Blog
Guides, tutorials and updates from the Dial team.
Featured
Most AI agents live entirely in text. But the real world runs on phone calls, SMS verification codes, and WhatsApp messages. Here's why giving your agent a real number changes everything.
Jensen Huang, Sam Altman, and Marc Andreessen all landed on the same number independently: one human, one hundred AI agents. The math is not a forecast. It's already happening in early enterprise deployments — and our communication infrastructure is completely unprepared.
All posts
Operators running automated WhatsApp numbers keep describing the same thing: not a steady trickle of bans, but waves that take out many accounts within days. Here is what the pattern looks like, what seems to trigger it, and why the one approach that avoids it entirely is the official one.
Five bold, clearly labeled predictions about how agents will communicate by 2030, with the reasoning behind each. No fake data, just arguments you can disagree with.
Ten crisp principles for builders putting agents in front of real people, each with a short reason why. Drawn from the needs agents actually have and the failures production teams keep hitting.
A pipe bursts at 3 AM and the plumber's phone goes to voicemail. This is a design walkthrough for an agent receptionist: what it says, what it captures, when it wakes a human, and what the owner reads over coffee.
Fax machines did not lose to a better fax machine. They lost to something with different assumptions. History keeps repeating this pattern, and agent communication is standing in the same spot.
A voice agent that passes your unit tests can still hang up on a customer mid-sentence. Here is a practical testing playbook: scripted personas, adversarial callers, transcript replay, latency budgets, sandboxes, and canaries.
Which channel should your agent use to reach a customer? The sticker price per message is the least interesting part of the answer. Here is a framework built on attention, failure cost, retries, and compliance overhead.
For a century, a familiar voice on the line was proof enough. Synthetic voices ended that, and caller ID never really worked as proof anyway. Here is what has to carry trust instead, and what you can do about it today.
Letting an agent send mail from a shared mailbox feels like a shortcut. It is a trap with at least four ways to go wrong: threading chaos, identity confusion, reputation spillover, and two agents auto-replying to each other until someone gets a very large bill.
One-on-one chat is easy: someone talks, you answer. Group chats ask an agent to read the room, know when to stay quiet, handle side conversations, and protect people who never asked to be part of its context. That is a lot harder than it looks.
Agents have no concept of 3 AM unless you give them one. Quiet hours belong in infrastructure policy, not in a prompt, and getting a recipient's local time right is harder than it looks. Here is how to build politeness in.
Your agent can reply in 400 milliseconds with four perfect paragraphs. Should it? An opinionated guide to message length, response delay, bursts, and typing indicators, and the cases where machine speed is the whole point.
Every number you retire does not disappear. It gets recycled to a stranger, who inherits its history, its spam score, and its wrong-number texts. A look at the afterlife of agent phone numbers, and why the number-per-agent model fills the graveyard so fast.
For years, giving an agent a phone meant weeks of SDK glue code. With the Model Context Protocol, the agent can discover communication as a set of tools and just use it. Here is what that shift looks like in practice.
Caller ID can tell you which carrier a call came from. It cannot tell you whether the caller is a person or a piece of software, or what that software is allowed to do. An attestation token is the missing piece, and here is what it should carry.
A customer replies STOP to a text. Three days later, the same company's email agent writes to them again. Treating consent as a protocol primitive, checked at dispatch and shared across channels, closes the gap. A builder's view, not legal advice.
Two versions of the same customer story. In one, an agent remembers Tuesday's call when the customer texts on Saturday. In the other, every channel is a stranger. A walkthrough of what session continuity changes.
Every voice agent hits a moment it should not handle alone. What happens next, a smooth warm transfer or a caller forced to repeat their whole story, is the part customers actually remember.
The phone rings, a warm voice says hello, and ten seconds later you realize you are talking to software. How that moment feels, and why the best time to disclose is before the person has to wonder.
Picture a restaurant booking agent calling a restaurant that answers with its own AI. Two polite machines, one phone line, and nobody willing to hang up first. Why agent-to-agent traffic needs a machine channel and not a spoken one.
The moment your agent can read an SMS, an email, or a WhatsApp message, it can be told what to do by a stranger. Every inbound channel is untrusted input, and most agent builders have not treated it that way yet.
