Can AI Assistants Make Phone Calls? Give Yours a Number
Can AI assistants make phone calls, yes, but ops leaders need the right infrastructure beneath the model to avoid costly build failures.
AI can make phone calls. But the model is the easy part. The real challenge is the five-layer infrastructure beneath it, and most teams only discover what they're missing after their first production failure.
Most teams assume the hard part is finding an AI smart enough to hold a conversation. The harder problem is building the infrastructure beneath that AI so it can actually pick up a phone, stay on the line, and complete a task without you hovering over it. That gap, between conversational intelligence and genuine telephony capability, is where most AI calling projects quietly fail without a human in the loop. The tool, the infrastructure beneath it, and the task complexity all determine which version of "can make calls" you're actually getting.
Google Duplex, announced in May 2018, was the clearest early proof that AI could conduct a real, autonomous phone conversation. But Duplex wasn't a general-purpose AI that someone simply "connected" to a phone number. It was a purpose-built system assembled from automatic speech recognition, text-to-speech, dialogue management, and carrier connectivity, designed for constrained, specific tasks. That distinction matters enormously for anyone evaluating AI calling capability today. Five components have to exist beneath the AI model before it can autonomously conduct a phone conversation:

- Carrier connectivity
- A real phone number
- A streaming speech-to-text engine
- A text-to-speech layer
- The LLM itself
Remove any one of these, and the call cannot happen.
None of these components come bundled with a large language model. ChatGPT, Claude, and similar models are text-in, text-out systems by default. Most teams discover this too late.
They license a capable model, assume the phone infrastructure is a commodity detail, and then spend weeks stitching together third-party providers for STT, TTS, and carrier access, introducing latency, data residency questions, and reliability risk at every seam. Platforms like Bland AI address this by owning the full stack, STT, LLM inference, TTS, and carrier access on self-hosted hardware, removing the latency seams and third-party audit exposure that come from renting each layer separately.
The harder question is what has to be true beneath the model for that call to actually complete, stay compliant, and scale past your first hundred concurrent lines.
Key takeaways#
- Conversational fluency in a chat window does not transfer to phone calls, the bottleneck is infrastructure latency, not model intelligence, and no prompt engineering fixes that.
- Most AI calling projects fail not because the AI can't talk, but because the stack beneath it can't handle real telephony: SIP trunking, real-time audio streaming, turn-taking, and sub-200ms response loops all have to work together.
- Checking appointment slots and qualifying a 12-minute Medicare lead are both 'phone calls', they require completely different infrastructure, and most AI assistants were only built for the first one.
- The FCC's February 2024 ruling makes AI-generated voices subject to TCPA prior written consent requirements for marketing calls, that is not a gray area, and state attorneys general can layer additional enforcement on top.
- Compliance exposure and infrastructure fragility kill AI calling projects before they reach production far more often than model capability does.
- Provisioning a dedicated phone number for an AI assistant is not a developer sprint, it takes roughly 30 seconds, and the real cost driver is the infrastructure layer underneath, not the rate card.
- Bland closes the gap by providing fully self-hosted AI phone calling infrastructure, one platform handling telephony, voice agents, and compliance architecture together, without routing your data through unvetted third-party providers.
How AI Phone Calling Actually Works Under the Hood#
That assumption is where production failures are born, because the bottleneck in AI phone calls is the compounding latency created by stacking five discrete architectural layers between a user's instruction and a live caller's ear.

How AI Phone Calls Actually Work - The Five-Layer Architecture Explained#
Understanding that pipeline matters before you choose a vendor, because no amount of model intelligence or prompt engineering can fix dead-air silences caused by stacking speech-to-text, LLM inference, and text-to-speech delays on top of one another.
Most buyers discover this only after their first production failure, when a vendor outage or a latency spike turns a live customer conversation into dead air. Teams that have built toward 24/7 coverage without adding headcount, or those trying to reduce call center costs by 50% or more, cannot absorb that kind of failure in production. Understanding the architecture before you buy is how you avoid that 2 a.m. surprise.
The Five-Layer Pipeline From User Instruction to Live Phone Conversation#
- Telephony layer establishes the carrier connection and routes the audio stream. Twilio's programmable voice infrastructure is one example of a telephony layer that teams commonly connect at this stage.
- Speech-to-text (STT) layer transcribes the caller's spoken words into text in real time, with no hidden per-transcript surcharge stacked on top of talk time.
- LLM reasoning engine reads that text, determines intent, and generates a reply; token charges are included in the per-minute rate on every Bland plan rather than billed separately.
