ResearchVoice agents

Who is calling?

AI agents are dialing into clinics and health systems. Some pass as human. Some speak for patients who never agreed. The first call swarms have already hit, and the front desk has no way to tell friend from fraud.

Published
November 2026
Type
Position paper
Topic
AI safety, patient access
Read
10 minutes
Scroll
Incoming callFront desk · 9:02 am
  • Is the phone number real?Can be checked
  • Does the caller know the patient's details?Can be checked
  • Is the caller a person or an agent?Guesswork
  • Did the patient authorize this caller?No way to check
What a clinic can and cannot establish about a caller today.
Summary

For as long as clinics have had phones, the person calling was a person. That is no longer a safe assumption.

In 2026 consumer AI agents learned to use the telephone. Meta's Muse now places calls to U.S. businesses for paying subscribers, and the examples Meta gives include insurance companies.1 Instinct and Google's Gemini offer something similar.1 Ferry Health launched in September with an agent that calls clinics, checks coverage and books appointments.2

These agents say they act for a patient. Most probably do. But a clinic that picks up the phone has no way to confirm it. The checks a front desk runs today test what the caller knows, and an agent holding a patient's details knows all of it.

10million +

people using Meta's Muse agent, which can now place phone calls for paying subscribers.

Meta, via TechCrunch [1]

100%

of AI agent calls are ended on detection at one large health system we spoke with. Its policy is to hang up, whoever the agent is calling for.

Arul provider survey [13]

0standards

in use today for an AI agent to prove, on a phone call, whom it is acting for.

Our review, Part III

This paper makes three points. Agent calls to healthcare are already arriving, and nobody is counting them. The current response, which is to hang up, fails patients with honest agents and does little to stop dishonest ones. And the gap can be closed, because the web has already solved a version of this problem. We propose a four-part framework, DIAL, for how a voice agent should identify itself and prove a patient's consent before a clinic acts on what it says.

Part I

Agents picked up the phone

The idea is not new. Google showed an assistant phoning a hair salon in 2018. What changed in 2026 is that the capability reached ordinary consumer products, at consumer prices, within a few months (Figure 1).

Figure 1Public announcements and reporting

In one autumn, calling a business became a feature of consumer AI

  1. Google demonstrates DuplexAn assistant books a salon appointment by phone. It stays a narrow feature for years.
  2. FCC rules AI voices are "artificial" under the TCPAThe ruling targets robocalls made to consumers. It does not address agents calling businesses for a consumer. [4]
  3. Muse and Instinct begin testing callsMeta opens calling to selected early users. Merchants hang up often enough that Meta tries human callers. [1]
  4. Ferry Health launches with $9MIts agent calls clinics to verify availability and coverage and to book visits, offered through health systems, plans and navigation vendors. [2]
  5. OpenAI launches dotsAlways-on agents that users can phone. At launch they do not place calls to businesses. [3]
  6. Muse calling opens to all paying U.S. subscribersFrom $20 a month. Meta's examples: the cable company, the DMV, a restaurant, an insurance provider. [1]

Purple pins are events that put agent-placed calls in front of healthcare organizations. This is a selection of public events, not a complete list.

Healthcare was not the target of most of these launches. It is the obvious destination. Booking an appointment, checking whether a doctor takes your plan, chasing a referral, disputing a bill: these are the calls people least want to make, and much of healthcare still runs on them.

We saw this from the other side. Over 2026 we surveyed 110 practices and health systems. One of them, the patient access team at a large academic health system, described it to us in detail in November. They first noticed calls from AI agents two to three months earlier. They had reviewed three or four recordings. They did not know the real volume, and believed it was undercounted, because nothing in their phone system flags an agent.13

Part II

The phone is still healthcare's front door, and it was already strained

Agent calls are arriving at a part of healthcare that was struggling before they showed up. Reported hold times at healthcare call centers run several times the industry's own target (Figure 2). Medical groups know it. In an MGMA poll, more than one in five named phone access as their top patient-access focus for 2026.10

Figure 2Average time on hold · seconds

Callers wait about five times longer than the target

Reported averageU.S. healthcare call centers264s
CMS limitMedicare Advantage and Part D plans120s
Industry targetHFMA50s

Sources: [12].

