Verdict: You can give an AI assistant a real phone number — so it calls businesses to book appointments, answers your incoming calls with a security code, reads from your knowledge base on demand, and updates your calendar — using three core pieces: a voice AI platform, a phone number provider, and an integration layer. The cheapest viable build costs about $1.15/month for the number plus $0.14/minute on Bland AI's Start plan, with no platform fee and no credit card required to start. Vapi ($0.05/min platform fee) and Retell AI ($0.07/min) are the main alternatives if you want more control over the model stack. The real work — and the real cost — is not the minutes; it is the prompting, the tool integrations, and the security layer that keeps your assistant from leaking sensitive data to a stranger who guessed a code.
Last verified: July 31, 2026. Pricing and feature details change often — especially per-minute rates and plan caps. Always confirm on the vendor's pricing page before committing.
- Best all-in value: Bland AI Start — $0.14/min covers LLM, STT, and TTS; no platform fee.
- Best for customization: Vapi — $0.05/min platform fee, but you pay separately for every model layer.
- Best for fast setup: Retell AI — $0.07/min, visual builder, ~600ms latency.
- Phone number: Twilio US local — ~$1.15/month (confirmed primary source).
- Security: A numeric access code is a minimum viable gate; real deployments need more.
What Does an AI Phone Assistant Actually Do?
An AI phone assistant is a voice AI agent with a dedicated phone number that can make outbound calls to real businesses, receive inbound calls from you or your clients, hold a natural conversation, and trigger actions in your other tools — calendar, CRM, email — based on what was said. Unlike a chatbot, it speaks and listens in real time, handling the back-and-forth of a real phone call: clarifying questions, confirming details, and adapting when the other party proposes something unexpected (like "7:00 is booked, can you do 8:00?").
The technology stack behind every modern voice agent has four layers:
- Speech-to-text (STT): converts the caller's speech into text in real time.
- Large language model (LLM): understands the conversation, generates responses, and decides when to call external tools.
- Text-to-speech (TTS): converts the AI's text responses back into natural-sounding voice.
- Telephony: the actual phone number and call routing — usually Twilio, either directly or through the voice platform's built-in integration.
Some platforms bundle all four layers into a single per-minute rate (Bland AI). Others charge a platform fee and let you bring your own STT, LLM, and TTS providers, so you pay for each component separately (Vapi, Retell AI). This distinction is the single biggest driver of your real cost per call.
How Much Does It Cost to Build an AI Phone Assistant in 2026?
The honest answer is that the advertised per-minute rate is the floor, not the ceiling. A realistic budget breaks down into three categories: the phone number, the AI voice platform, and your time building and maintaining the agent.
Phone Number Costs (Twilio)
A Twilio US local number costs approximately $1.15/month as a monthly recurring charge, confirmed on Twilio's own pricing documentation. Toll-free numbers cost about $2.15/month. You also pay per-minute usage on top: inbound calls to a local number run about $0.0085/minute, and outbound calls within the US cost about $0.014/minute. Call recording adds $0.0025/minute, and transcription costs $0.05/minute if you use Twilio's built-in service rather than your voice platform's. (Source: Twilio Pricing — Phone Numbers resource, Twilio Help Center — How much does a phone number cost)
AI Voice Platform Costs Compared
| Platform | Headline Rate | What's Included | Platform Fee | All-in Estimate | Free Tier |
|---|---|---|---|---|---|
| Bland AI (Start) | $0.14/min | LLM + STT + TTS | $0 | ~$0.14/min | 2 credits + inbound number ($15/mo value) |
| Bland AI (Build) | $0.12/min | LLM + STT + TTS | $299/mo | ~$0.12/min | None |
| Vapi (Build) | $0.05/min | Platform orchestration only | $0 | ~$0.07–$0.25/min | $10 credits |
| Retell AI (Pay-as-you-go) | $0.07/min | Voice engine only | $0 | ~$0.07–$0.31/min | $10 credits, 20 concurrent calls |
| Synthflow | $0.09/min + $0.05/min (GPT-4.1) | Voice engine + GPT-4.1 | Varies | ~$0.14/min | None |
| Ringly.io (Grow) | $349/mo | 1,000 minutes included | Included (~$0.35/min effective) | $0.19/min overage | 14-day trial |
(Sources: Bland AI pricing page, Vapi pricing page, Retell AI pricing page, Ringly.io comparison. Pricing verified July 2026.)