- Text-to-speech (TTS) engine converts the reply back into audio, using premium voices or clones (also included in the per-minute rate, with up to 15 voice clones on Scale and 1 on Start).
- Providers such as ElevenLabs supply high-fidelity voice synthesis at this layer.
- Orchestration layer manages the state of the conversation across turns, routes tool calls such as calendar updates or CRM writes, and decides when to transfer or end the call.
Each of these layers must stay connected to external systems without introducing additional delay, which is why Bland's integrations platform and conversational pathways are available on every paid plan to keep the orchestration layer connected to external systems, including Amazon Connect, for teams that want to add AI voice inside existing call flows without migrating to a new platform entirely.
Why Latency Is the Silent Killer Across STT, LLM and TTS#
The delay problem begins the moment a caller finishes speaking. According to the Soniox Speech-to-Text Benchmarks (Pipecat STT Benchmark), even a best-in-class STT model carries a median latency of 260ms, with P95 performance at 305ms. The LLM then needs time to generate its first token, and the TTS layer adds its own rendering delay before a single word reaches the caller's ear.
260ms
Median latency for best-in-class speech-to-text
Accuracy compounds the problem alongside speed. Soniox records a 1.34% semantic Word Error Rate on 1,000 real-world samples for its leading model, which sounds small until a misheard medication name or policy number triggers the wrong action (Soniox Speech-to-Text Benchmarks (Pipecat STT Benchmark)). As industry research notes, each step in the pipeline compounds delay, and cumulative lag determines whether a conversation sounds natural or broken.
1.34%
Semantic word error rate on real-world samples
At scale, those margins become critical. A team handling thousands of outbound sales calls, follow-up reminders, or inbound support interactions per day, Bland's Scale plan supports up to 1,000 concurrent calls per hour and 5,000 per day, cannot tolerate a slow transcription layer or unpredictable LLM latency at scale.
AI Phone Agents vs Traditional IVR in Multi-Turn Conversation#
AI phone agents and traditional IVR systems are architecturally incompatible. IVR routes callers through a fixed decision tree and cannot handle a caller who answers "sometime next week, maybe Thursday if you have mornings free" when the script expects "press 1 for Monday, press 2 for Tuesday." AI phone agents handle that kind of open-ended, multi-turn exchange natively.
Bland's conversational pathways are available on every plan and are designed to manage exactly this kind of multi-turn, branching logic, including knowledge bases (10 on Start, up to 100 on Scale, and unlimited on Enterprise) that let the agent draw on accurate, scoped information rather than hallucinating answers under latency pressure.
For teams with stricter compliance or infrastructure requirements, Bland's Enterprise tier adds dedicated infrastructure, on-prem and VPC deployment options, data residency controls, SSO, BAA availability, and compliance documentation available under NDA, with a forward-deployed engineering team that scopes, builds, and goes live with the first agent through a 28-day deployment framework.
For prospects or customers who are easier to reach via text, or where multi-channel outreach improves response rates, Bland SMS extends the same automation logic beyond voice into a complementary channel.
What AI Assistants Can Do on Phone Calls Depends on Which Type You Have#
Calling a salon to check appointment slots and qualifying a Medicare lead over a 12-minute conversation are both "phone calls," yet the tools capable of handling each differ entirely.

Built-In Device Assistants Offer One Tap and Zero Autonomy#
Built-in device assistants, Siri and Google Assistant, can initiate a call with a voice command, but their involvement ends the moment the other party picks up.
They cannot listen, respond, handle a hold queue, or complete any task on the other end. A user who asks Siri to "call the clinic and check if they have a Friday opening" will hear the phone ring and then be on their own, because initiating a call and conducting one are not the same thing.
Those are structurally different functions, and no amount of model improvement changes the fact that device assistants were designed for device control, not live telephony.
Consumer Auto-Dialers Are Experimental and Geography-Locked#
The closest consumer tools come to autonomous calling is Google's agentic calling, which lets Google place calls to local businesses on a user's behalf to gather information like pricing, stock availability, and opening hours, then delivers a summary by text or email; the user is never on the line. In practice, it launched as a limited beta restricted to select U.S. markets, not a general release available to most users.
A deeper problem surfaces even when these tools work as advertised: they struggle to understand nuanced or specific customer inquiries, making them poorly suited to anything beyond simple, predictable exchanges.
