Get AI summaries of any video or article — Sign up free
I found an AI Agent that makes Phone Calls for you thumbnail

I found an AI Agent that makes Phone Calls for you

MattVidPro·
5 min read

Based on MattVidPro's video on YouTube. If you like this content, support the original creators by watching, liking and subscribing to their content.

TL;DR

Phone Call GPT can place outbound calls and conduct natural-sounding conversations using a user-defined identity prompt and conversation goal.

Briefing

Phone Call GPT is an AI service that can place realistic, voice-to-voice phone calls on a user’s behalf—handling conversations, collecting information, and even completing simple transactions—while charging per call through a credit system. In practical tests, it successfully ordered a large cheese pizza and later completed a scripted pickup conversation for a “40ft tall lemon statue,” showing that careful prompting can drive reliable outcomes.

The setup runs through the Freedom GPT website, where users choose from multiple large language models (including OpenAI options and Dolly 3), then enter an identity prompt (who the caller is) plus a goal for the call. Users can add one or more phone numbers and run simultaneous calls. The interface also lets users select from multiple voice options and generate a conversation starter. A key operational detail is that the AI can record calls if the user enables a quality-assurance disclaimer that informs the other party the call is being recorded.

In the pizza test, the AI placed the call and conducted a natural-sounding exchange: confirming the business (“Freedom Pizza”), taking the order, and repeating the total (32.88) and customer name (Matthew). However, it also showed failure modes that matter on real phones. It made an incorrect assumption about add-ons (it offered fries and drinks without the same level of confirmation a human would require), and it did not automatically hang up when the order was complete—requiring manual intervention or relying on the other party to end the call.

A second scenario improved results through stricter instructions. When prompted to deny any extra items and to stick only to pickup of the lemon statue, the AI handled the conversation cleanly: it confirmed the address (835 Production Avenue), stated a readiness window (“in 3 days or so”), and refused the “extra lime juice fountain” when asked. This time, the call ended automatically, suggesting that stronger constraints and clearer refusal rules can reduce drift.

Beyond scripted transactions, the service is positioned for business automation—customer service callbacks, appointment scheduling, and outbound follow-ups—plus personal use cases like ordering food, contacting businesses when someone is busy or driving, or helping people who struggle with phone communication. The transcript also flags a major near-term limitation: while the text-to-speech sounds convincing, the emotional delivery can swing unnaturally, and transcription/understanding can hiccup when the other side speaks uncertainly.

Pricing is presented as enterprise-focused at $1,000 a month, with 75 credits per call and a pay-as-you-go option. The lowest bundle is 1,000 credits for $10, roughly equating to about a cent per credit, translating to around 7–12 calls depending on credit usage. The overall takeaway is that Phone Call GPT is already functional for structured tasks, but performance depends heavily on prompt quality—and future gains likely hinge on better voice-to-text accuracy, more controlled “robot” expressiveness, and smoother call lifecycle handling.

Cornell Notes

Phone Call GPT places outbound phone calls using a chosen voice and a prompt that defines the caller identity and the conversation goal. In tests, it successfully ordered a pizza and later arranged a pickup for a “40ft tall lemon statue,” including refusing add-ons when instructed. Results improved when prompts explicitly denied extras and constrained the conversation. The system can record calls with a quality-assurance disclaimer, and it supports multiple phone numbers and simultaneous calling. While speech sounds realistic, transcription and emotional delivery can be inconsistent, and call hang-up behavior may require attention.

How does Phone Call GPT decide what to say during a call?

Users provide two main prompt elements: an identity (e.g., “You are Matthew”) and a goal (e.g., placing an order for a large cheese pizza for pickup). It also uses a conversation starter (like “Hey, is this Freedom Pizza?”) and can be configured with a selected voice. The transcript emphasizes that the AI’s behavior depends heavily on how carefully the prompt constrains what it should and shouldn’t do—especially when asked about extras.

What went right in the pizza-order test, and what went wrong?

It reached the business, confirmed the order context, and completed a coherent ordering flow: it asked about sides and drinks, produced a total of 32.88, and repeated the order name as Matthew. The problems were practical: it made an assumption about add-ons (offering fries and drinks) without the same level of confirmation a human would likely require, and it failed to automatically hang up at the end of the transaction.

Why did the lemon-statue scenario perform better?

The second test used stricter instructions: it required denial of any extra items and limited the conversation to pickup of the statue only. When the business offered an “extra lime juice fountain,” the AI refused appropriately. It also confirmed the pickup address (835 Production Avenue) and stated a readiness timeline (“in 3 days or so”). This time, the call hung up automatically, suggesting the prompt constraints reduced conversational drift.

What technical or UX limitations show up in the transcript?

The voice-to-speech sounds realistic, but emotional delivery can be erratic, and transcription/understanding can hiccup—especially when the other party’s speech is unclear. The transcript also notes that hang-up behavior wasn’t consistent: one call didn’t end automatically, while the next did. These issues matter because phone interactions require both accurate comprehension and clean call termination.

What are the main use cases suggested, and how do they differ for businesses vs. individuals?

For businesses: automating inbound/outbound calls, scheduling appointments, and making callback or notification calls. For individuals: ordering food, contacting businesses when busy (e.g., driving or in meetings), and potentially helping people who are mute or have extreme social anxiety. The transcript also suggests a near-term integration path where a chat interface coordinates the phone agent and collects missing details from the user.

Review Questions

  1. In the transcript’s tests, what specific prompt changes led to better refusal behavior (denying extras) in the lemon-statue call?
  2. What two categories of issues appear to limit reliability on real phone calls (speech/emotion vs. transcription/turn-taking), and how did they show up in the examples?
  3. How does the credit/pay-as-you-go model affect how you might estimate cost per call compared with a flat monthly enterprise plan?

Key Points

  1. 1

    Phone Call GPT can place outbound calls and conduct natural-sounding conversations using a user-defined identity prompt and conversation goal.

  2. 2

    Simultaneous calling is supported by allowing multiple phone numbers, and users can choose from multiple voice options.

  3. 3

    Call recording is possible via a quality-assurance disclaimer, which informs the other party the call is being recorded.

  4. 4

    In a pizza-order test, the AI completed the transaction but made questionable assumptions about add-ons and did not auto-hang up.

  5. 5

    In a constrained lemon-statue test, clearer instructions (including denying extras) improved accuracy and led to automatic hang-up.

  6. 6

    The service is positioned for business automation (customer service, scheduling, callbacks) and personal tasks (ordering, contacting businesses when busy).

  7. 7

    Pricing combines an enterprise monthly tier with a pay-as-you-go credit system, with 1,000 credits for $10 cited as the cheapest bundle.

Highlights

Phone Call GPT completed a full pizza ordering exchange, including a stated total of 32.88 and the customer name Matthew—showing it can handle structured transactions.
Prompting constraints mattered: when told to deny any extras, the AI refused an offered “lime juice fountain” during the lemon-statue call.
Realistic voice output is a strength, but emotional delivery and transcription hiccups can still undermine reliability on live calls.

Topics

Mentioned