The key insight: Bland AI's Start plan is the cheapest entry point because $0.14/minute is genuinely all-inclusive — it covers the LLM, speech-to-text, and text-to-speech in a single rate. Vapi advertises $0.05/minute, but that covers only its orchestration layer; once you add a speech-to-text engine ($0.01/min), an LLM ($0.02–$0.20/min), a TTS provider ($0.04/min), and telephony ($0.01/min), the realistic all-in cost lands between $0.07 and $0.25 per minute, and can exceed $0.30 with premium voices and large models. (Source: AI Agent Square — Vapi Pricing 2026)
Retell AI sits in between, with a $0.07/minute base rate on its Pay-as-you-go plan that can climb to $0.31/minute depending on your model and voice choices. Its advantage is a visual drag-and-drop builder and native integrations with HubSpot, Twilio, and n8n, and a reported latency of ~600ms — meaningfully faster than some competitors. (Source: Retell AI pricing, Ringly.io platform comparison)
A Worked Cost Example
Say you want an assistant that makes 10 outbound calls per day, averaging 3 minutes each, plus receives 5 inbound calls per day at 2 minutes each. That is about 40 minutes/day, or roughly 1,200 minutes/month.
| Component | Monthly Cost (Bland AI Start) | Monthly Cost (Vapi, standard stack) |
|---|---|---|
| Phone number (Twilio local) | $1.15 | $1.15 |
| Outbound AI minutes (~900 min) | 900 × $0.14 = $126.00 | 900 × ~$0.11 = $99.00 |
| Inbound AI minutes (~300 min) | 300 × $0.14 = $42.00 | 300 × ~$0.11 = $33.00 |
| Telephony per-minute (Twilio) | ~$0.014 × 900 + ~$0.0085 × 300 ≈ $15.15 | ~$15.15 |
| Platform fee | $0 | $0 |
| Total | ~$184.30/month | ~$148.30/month |
Vapi can come out cheaper on paper because you can choose economy models, but the trade-off is that you are responsible for assembling, configuring, and maintaining the STT + LLM + TTS pipeline yourself. Bland bundles it. Retell falls in between. The right choice depends on whether you want to own the stack or have it ready to go.
How Do You Give an AI Assistant a Phone Number?
Giving an AI assistant a phone number requires three steps: getting a real phone number from a telephony provider, connecting that number to a voice AI platform that handles the conversation, and configuring the agent's prompt and tools.
Step 1: Get a Phone Number from Twilio
Twilio is the standard telephony provider behind virtually every voice AI platform. Bland AI, Vapi, and Retell all support "Bring Your Own Telephony" (BYOT) with Twilio, or they offer built-in Twilio integration at pass-through cost. A real phone number — one that shows up on caller ID as an actual number, not a Google Voice number — costs about $1.15/month for a US local number. (Source: Twilio Help Center — How much does a phone number cost?)
To get a number:
- Sign up for a free Twilio account (you get $15 in free credits).
- Search for and purchase a local phone number in the Twilio Console. The number's monthly recurring charge is displayed before you buy.
- Configure the number's voice webhook to point to your voice AI platform's inbound endpoint — this is what routes incoming calls to your AI agent.