Enterprise Voice Agents Where Multi-Turn Calling Actually Lives#
Enterprise AI voice agents handle tasks that consumer tools cannot approach: multi-turn qualification conversations, appointment reminders with calendar writes, inbound routing with CRM logging, and outbound campaigns at volume. This is where the types of AI phone call tools diverge most sharply: consumer tools report back a summary; enterprise voice agents write data into systems of record and hand off to humans mid-call when escalation logic fires.
Bland is purpose-built for this tier, most beneficial when a business handles high call volumes or needs 24/7 phone coverage without scaling headcount. Its Conversational Pathways architecture is the specific capability that closes the gap consumer tools leave open: it is most valuable when call scripts have multiple conditional branches or require dynamic routing based on caller responses, enabling agents to adapt in real time rather than stall on an unexpected answer.
Rather than a linear script that breaks on the first unexpected answer, Bland agents navigate the full decision tree, automate the highest-volume and highest-stakes calls end to end, and integrate with existing CRM, CCaaS, and telephony systems so data lands where it belongs without manual intervention, eliminating the reconciliation work that typically follows high-volume call campaigns.
For teams already running on Amazon Connect, Bland's Amazon Connect Integration lets you add AI voice to existing inbound and outbound call flows without migrating to a new platform, preserving the infrastructure already in place.
Why Two AI Assistants Making Phone Calls Can Produce Completely Different Results#
The answer lies in infrastructure: stitched-together third-party components introduce latency, failure points, and compliance gaps that make enterprise-grade calling unreliable. Bland's unified calling stack, STT, LLM, TTS, and telephony on a single controlled infrastructure, closes that gap. On Enterprise plans, data residency controls and compliance documentation are available under NDA, making legal sign-off tractable rather than a six-month negotiation; a forward-deployed engineering team scopes, builds, and tests the first agent within a defined 28-day deployment framework before go-live.
Device Assistant (Siri/Google)#
Can initiate a call autonomously: ✅
Stays on the line and converses: ❌
Handles multi-turn dialogue: ❌
Handles conditional branching (e.g. Conversational Pathways): ❌
Writes data to CRM/calendar: ❌
Available outside select U.S. markets: ✅
Complies with TCPA/FTC disclosure requirements: N/A
Supports inbound + outbound on same number: ❌
Integrates with Amazon Connect / CCaaS: ❌
Requires developer integration: ❌
Consumer Auto-Dialer#
Can initiate a call autonomously: ✅
Stays on the line and converses: ❌
Handles multi-turn dialogue: ❌
Handles conditional branching (e.g. Conversational Pathways): ❌
Writes data to CRM/calendar: ❌
Available outside select U.S. markets: ❌
Complies with TCPA/FTC disclosure requirements: Varies
Supports inbound + outbound on same number: ❌
Integrates with Amazon Connect / CCaaS: ❌
Requires developer integration: ❌
Enterprise Voice Agent#
Can initiate a call autonomously: ✅
Stays on the line and converses: ✅
Handles multi-turn dialogue: ✅
Handles conditional branching (e.g. Conversational Pathways): ✅
Writes data to CRM/calendar: ✅
Available outside select U.S. markets: ✅
Complies with TCPA/FTC disclosure requirements: ✅ (platform-dependent)
Supports inbound + outbound on same number: ✅
Integrates with Amazon Connect / CCaaS: ✅
Requires developer integration: Varies
Current Limits of AI Phone Calling Every User Should Know Before Dialing#
Compliance exposure is the limitation that ends AI calling projects before they ever reach production.

Disclosure Laws and Regional Locks That Ground AI Callers Before They Dial#
AI phone tools are legally required to identify themselves as AI at the start of a call. This is not a best practice. Under Section 5 of the FTC Act, failing to disclose a caller's non-human nature constitutes an unfair or deceptive practice, and the FTC's guidance for AI chatbots and voice agents explicitly lists deceiving users into thinking they are speaking with a human among its prohibited behaviors.
California's SB 1001 prohibits any bot from denying it is a bot when sincerely asked, with civil penalties attached to violations. Other states are moving in the same direction at different speeds, creating a patchwork that an AI calling platform must handle jurisdiction by jurisdiction.
Where Conversational Reliability Actually Breaks Down - IVR Trees, Abrupt Hang-Ups, and Loop Failures#
They surface in the plumbing around it. Interactive voice response systems rely on DTMF tone detection. It just stops progressing.
They are the normal failure modes of any system tested in controlled conditions but not stress-tested against real caller behavior.