Providers are responding with automation of their own. Three quarters of health systems in one 2026 survey had deployed at least one AI tool, and calls sit near the top of the front-office automation list (Figure 3). Meanwhile practices still spend about 13 hours of physician and staff time a week on prior authorization alone, much of it on the phone with insurers.11

Figure 3Share of respondents, four separate surveys

Providers are automating their own end of the line

Health systems with at least one AI tool deployed120 health systems, Feb 2026 [9]75%
Large systems using AI agents for eligibility900+ beds, same survey [9]55%
Medical groups: calls are the top automation priorityMGMA poll, Feb 2026 [10]27%
Medical groups: phone access is the top access focusMGMA poll, Dec 2025 [10]22%

Each bar is a share of its own survey, so the bars are not directly comparable with each other.

Put the two trends together and the direction is clear. Patients are sending agents to call. Clinics and insurers are deploying agents to answer. Within a few years a large share of healthcare phone calls may have software on both ends, speaking English to each other over a channel designed for people, with no agreed way to say who they are.

Part III

The checks we have test knowledge, and agents know everything

Before a clinic discusses a patient with a caller, privacy rules require it to verify who the caller is and whether they have the authority to ask.5 In practice that means a short list of questions. Follow one call to see where it stops working.

Call in progressFigure 4
Caller ID: valid number, not spoofed
CallerHi, I'm calling to book a follow-up visit for a patient of the practice.
Front deskCan I have the name, date of birth and ZIP code?
CallerOf course. [Gives all three correctly, then the street address.]
Knowledge checks passed: 5 of 5 · Authority established: none
Front desk suspects an AI caller and ends the call.
An illustrative call, based on the verification steps providers described to us. [13]
The ring

The number checks out.

Phone networks can confirm that a number has not been spoofed. That is all they confirm. A fraudster can rent a clean, verified number as easily as anyone else.7

The questions

The caller knows the answers.

One large health system we spoke with verifies callers with caller ID, name, date of birth, ZIP code and address. A patient's own agent has all of these. So does anyone who has bought a leaked record.

The gap

Three different callers look identical.

An agent the patient set up. An agent a relative set up without asking. A script run by a fraud ring. All three pass the same five checks, and none of them has shown that the patient agreed to the call.

The outcome

So the clinic hangs up.

Staff are trained to listen for callers that do not sound human, and the policy is to end those calls. The patient with an honest agent loses the appointment. The dishonest caller tries again with a more convincing voice.

Listening for a synthetic voice is a short-term defense. Reporting on Muse notes that AI voices are getting harder to tell from human ones,1 and some builders are working to make their agents pass as human on purpose. Detection can catch today's common voice models. It cannot say anything about consent (Figure 5).

Figure 5Our assessment of each check

Nothing a front desk can ask today shows that the patient agreed

CheckCaller knows the patient's detailsWho operates the callerPatient authorized this caller
Caller IDNumber is valid, not spoofedNoPartlyNo
Name and date of birthYesNoNo
ZIP code and addressYesNoNo
Listening for a synthetic voiceNoPartlyNo
Faxed release signed by the patientPossible today, easy to forgeNoNoPartly
An agentic healthcare proxySigned by the patient; see Part VIYesYesYes

"Partly" means the check gives a signal that can be faked or that identifies a number or a voice model, not an accountable operator. This is our judgment, not a measured result.

The web has met this problem already. Websites needed to tell a real shopping agent from a scraper, and the answer was cryptographic: agents sign their requests, and a site can check the signature against a published key. Web Bot Auth does this for crawlers and agents, and Visa's Trusted Agent Protocol builds on it to tie an agent to a specific consumer's instruction.8 The phone has caller-ID attestation, which vouches for a number and says nothing about who is using it or why.7 We found no equivalent standard for voice.

Part IV

The risks are not hypothetical

Voice fraud against call centers was climbing before consumer agents arrived. The figures below come mostly from a security vendor, so read them as that vendor's view. They still describe the environment these new callers are entering.

$893million

in reported losses to AI-enabled fraud in 2025, across 22,364 complaints. The FBI tracked the category separately for the first time.

FBI, as cited by Pindrop [6]

50% +

of the time, fraudsters get past knowledge-based questions, according to one vendor's analysis.

Pindrop [6]

4500calls

in a single reported attack on a healthcare organization, exposing $18 million in account value before it was contained.