Step 2: Connect the Number to a Voice AI Platform
Each platform integrates Twilio differently:
Bland AI: Purchase an inbound phone number directly from Bland's developer portal (they provision it via their built-in Twilio integration). New accounts on the Start plan get an inbound number valued at $15/month included. Alternatively, connect your own Twilio number via BYOT — BYOT customers pay no transfer fees, only the per-minute AI rate. (Source: Bland AI pricing)
Vapi: Vapi offers 10 free US phone numbers on its Build plan. You can also BYO Twilio number and configure it to route to Vapi's webhook endpoint. (Source: Vapi pricing, AITrendTool Vapi review)
Retell AI: Retell charges $2/month for a phone number and offers native Twilio integration. You configure the number's webhook to point to Retell's call-handling endpoint. (Source: Retell AI pricing)
Step 3: Configure Your Agent's Prompt and Tools
This is where the real work happens. A voice agent is fundamentally a prompt with tools. The prompt defines the agent's persona, what it knows, what it can say, and what it must never reveal. The tools let it reach into your calendar, CRM, email, or knowledge base during a live call.
On Bland AI, you configure agents through their developer portal with Conversational Pathways — branching conversation logic that defines how the agent responds in each scenario. You can set up different agents for different purposes: one for outbound appointment booking, one for answering your hotline, one for lead qualification.
On Vapi and Retell, you configure the agent programmatically through their API, or via a visual builder in Retell's case. You specify which LLM to use (GPT-4, Claude, Gemini — Retell supports all three), which TTS voice, and which external tools the agent can call.
What Should the Agent's Prompt Include?
A well-structured voice agent prompt is the difference between an assistant that sounds like a real person and one that sounds like a robot reading a script. Based on the prompts documented across these platforms, an effective voice agent prompt should include:
A persona definition: Who is the agent? "You are Jarvis, a professional personal assistant calling on behalf of [name]." Give it a name, a role, and a tone.
Context about the caller or business: Who is the agent representing? What business are they calling? What information should the agent have ready?
Conversation rules: What should the agent say first? How should it introduce itself ("Hi, this is Jarvis, [name]'s assistant")? What questions is it trying to get answered? What should it do if the person on the other end asks something it doesn't know?
Tool definitions: What external APIs or services can the agent call during the conversation? Calendar booking, CRM lookups, sending a quote — these are defined as tools the LLM can invoke.
Security boundaries: What information must the agent NEVER reveal? Credit card numbers, passwords, personal addresses, or anything sensitive should be explicitly blocked in the prompt. The agent should also require an access code before sharing any information from your knowledge base.
Escalation rules: When should the agent check with its boss before committing to something? (e.g., "I'll check with my boss and call you back" is a common pattern for appointment confirmations.)
The prompting principles here are not unique to voice — the same context-engineering fundamentals apply whether you are writing prompts for a voice agent or a text model. If you want to go deeper on how to write effective prompts for modern models (Claude Opus 5, GPT-5, etc.), see our Claude Opus 5 prompting guide — the effort dial and scope-capping techniques translate directly to voice agent system prompts.
How Do You Secure an AI Phone Assistant?
When you give an AI assistant a phone number, anyone who dials that number can talk to it. You need a security layer that gates access to your information without making the agent unusable for you.
The Access Code Pattern
The minimum viable security pattern is a numeric access code: when someone calls the hotline, the agent asks for a code before doing anything else. If the caller provides the wrong code twice, the agent hangs up and notifies you that an unknown caller tried to access the line.
This pattern is simple to implement in the agent's prompt — but it has real limitations:
- Codes can be guessed, brute-forced, or overheard.
- A 4-digit code has 10,000 combinations, which is trivially exhaustible with automation.
- The code is spoken aloud, so anyone nearby can hear it.
Beyond the Code: Real Security Measures
For production deployments, the access code should be just one layer of a multi-factor approach:
Caller ID allowlisting: Configure your Twilio number to reject calls from numbers not in your approved list. Twilio's TwiML
<Reject>verb can filter incoming calls by caller ID before they ever reach the AI.Short, rotating codes: Use 6+ digit codes and rotate them periodically. Limit to 2 incorrect attempts before the agent disconnects.