The Hidden Infrastructure Ceiling - Third-Party Audio Routing, Data Residency Risk, and Integration Gaps#
A healthcare intake agent that routes raw call audio through a third-party LLM API creates a HIPAA audit exposure the IT team discovers during a compliance review, not during the demo.
The integration layer adds its own failure mode. Direct connections to Salesforce, ServiceNow, or Zendesk allow an AI phone agent to retrieve and update records mid-call, but when those connections break or were never built correctly, the result is a manual handoff that costs operations teams hours per week in reconciliation work.
A smarter LLM cannot fix a disclosure architecture that was never built, a DTMF handler that misfires, or an audio routing path that empties into a provider you cannot name in your compliance documentation.
How to Give Your AI Assistant Its Own Phone Number Right Now#
Three minutes.
That is roughly how long it takes to provision a dedicated US phone number through AI phone number and attach it to the AI assistant already open in your browser tab. The reality is that a dedicated number with inbound and outbound calling plus SMS can be live inside a single working session, for $9.99 the first month, then $19.99 a month, with no telephony engineering required, though texting volume is subject to fair-use limits and call volume is capped, making this the right fit for personal delegation and small-team workflows rather than high-volume operations.
"AI systems (e.g. Gemini) can hallucinate and output real private phone numbers as placeholder or mock contact data, causing unsolicited calls and messages to innocent individuals."
— what we hear from AI safety advocates

Bland AI's noise cancellation can be configured at three levels: the agent (Persona) level, per individual call, or per inbound phone number.
The Agent Phone Plan and What You Actually Get for $9.99, then $19.99 a month#
The plan provisions one dedicated US number with inbound and outbound calling plus SMS. Fair-use limits apply to texting volume, so bulk SMS campaigns are outside its scope, and call volume is similarly capped to suit personal and small-team use rather than high-frequency operations.
Noise cancellation filters background noise upstream of the transcriber, achieving a 16% lower word error rate across live calls, as shown in Bland's noise cancellation documentation, and can be tuned at the agent, per-call, or per-number level to match your environment.
Our research found that after the free tier, Bland Speech pricing per character is $0.015 per 1,000 characters, with that same rate applying both in the studio and through the API.
Which AI Apps Connect and How Each One Plugs In#
Claude Desktop connects via Model Context Protocol (MCP), which has become the dominant standard for giving AI assistants access to external tools. You install the Bland MCP server, point Claude Desktop at it, and the assistant can initiate outbound calls and send texts as part of any workflow you describe in natural language. The practical result: you can tell Claude to call a contractor, confirm a quote, and text you the summary, without touching your phone.
Pay-Per-Call MCP Credits as a No-Number Alternative#
If you only need outbound calls occasionally and have no use for inbound or texting, MCP credits let you pay per call without committing to a monthly number. There is no persistent caller ID, so the number presented to recipients may vary between calls. For one-off tasks like checking business hours or confirming a delivery window, that is fine.
Five Real Phone Calls Your AI Assistant Can Make Right Now#
According to Nextiva's 2024 research, 60% of customers prefer a callback over waiting on hold, which means an AI assistant placing that call on your behalf is often a welcome interaction rather than an intrusive one. None of these require a complex script, just a plain-language instruction to your assistant.
What AI Phone Calling Costs and Why the Infrastructure Layer Determines Real Value#
Buying an AI phone calling platform on rate-card alone is roughly as useful as hiring a call center based on agent wages while ignoring QA, telecom, rework, and training costs. Those hidden layers matter just as much in AI calling, which is why understanding what sits above the per-minute rate floor is essential before any meaningful comparison.
Bland Speech v3 was trained on over 100 million real human conversations, teaching the model conversational speech patterns rather than polished studio delivery.

The Market Rate Map From Consumer Plans to Enterprise Infrastructure#
The pricing tiers span a wide range: $99 per month for a dedicated number with bundled calling; $0.14 per minute with no platform fee and a 100-call daily cap; $0.12 per minute at a $299 monthly platform fee; and $0.11 per minute plus a $499 monthly platform fee, supporting up to 5,000 calls per day, 1,000 calls per hour, and 100 concurrent calls. Each tier reflects a different assumption about volume, infrastructure ownership, and what the headline rate actually includes.
Why the Per-Minute Rate Is a Misleading Comparison Metric for High-Volume Operations#
The gap between the advertised rate and the true cost of a call becomes most visible at scale. A team picks a nominally cheaper platform, launches at scale, and then discovers that every call routes LLM inference through a frontier provider billed separately in tokens. Teams running outbound sales campaigns or 24/7 inbound support at volume have encountered this pattern acutely: the platform rate looks contained, then a separate model-provider invoice surfaces that was never part of the budget model.
