, then Advanced settings, then "Google automated calls and text messages," with separate toggles for "Bookings and inquiries from customers," "Keep your profile up to date," and "Post to your profile on your behalf."² An unverified business must verify before it can opt out.² A verbal refusal on the call also works: Google's page suggests wording such as "Please stop calling my business."² Google adds that "after opting out, you may still occasionally get manual calls from Google to confirm your business information."² Google does not document what the requesting customer sees when a business has opted out; the reasonable inference is that the summary is assembled from whoever did respond, but Google has not said so.

Handle the phone-tree case separately. Google documents a distinct opt-out for calls made to map a business's phone tree: call +1-650-206-5555 and leave a voicemail with the business name and phone number.²

Assume the call is recorded. Google states plainly that "the calls are monitored and recorded for quality assurance."² Anything an employee improvises about pricing on one of these calls is captured, summarized, and delivered to a customer next to competitors' answers.

The bottom line

The story here is not that Google built a robot that phones a barbershop. It is that in August 2026, AI agents calling businesses has become normal enough that both ends of the call are automating independently, and the interface they meet on is the oldest one available.

Google's own documentation says the quiet part: when a structured feed exists, the agent reads the feed and does not dial.² Every customer-triggered call is a measurable cost — $0.06 a minute all in at xAI's published rate, and at least $0.080 a minute before LLM and telephony at ElevenLabs' list price⁸⁴ — incurred because a machine-readable endpoint was missing. The agent-to-agent protocol that would remove that cost shipped a stable 1.0 and cleared 150 organizations,⁹ and no public source documents it reaching the kind of business generating these calls.

For a business owner, the decision this month is narrow and practical: publish the structured data, or answer the phone well. For everyone building agents, the lesson is less comfortable. When your agent has to speak out loud to another agent, that is not a capability milestone. It is a bill for a missing interface, and someone is paying it by the minute.