Tool-level access control: Even if someone gets past the code, the agent should be configured so that sensitive tools (banking, credit card data, personal address) are never callable through the phone interface. This is enforced in the tool definitions, not just the prompt.
Audit logging: Log every call — who called, when, what was discussed, what tools were invoked. Bland AI and Retell both provide call transcripts and post-call webhooks for this; Twilio's call recording and transcription can supplement it.
Data residency and compliance: If you are handling healthcare data, Bland AI is HIPAA-eligible with a signed BAA on Enterprise plans. Vapi offers HIPAA-compliant options on its Enterprise contract. Retell AI offers compliance documentation on request. (Source: Bland AI pricing, Vapi pricing)
How Do You Connect the Assistant to Your Tools?
The power of a phone-based AI assistant is not that it can talk — it is that it can do things while it talks: check your calendar, look up a client in your CRM, create a quote document, send a status update via Telegram. This is accomplished through tool integrations.
Calendar Integration (Google Calendar)
Your agent needs to read and write calendar entries so it can book appointments and check your availability. This is done through Google Calendar API, exposed to the voice agent as a custom tool. When the agent gathers appointment details from a business on a call, it calls the calendar tool to create an event with the time, location, and a reminder.
CRM and Knowledge Base Access
The agent should be able to look up client information during a call. This means connecting it to wherever your client data lives — Google Sheets, a CRM like HubSpot, or a local knowledge base. When you call the hotline and say "pull up Mike Johnson's information," the agent queries your data source and reads back the relevant details: what service they're interested in, the quote amount, the last interaction date.
On Bland AI, this is done through Custom Tools — external API endpoints you define that the agent can call live during a phone call. On Vapi, you use their custom tools API. On Retell, native HubSpot and Twilio integrations handle much of this out of the box, and you can add custom integrations through their API.
Telegram for Async Communication
Not every interaction needs a phone call. Telegram is a free, reliable alternative for text and voice-note communication with your assistant. You can send the agent a voice note from Telegram, and it can respond with one — useful when you are in a meeting and cannot take a call. This is implemented through the Telegram Bot API, which is free and well-documented, paired with your agent's LLM and TTS to handle voice-note input and output.
Document Generation
When a new lead comes in, your agent can generate a quote document on the fly — pulling from your business's template, filling in the client name, price, and service description, and saving it to your Google Drive. This requires connecting Google Docs and Google Sheets APIs as tools, with a prompt that instructs the agent on your standard quote format. The agent does not just create the document; it also stores the lead in your knowledge base so it remembers the interaction next time.
How Do You Build an Outbound Call Agent That Books Appointments?
To build an outbound calling agent that can call a business, gather information, and book an appointment, follow these steps on Bland AI (the most straightforward platform for this use case):
Step 1: Create a Bland AI Account
Go to app.bland.ai and sign up. You will automatically receive 2 free credits and an inbound phone number (a $15/month value) on the Start plan. No credit card required. (Source: Bland AI pricing, Bland AI starter guide)
Step 2: Define Your Agent's Prompt
Write a prompt that includes the persona, the task, conversation rules, and tools. For an appointment booking agent, the prompt should cover:
- How to introduce itself ("Hi, this is [Agent Name], [Your Name]'s assistant")
- What information to gather (price, time, any constraints)
- When to call back to confirm (after checking with its boss)
- How to confirm the booking when calling back
- Natural conversation handling (adapting to questions, clarifying details)
Step 3: Send the Call
Use Bland's Send Call API to place an outbound call to the target business. You pass the phone number, the agent's prompt, and any variables (like the car make and model for a mechanic inquiry). The API returns a call ID you can use to track the call.
Step 4: Handle the Conversation
The agent makes the call, introduces itself, asks its questions, and records the answers. If it doesn't know something the business asks, it says so naturally ("What do you recommend for that vehicle?") rather than freezing. After the call ends, Bland sends a post-call webhook with the transcript and extracted analysis.