Imagine an agent named Juno, a brand new number, and a to-do list of 400 appointment reminders. By hour three, things get interesting. A hypothetical day in the life of a number nobody warned about the real world.
Calling a voice API takes an afternoon. Everything underneath it - number provisioning, carrier registration, reputation monitoring, per-country compliance, rotation, retries - is the other 95% of the work, and it never stops being work.
The voice model is nearly indistinguishable from a human. The call still feels like talking to a machine. The reason is almost never the voice itself - it's the 700 milliseconds of dead air after every sentence, and what happens when the caller tries to interrupt.
No-shows cost service businesses real money, and the fix is a well-timed reminder with an easy way to confirm or reschedule. Here is a working reminder flow in under 40 lines - schedule, send, and handle the reply.
Every phone number type trades off differently between trust, throughput, setup time, and cost. Picking the wrong one is a common reason a new calling or texting workflow underperforms before a single word of the script matters.
WhatsApp does not work like SMS. There is a 24-hour window that determines whether you can send freely or need a pre-approved template, a quality rating that can get your number restricted, and opt-in rules that are stricter than almost any other channel.
For about a week, a fraud ring tried to route free Dial accounts into a rotating set of international premium-rate numbers to collect revenue share at our expense. Here is what International Revenue Share Fraud actually is, how we caught it, and what we changed.
Your agent just navigated a signup form, filled in every field correctly, and clicked submit without a single mistake. Then it hit a text box asking for a six-digit code, and the whole task ended right there. This is the most common failure mode nobody talks about.
The verification code was sent two minutes ago. The user is still staring at four empty boxes. It is not one bug - it is five different failure points stacked on top of each other, and most verification flows are designed around none of them.
Your AI agent's calls are reaching people with a "Scam Likely" label - or not reaching them at all. Here is what causes it, what Dial handles on your behalf, and the one thing only you can control: how your agent actually calls.
TCPA applies to calls to US numbers - not calls from US numbers to international recipients. The rules for AI voice agents are more complex than SMS: live vs. automated distinctions, AI disclosure mandates, recording consent laws, and DNC registries in 20+ countries. Every major market mapped.
10DLC is a US domestic regulation. It applies when your US long code sends to a US number - not when you send to Israel, India, or anywhere outside the US. But the destination country has its own rules. Here is every major market mapped, with the exact regulation name and enforcement type.
The global phone numbering system was designed in 1984 for humans calling other humans. AI agents are consuming numbers at a rate the infrastructure was never built for - and the bottleneck is not the number pool itself. It is the routing databases, registration queues, and reputation systems underneath it.
Everyone is benchmarking AI calling against human agents. The actual competition is the voicemail box. 34% of outbound AI calls go straight to it, and all your conversation design doesn't help a single one of them.
Carrier analytics systems score every number before it rings. AI agents trip every signal: volume spikes, low answer rates, no inbound history, identical audio. Most developers never find out — they just see no answer in the logs.
A 2025 SurveyMonkey study found 79% of Americans prefer human customer service over AI. Most developers read that as a ceiling. It isn't. People don't hate AI calls — they hate bad AI calls. Here's the difference, and exactly what to build differently.
Most developers building AI voice agents don't know which protocol stack they're on — or what that choice costs them in latency, reliability, and reach. Here's a plain-language breakdown of the three layers, how they interact, and what the right architecture looks like in 2026.
For 15 years, RCS sat mostly dormant — a richer SMS standard that carriers kept implementing unevenly and Apple refused to support. In September 2024, iOS 18 shipped with RCS. Global traffic jumped fivefold. Here's what actually changed, why it matters for AI agents, and what the infrastructure gap still is.
The internet needed TCP/IP. The web needed HTTP. The cloud needed virtualization. Every technology wave has required a new infrastructure layer to be built before the wave could break at scale. The agentic era is no different — and the infrastructure layer it needs doesn't exist yet.
Article 52 of the EU AI Act creates mandatory disclosure requirements for AI systems interacting with humans in real-time conversational contexts. If your agent makes calls or sends messages to EU residents, you have a compliance obligation that most engineering teams haven't addressed.