Pindrop [6]

We see four distinct risks.

  • Disclosure to the wrong party. An agent that passes the knowledge checks can be told a patient's appointments, results or balance. If the patient never authorized it, that is a privacy breach, and the clinic made it.
  • Fraud at scale. A person can make a few dozen fraudulent calls a day. An agent can make thousands. One health system we spoke with already treats fraudulent portal calls as a known problem, and worried about agents being pointed at Medicare and Medicaid billing.
  • Flooding. Call volume itself becomes an attack. We heard one first-hand account of an overnight wave of automated calls that overwhelmed a healthcare company's referral line.
  • Honest agents locked out. When the only safe policy is to hang up, the people hurt most are the ones who most need help making calls: older adults, people who are very sick, and caregivers managing someone else's care.

The last risk is easy to miss, and it is the one we care about most. A blanket refusal protects the clinic. It does nothing for the patient who cannot spend an hour on hold.

Part V

The opportunity is as large as the risk

It would be a mistake to read this paper as an argument against voice agents. An agent that can wait on hold, book the visit, confirm coverage and chase the referral gives time back to patients and to front-desk staff alike. That is the work Arul exists to do.

That health system expects that, over time, these interactions will move to digital channels that are easier to secure. We agree with the direction. But the phone will not disappear in the meantime. Many independent practices have no other front door, and agents will go where the appointment is.

The position in one line

The goal is not to keep agents off the phone. It is to make an honest agent provable, so that a clinic can say yes with confidence.

Part VI

DIAL: a framework for voice agents in healthcare

We propose four requirements for any voice agent that calls a healthcare organization on a patient's behalf. Each one answers a question the front desk cannot answer today. We are building Arul's own agents to meet them, and we are publishing them so that others can test them, criticize them and improve them.

  1. D
    Disclose

    The agent says it is an AI, who operates it, and whom it is calling for, in its first sentence. It never tries to pass as a person.

    Answers: is this a person?
  2. I
    Identify

    The operator is a known, accountable organization. The agent presents a credential the clinic can check against a public registry, not only a phone number.

    Answers: who runs this caller?
  3. A
    Authorize

    The agent carries an agentic healthcare proxy, signed by the patient with their legal signature. It says what the agent may do, for which provider, until when. The patient can revoke it at any time.

    Answers: did the patient agree?
  4. L
    Log

    Both sides keep a record of what was asked and what was done. The patient can see every call made in their name.

    Answers: what happened?

The agentic healthcare proxy

The A in DIAL rests on one document. We call it an agentic healthcare proxy: a form in which a patient authorizes a named AI agent to carry out specific administrative tasks for them. It is Arul's own form, written for this purpose.

  • A legal signature a clinic can check. The patient signs with their legal signature. A clinic can match it against the signature on the patient's driver's license, or against medical documents it already holds that carry the same signature.
  • Specific and limited. The form names the agent and its operator, the tasks allowed, the provider, and an end date. The patient can revoke it at any time.
  • Not a healthcare proxy. A healthcare proxy names a person to make medical decisions when the patient cannot. An agentic healthcare proxy gives software permission to handle administrative work while the patient stays in charge. It grants no authority over medical decisions.

How much proof, for which request

Not every call needs the same scrutiny. Asking whether a clinic is taking new patients is not the same as asking for lab results. We suggest matching the proof to the request (Figure 6).

Figure 6Proposed, for discussion

The more a request reveals or changes, the more the agent must prove

TierExample requestsWhat the clinic should requireDIAL
0 · PublicOpening hours, location, whether the practice takes a plan or new patientsDisclosure only. Nothing about a patient is shared.D
1 · SchedulingBook, move or cancel a visit for a named patientA known operator and an agentic healthcare proxy that covers scheduling.D I A L
2 · Protected informationResults, balances, records, referral statusEverything in tier 1, plus the patient confirming in the moment, on the line or on their own device.D I A L + patient
3 · Not by agentChanging contact details or portal access, controlled-substance requestsThe patient, in person or through an authenticated digital channel.None

Tiers and examples are a starting point. Each organization's privacy and compliance teams decide where its own lines sit.

How it could work on a real call

None of this requires new telephone technology. The options below came up in our work with providers' access and security teams, in rough order of how soon they could be used.

