- Dr. Serdar Özcan
- 0 Comments
- 213 Views
AI agents have started calling businesses. Our view: AI should answer the phone and leave the dialing to people.
In the space of two days, AI agents that make phone calls reached consumer betas and early access, and a regulator moved to make automated calls register with phone operators first. On September 17 Meta's Muse and the startup Instinct both added outbound calling to their consumer agents. A day later India's telecom regulator finalised rules under which calls made with artificial voices must be declared in advance, or be treated as spam. At TAO AI LAB we build assistants that answer calls, and we have a clear position on the other direction.
|
730K+
MUSE DOWNLOADS IN THE US
|
$10B
INSTINCT'S REPORTED VALUATION
|
42B
SPAM CALLS, TRUECALLER USERS, 2025
|
12B
OF THEM BLOCKED BY TRUECALLER
|
What happened on September 17
Two consumer AI agents gained the same ability on the same day. Meta's Muse can now place calls to U.S. businesses. Meta's Ryan Fox announced on X that the company had "expanded the Muse beta for outbound calls to US businesses," and the feature goes first to users who ask Muse to make a call.
Instinct launched Instinct Concierge, which founder Noah Shinn described as "a white glove service." TechCrunch lists what it can do: book restaurants that don't take online reservations, get on a dentist's cancellation list, sort out a cable bill. It is rolling out in early access.
The interest and money behind this are large. According to TechCrunch, Muse passed 730,000 U.S. downloads, while the Meta AI app was downloaded 707,000 times in its first five days after launch. Instinct raised $350 million at a $2.5 billion valuation in August, and TechCrunch says it now draws investor interest at a reported $10 billion valuation.
The question the announcements leave open
Neither announcement, as reported, says what the person who picks up the phone hears. Does the agent say it is an AI? Does it say who it is calling for? TechCrunch's coverage does not address disclosure, and we could not find it in the other reports we read.
That gap matters to us more than the feature itself. A restaurant host answering the phone at 7 pm has a few seconds to decide how to treat the caller. They deserve to know whether a person or a program is on the line.
India draws a line
On September 18 India's Telecom Regulatory Authority (TRAI) finalised the Telecom Commercial Communications Customer Preference (Third Amendment) Regulations, 2026. Once they are formally notified, calls "initiated by an application, software system or automated platform without direct human dialing" will fall under an application-to-person (A2P) framework, which explicitly includes robocalls and "pre-recorded/artificial voice technologies." MediaNama reports that the regulations have not yet been notified and take effect in stages after that.
Anyone using such calls must declare it, along with the numbers involved, to their telecom operator in advance. Calls made without that declaration will be treated as spam. The scale explains the urgency: Truecaller says its users in India faced around 42 billion spam calls in 2025, and that it blocked nearly 12 billion of them, according to TechCrunch.
The part that interests us most is the person answering. ETV Bharat reports that robocallers "will now have to declare upfront that the call is an automated one," so recipients can see before accepting whether a human or a machine is calling. TechCrunch and MediaNama describe only the declaration to the operator, so we are taking this detail from one report. If it holds, India has put into regulation the principle we argue for below.
Our position: AI should answer, people should dial
This is an opinion, and we want to label it as one. We find an AI that starts a call off-putting. When a synthetic voice rings you with something to sell, it takes your time on terms you never agreed to. We don't think AI should be making cold calls.
The agents announced this week are a different case, and we should be fair about that. Muse and Instinct call because their user asked them to book a table or fix a bill. That is an errand, and it is closer to a personal assistant than to a call centre. Our discomfort is smaller there. It does not go away, because the person answering still did not choose to talk to software.
Answering is where we think AI earns its place. A business that misses calls at lunch or after hours can lose the people who were trying to reach it. An assistant that picks up, says clearly that it is an AI assistant, handles the request and hands over to a person when needed serves someone who chose to call. The consent sits with the caller. That is the direction we build in.
What changes for a business phone line
If these agents spread, some of the calls a business receives will come from software acting for a customer. We think that raises practical questions for anyone running a phone line. Can your staff tell when an agent is calling? Will you take a booking from one? What happens when you ask it a detail its user never gave it?
We don't have settled answers, and we are not going to pretend otherwise. In our view, the first seconds of a call set the tone. Saying clearly who and what is on the line costs a sentence and saves confusion on both sides.
TAO AI LAB note
Editor's comment
We write about voice assistants often, because voice is a field that moves quickly. It is one of four areas we work in. The others get less space here, and they deserve a mention:
- Voice AI assistants and virtual receptionist solutions: assistants that answer real business phone lines, in Turkish and English.
- AI-supported social loneliness and anxiety management applications: digital wellbeing tools designed around evidence-based methods such as CBT, with a bridge to a human specialist when needed.
- Prediction models on big data: machine and deep learning for time series forecasting, anomaly detection and risk scoring.
- TUBITAK 1707 project support and R&D partnership.
The principle this week's news raised already appears in our own design work: people should know what they are talking to. Our wellbeing application design states openly that it does not make a medical diagnosis and does not replace treatment. An AI that answers the phone should be just as clear about what it is. We keep building with that principle at the centre.
What this does not settle
Three limits. First, we could not open Meta's or Instinct's own announcements; the details come from TechCrunch and the posts it quotes. Second, TRAI's press release page did not load for us, so the rule's wording comes from Indian press coverage, and only one of those reports says the person answering will be told the call is automated. Third, Turkey has no equivalent rule that we know of, and any guess about whether one will follow would be speculation.
Frequently asked questions
Can AI agents make phone calls for you now?
Yes, in limited form. Since September 17, 2026 Meta's Muse can call U.S. businesses for users who ask, and Instinct Concierge is in early access for tasks like restaurant bookings.
What did India decide about AI voice calls?
TRAI's rules, finalised on September 18, 2026 and not yet formally notified, classify calls using artificial voices as A2P calls. They must be declared to the operator in advance, or they will be treated as spam.
Is an agent calling on your behalf the same as a cold call?
We don't think so. It runs an errand you asked for. The person answering still deserves to be told it is an AI.
What does a virtual receptionist do differently?
It answers calls people chose to make, says it is an AI assistant, handles the request and passes the call to a person when needed.
We build AI that picks up the phone, and we would rather leave the dialing to people. One question for you: if an AI called your business tomorrow to book a table for a customer, would you want it to say so in the first sentence? Tell us in the comments.
Sources
- TechCrunch, "Rival AI agents, Instinct and Meta's Muse, both add the ability to make calls" (September 17, 2026)
- TechCrunch, "India forces caller-ID apps to feed spam reports to telcos" (September 18, 2026)
- ETV Bharat, "TRAI tightens regulations on spam callers" (September 18, 2026)
- MediaNama, TRAI Third Amendment Regulations 2026 (September 19, 2026)
- TAO AI LAB, voice AI assistants and virtual receptionist solutions
- TAO AI LAB, AI-supported social loneliness and anxiety management applications
- TAO AI LAB, prediction models on big data
- TAO AI LAB, TUBITAK project support and R&D partnership