Step 5: Confirmation Call-Back
If the business offered a time, the agent checks with you (via text, Telegram, or a second phone call) to confirm. Once you approve, it calls the business back to finalize the booking. This two-call pattern — gather, check, confirm — prevents the agent from booking appointments you cannot keep.
Step 6: Calendar Update
After the booking is confirmed, the agent calls your calendar tool to create the event with a reminder. This is where the tool integration pays off: the appointment goes straight from a phone call to your calendar without you touching anything.
Which Platform Should You Choose for an AI Phone Assistant?
| Your Situation | Best Platform | Why |
|---|---|---|
| You want the cheapest all-in rate and no platform fee | Bland AI Start | $0.14/min covers everything; no card needed; 2 free credits |
| You want full control over every model layer | Vapi | $0.05/min platform fee; bring your own STT, LLM, TTS |
| You want a visual builder and fast setup | Retell AI | Drag-and-drop builder; ~600ms latency; native HubSpot/Twilio |
| You want a no-code, done-for-you agent | Synthflow or Lindy | Visual deployment; no engineering needed |
| You run a Shopify store and want support calls | Ringly.io | Shopify-native; resolves ~73% of calls without escalation |
The recommendation for builders and small businesses: Start with Bland AI's Start plan if you want the simplest, cheapest path. It bundles the entire AI stack into one rate, gives you a phone number for free, and requires no credit card to begin. You can build a functional outbound calling agent within a day. Move to Vapi or Retell only if you need granular control over the model stack or deeper native integrations.
The recommendation for agencies: If you are building voice agents for clients — a legitimate and growing business model — Vapi's combination of flexibility and per-minute scaling is well-suited, but budget for engineering time. A developer documented spending 100+ hours building a single Vapi agent, and found the AI was only ~20% of the effort; the rest was error handling, CRM integration, and workflow automation. (Source: HappyRobot — Vapi AI Pricing)
What Are the Hidden Costs and Gotchas?
The headline rate is not your real cost. With Vapi and Retell, the advertised per-minute rate covers only the platform layer. You pay separately for STT, LLM, TTS, and telephony. Budget for the all-in number, not the headline. (Source: AI Agent Square — Vapi Pricing 2026)
Engineering time is the real expense. Configuring the prompt, building tool integrations, testing edge cases, and maintaining the agent over time can consume dozens of hours — often far more than the per-minute API cost. A Vapi developer's account of 100+ hours on a single agent is not unusual.
Per-minute costs compound with call length. A 30-second qualify-and-transfer call is cheap. A 5-minute conversation is 10x the cost. Optimize your prompts for efficiency — get in, get the information, get out.
Carrier surcharges on SMS. If your agent sends SMS follow-ups, Twilio charges carrier pass-through fees ($0.003–$0.005/message) on top of the $0.0083/message rate. A2P 10DLC registration is mandatory for legitimate business messaging in the US. (Source: AutomationAtlas — Twilio Pricing 2026)
Concurrency limits. Bland's Start plan allows 10 concurrent calls and 100 calls/day. The Build plan ($299/mo) raises that to 50 concurrent and 2,000 calls/day. If you need high volume, factor the platform fee into your budget. (Source: Bland AI pricing)
Security is not optional. A phone number is public. Anyone who knows it can call. Your agent must gate access, block sensitive data, and log every interaction. Do not skip the security layer to save time.
Pricing changes. Bland AI had a pricing revision in December 2025. Vapi's component costs shift as model providers update their rates. The pricing in this article was verified in July 2026, but you should confirm current rates on the vendor's pricing page before building. (For more on how AI pricing shifts and how to adapt, see our coverage of the GPT-5.6 Luna price cut.)
What This Means for You
If you run a small business, agency, or solo operation, an AI phone assistant is now accessible enough to build in an afternoon — not a quarter. For about $1.15/month (the phone number) plus $0.14/minute on Bland AI's free-to-start plan, you can have an agent that calls businesses to book appointments, answers your calls with a security gate, reads from your knowledge base, and updates your calendar. The technology is real and deployable today.