  • Patient on the line. The agent waits on hold and, when a person answers, brings the patient into the call to be verified in the usual way. Slow, but it works today.
  • A confirmation on the patient's device. Mid-call, the clinic sends a text or app prompt to the number on file, and the patient approves the specific request.
  • A credential exchanged before the conversation. The agent presents a signed credential at the start of the call, the way a browser presents a certificate, and the clinic's phone system checks it before a person or an agent picks up.
  • A digital door for agents. The clinic answers agent calls with a prompt that sends them to an authenticated digital intake, where consent can be checked properly.

A legal model already exists for the idea underneath all four. A patient can name a personal representative, and a clinic must treat that person as it would the patient once their authority is verified.5 An agentic healthcare proxy is a narrower, time-limited version of the same idea. Whether the law will treat it that way is an open question, and one for regulators and counsel, not for us.

What comes next

What we are doing, and what we are asking for

  • Measuring. Nobody we spoke with could say how many agent calls they receive. We are working with provider partners on simple proxies, starting with the pattern of short calls ended by staff.
  • Building to DIAL. Arul's agents disclose that they are AI, and patients approve anything sent in their name. We are extending that to a credential and an agentic healthcare proxy a clinic can check.
  • Sharing what we learn about detection. We will publish our methods for recognizing synthetic callers, with their error rates, in a follow-up note.
  • Asking for partners. If you run patient access, security or compliance at a provider or a plan, or you build voice agents, we would like to test this framework with you.

Work with us on DIAL

Sources and limits

Where the evidence comes from, and how far to trust it

Limits

  • This is a position paper, not a study. We did not measure agent call volume. As far as we can tell, nobody has.
  • Our field evidence is a survey of 110 practices and health systems. It covers independent practices and large health systems. The account we quote in detail comes from one health system.
  • Several figures come from vendors. Fraud statistics from a security company and adoption statistics from secondary reports serve those authors' interests. We say so where we use them.
  • The framework is untested. DIAL is a proposal. It has not been piloted with a provider.
  • We have a stake. Arul builds agents that act for patients. We and our patients benefit if honest agents are accepted.

Sources

  1. S. Perez, "Muse can now place calls to U.S. businesses, if you pay," TechCrunch, 12 November 2026. techcrunch.com
  2. "Ferry Health, agentic AI scheduling platform backed by a16z, launches from stealth," Fierce Healthcare, September 2026. fiercehealthcare.com
  3. "OpenAI Announces Always-On AI Agents Named Dots," OfficeChai, 2026. officechai.com
  4. FCC Declaratory Ruling, CG Docket No. 23-362, 8 February 2024, as summarized by Cooley. cooley.com
  5. HIPAA Privacy Rule, 45 CFR 164.514(h) (verification) and 164.502(g) (personal representatives). ecfr.gov
  6. Pindrop, healthcare voice-fraud articles citing its Voice Intelligence and Security Report and the FBI's 2025 Internet Crime Report. pindrop.com Vendor source.
  7. First Orion, "Why STIR/SHAKEN Cannot Stop AI Voice Fraud and How Call Authentication Can." firstorion.com Vendor source.
  8. Akamai, "Redefine trust: web bot authentication," November 2025, on Web Bot Auth and Visa's Trusted Agent Protocol. akamai.com
  9. "75% of Health Systems Have Adopted AI," Agentman, reporting a February 2026 survey of 120 health systems. agentman.ai Secondary source.
  10. Parlance, "Healthcare AI Falls Short on Patient Phone Access," citing MGMA Stat polls of December 2025 and February 2026. parlancecorp.com Secondary source.
  11. American Medical Association, prior authorization physician survey. ama-assn.org
  12. Commure, "Healthcare Call Center Metrics: Definitions and Where Each Target Comes From." commure.com Vendor source.
  13. Arul Health, survey of 110 practices and health systems, 2026. The detailed account is from the patient access and security teams of a large U.S. academic health system, November 2026. Organizations are not named.
Cite this paper Arul Health Research (2026). Who is calling? Voice agents at healthcare's front door. https://arulhealth.org/research/who-is-calling

Arul is not a medical service and does not give medical or legal advice. Product names belong to their owners and are mentioned for identification only. Questions: support@arulhealth.com.

More from Arul Health Research

All research