Latency compounds the problem. Slower response times extend call duration, directly inflating per-minute costs even before any secondary invoices are factored in.
Stack Ownership as a Pricing Signal for Enterprise Buyers#
The most reliable signal that a platform's per-minute rate is actually all-in is whether the vendor owns every layer of the stack, from speech recognition and LLM inference through to voice synthesis and carrier access.
Bland's fully self-hosted infrastructure is built on this principle. STT, LLM inference, TTS, and carrier access are all operated on its own hardware, so the $0.11 per minute on the Scale plan reflects the complete cost of a call rather than a floor that a separate API invoice quietly raises each month. The voice layer is built on Bland Speech v3, which was trained on over 100 million real human conversations, so premium voices and clones are included in the per-minute rate across every paid plan rather than billed as an add-on through a third-party TTS provider.
That stack ownership also matters for teams layering AI calling on top of existing contact center infrastructure. Bland's Integrations Platform is designed for this scenario, most beneficial when a business already has a contact center platform or CRM and wants to add AI voice without migrating to a new system. The integration sits alongside existing tooling rather than displacing it. In that architecture, the per-minute cost is additive to existing infrastructure spend, which makes the all-in nature of Bland's rate directly relevant to the total cost model.
For lower-volume use cases the distinction may be manageable, but at 5,000 calls per day the math inverts quickly, and the compliance exposure from unvetted third-party data routing becomes a separate, non-trivial cost category entirely, one that broader contact center economics consistently identify as a systemic underestimate in total cost of ownership.
Next steps#
If your AI assistant can hold a conversation but falls silent the moment it needs a real phone line, the path forward starts with owning the infrastructure beneath the model, not upgrading the model itself.
The compounding latency problem across speech-to-text, LLM inference, and text-to-speech means dead air is an infrastructure failure, not a prompt engineering problem. And as the compliance picture makes clear, legal exposure from TCPA consent requirements and FTC disclosure mandates is baked into the dialing architecture before a single call goes out, not determined by what the AI says during it. Together, these realities point to one action: get your AI a dedicated number on a stack that controls every layer, so you stop renting someone else's ceiling.
Start with an AI phone number from Bland. Real-time transcription, voice, and LLM inference are included in a single per-minute rate, inbound and outbound run on the same line, and the disclosure and consent architecture is built in from day one.
Frequently Asked Questions#
What's the real difference between an AI phone agent and a traditional IVR system?#
AI phone agents and traditional IVR systems are architecturally incompatible. IVR routes callers through a fixed decision tree and breaks the moment a caller gives a natural, unpredictable answer, like "sometime next week, maybe Thursday if you have mornings free", whereas an AI phone agent understands varied phrasing across multiple turns without losing context, updating its understanding with each new piece of information rather than resetting to a menu prompt.
Why does an AI call sometimes have awkward pauses or dead air?#
Dead air is a latency problem, not a model intelligence problem. Even a best-in-class speech-to-text engine carries a median latency of 260ms, and that's before LLM inference and text-to-speech rendering add their own delays on top, stack all three and you routinely exceed the threshold where a human perceives an unnatural pause and either repeats themselves or hangs up. No amount of prompt engineering fixes a silence caused by stacking those infrastructure delays.
Can I just connect ChatGPT or Claude to a phone number to make AI calls?#
No, ChatGPT, Claude, and similar models are text-in, text-out systems by default and come with none of the required telephony infrastructure. Five components must exist beneath the AI model before it can autonomously conduct a phone conversation: carrier connectivity, a real phone number, a streaming speech-to-text engine, a text-to-speech layer, and the LLM itself, and none of them are bundled with a large language model.
Do AI callers have to disclose that they're not human?#
Google Duplex triggered disclosure requirements in several states, establishing a legal precedent that any AI caller must identify itself, and that compliance layer has slowed consumer AI calling rollouts considerably. The post notes this as an established legal baseline, though the specific rules vary by jurisdiction.
Can an AI phone agent handle a caller who interrupts or talks over it mid-sentence?#
The post identifies handling silence, hold music, and mid-sentence interruptions without breaking as a key differentiator between enterprise voice agents and simpler tools, it's part of what sits in the orchestration layer beneath the model. Two AI assistants running similar language models can produce radically different results on a live call because of exactly these infrastructure-level capabilities, not the model itself.