But the competitive moat is not the minutes or the phone number — those are commodities. The moat is your prompt, your tool integrations, and your security layer. The businesses that get real value from this stack are the ones that invest in a well-structured agent prompt, connect it to the right data sources, and never skip the security gate. The agent that sounds like a professional assistant, knows your business, and protects your data — that is the one worth building.
If you are also running AI models at the edge or on constrained hardware, note that voice agent inference is fundamentally the same problem as any other LLM serving — just with tighter latency requirements. Our guide to running a 29M-parameter LLM on an $8 ESP32-S3 microcontroller covers the same fundamental trade-offs (model size, quantization, throughput) from the embedded side.
FAQ
Q: How much does it cost per month to run an AI phone assistant? A: A minimal deployment on Bland AI's Start plan costs about $1.15/month for a Twilio US local number plus $0.14 per minute of call time, with no platform fee. For 1,200 minutes/month (roughly 30 calls/day), expect to pay around $170–$185/month all-in. Vapi and Retell can cost slightly less or significantly more depending on your model choices, because their headline rate covers only the platform layer — you pay separately for STT, LLM, and TTS.
Q: Do I need to know how to code to build an AI phone agent? A: No — but it helps. Bland AI's Start plan requires no credit card and lets you configure agents through their developer portal with prompts, not code. A basic outbound calling agent can be live within a day. For deeper tool integrations (calendar, CRM, custom APIs), some scripting is needed. Platforms like Synthflow and Retell AI offer visual builders for teams that want to avoid code entirely. (Source: Bland AI starter guide)
Q: What is the difference between Bland AI, Vapi, and Retell AI? A: Bland AI bundles the entire AI stack (LLM, STT, TTS) into a single per-minute rate starting at $0.14/min with no platform fee — simplest and most predictable. Vapi charges $0.05/min for its orchestration layer only, and you pay separately for every other component — maximum flexibility but more complexity. Retell AI sits in between at $0.07/min with a visual builder and native integrations, reported latency of ~600ms, and a G2 rating of 4.8/5.
Q: Is it legal to have an AI make phone calls for me? A: Voice AI platforms require you to comply with calling laws and consent requirements in your jurisdiction. Bland AI's documentation explicitly states that users are "responsible for compliance with calling laws and consent requirements." In the US, this means following TCPA regulations for outbound calls, including obtaining consent for marketing calls and honoring do-not-call lists. For informational calls (like booking your own appointment), the legal risk is lower, but you should still review applicable regulations. (Source: AI Agent Square — Bland AI Review 2026)
Q: Can an AI phone assistant access my calendar and CRM? A: Yes — through tool integrations. On Bland AI, you define Custom Tools (external API endpoints) that the agent can call live during a conversation. On Vapi, you use their tools API. On Retell AI, native HubSpot and Twilio integrations handle common cases, and you can add custom integrations through their API. The agent is configured to know which tools it can call and when, so it can check your calendar for availability and book an appointment during a live call.
Q: How do I secure my AI phone assistant so strangers cannot access my data? A: At minimum, require a numeric access code before the agent shares any information. Follow it with caller ID allowlisting (reject calls from unknown numbers at the telephony layer), tool-level access control (sensitive APIs like banking are never callable through the phone interface), and full audit logging of every call. For healthcare data, use a HIPAA-eligible platform with a signed BAA (Bland AI Enterprise, Vapi Enterprise). Never rely on the prompt alone for security — enforce it at the tool and telephony layers.
Q: Can I use Telegram instead of phone calls to interact with my AI assistant? A: Yes. Telegram's Bot API is free and supports both text and voice notes. You can connect it to your AI agent's LLM and TTS so you can send a voice note and receive a spoken response — useful when you cannot take a phone call. This adds an asynchronous channel alongside your phone number, which is especially useful for sending status updates, quotes, and follow-ups.
Discussion
0 comments