The demo works. The prototype impresses. Then you try to ship it, and you discover that between your working agent and a production deployment that can actually communicate, there's 6–10 weeks of carrier registration, domain warm-up, and compliance setup. Most teams don't see it coming.
A phone number is identity rooted in physical possession — you have a SIM card, so you have a number. AI agents don't have SIM cards. They need identity rooted in cryptographic proof — persistent across deployments, portable across infrastructure, verifiable by anyone.
The CPaaS market was valued at $18 billion in 2024 — built on human communication at human scale. If the 1:100 agent ratio materializes, agent communication will dwarf that by an order of magnitude. The agent communication market doesn't have a name yet. But the math is not hard.
If your agent only communicates via SMS and voice, it's invisible to most of the world. WhatsApp has 2 billion active users in 180+ countries. Outside North America, it's the default communication channel — not SMS. Building for agents without WhatsApp is building for half the planet.
When AI agents coordinate with each other, they do it through API calls and shared state stores. Unverified. Unauditable. Invisible to the humans overseeing the system. ACES defines an authenticated, auditable, session-persistent primitive for agent-to-agent interaction — and it may be the most important channel nobody is talking about.
For a decade, voice calls felt like a legacy channel. Then LLMs got fast enough for real-time conversation, and suddenly calling is the most natural interface again. There's a reason ACES starts with phone — and it's not nostalgia.
A phone call, an SMS, an email, and a WhatsApp message are not just different channels. They're different protocols, different encoding, different compliance requirements, different trust frameworks. The Channel Abstraction Layer normalizes all of them into a single interface so your agent never has to care.
TCPA, CAN-SPAM, GDPR, and the EU AI Act apply simultaneously to AI agents communicating across channels. None of these frameworks were written with agents in mind. The right answer isn't a bigger legal team — it's compliance baked into the infrastructure layer.
Twilio is excellent infrastructure. It's also human infrastructure — every rate limit, reputation model, and identity primitive it exposes was designed for a world with humans on at least one end of every communication. For agents, that's the wrong starting point.
Every team building AI communication agents runs into the same five failure modes, in roughly the same order. Reputation decay. Volume throttle. Identity ambiguity. Cross-channel fragmentation. Compliance exposure. None of them are bugs. All of them are infrastructure.
All human communication on Earth — every SMS, email, WhatsApp message, and phone call — totals roughly 437 billion events per day. AI agents, at the 1:100 ratio, will generate 6.5 trillion. That's not a scalability challenge. It's a category error.
The phone number was the identity anchor of the telephone era. The email address was the identity anchor of the internet era. The Agent Communication Number (ACN) is the identity anchor for the agentic era — persistent, cryptographically rooted, and designed to work across every channel simultaneously.
STIR/SHAKEN is a real improvement for phone call authentication. But it only answers one question: did this call originate from a legitimate carrier customer? It says nothing about whether the caller is a human or an agent, what the agent is authorized to do, or whether the communication is legitimate.
Your agent called a lead last Tuesday about a contract. The lead texts back on Saturday with a question. Your agent has no idea who this is or what they discussed. Every channel switch is a cold start. This is the defining failure mode of agent communication today — and it's entirely infrastructure.
A phone number used by an AI agent for outbound calls accumulates spam complaints at 10–100× the rate of a human-operated number. Carriers update their databases within hours. A clean number on Monday is flagged by Wednesday. Most teams never see it coming until calls stop connecting.
Teams build email agents, those agents send at agent scale, and one day the CEO notices that the company's email isn't being delivered. It takes weeks to recover from domain blacklisting. Most teams never see it coming.
AI agents go from zero to deployed in the time it takes to run a pip install. Getting permission to use a phone number for that agent takes 6–10 weeks. This is not a minor inconvenience. It's a categorical incompatibility with how agents work.
Phone numbers assume SIM card possession. Email assumes human typing patterns. SMS rate limits assume human-speed outreach. Every layer of communication infrastructure was built with one unexamined assumption: the entity communicating is a person.
A step-by-step tutorial on automating SMS verification workflows. We'll sign up for a real service, intercept the OTP, and complete verification — all from a Claude Code session.