Null StudioNullStudio

Blog · November 13, 2026 · 9 min read

Voice and SMS AI Agents: Running One Conversation Across Calls and Texts

By the Null Studio team

TL;DR: Most businesses buy voice AI and SMS automation as two separate things: a voice agent that answers calls, and a texting tool that sends confirmations or a "sorry we missed you" message. Customers do not think in channels: they call, text back an hour later, then call again, and expect you to remember. A combined voice and SMS AI agent treats calls and texts as one conversation with one memory, one set of rules and one record in your CRM. It texts what is hard to say on the phone (links, addresses, forms, payment requests), calls when a text thread stalls or gets complicated, and hands off to a person with the full history attached. This guide covers when each channel wins, how shared context actually works, the US texting rules that quietly break most setups, and what to ask a vendor before you buy.

Why two channels as one system

The usual setup looks like this. A voice agent answers the phone and books appointments. A separate tool sends confirmation texts, reminders and missed-call text-backs. The two share a phone number at best and nothing else.

That works until a customer crosses channels, which happens constantly:

Each gap costs a booking or some goodwill. Our guide to the missed-call revenue leak covers text-back on its own. This article is about one agent that works both channels with shared context.

When to talk and when to text

Each channel is good at different things. Designing a combined agent starts with deciding which job belongs where.

Voice is better for

Text is better for

The handoffs between them

The value of a combined system is in moving between channels on purpose:

  1. Call to text. During a call the agent says "I'll text you the address and a link to the intake form," and the text arrives before the call ends. The caller does not have to write anything down.
  2. Text to call. A text thread that goes past a few back-and-forth messages, or where the customer writes something like "it's complicated," triggers an offer: "Easier to talk it through? I can call you now." The call opens already knowing what was said in the thread.
  3. Missed call to text to call. A missed call gets an immediate text. If the customer replies with intent, the agent either answers by text or calls back, depending on what they asked.
  4. Either channel to a human. When the agent should not handle something, the person who takes over sees the transcript of every call and the full text thread in one place. How that handoff should work is covered in voice AI human handoff.

What shared context actually means

"Omnichannel" gets promised a lot. In practice it comes down to a few concrete pieces of engineering.

One identity per customer

The phone number is the natural key, but it is not perfect: households share numbers and businesses call from main lines. The system needs a clear matching rule and a CRM lookup so a returning customer is recognized, not treated as a new lead every time.

One conversation record

Every call transcript and text message goes into a single timeline for that customer. When a new call or text arrives, the agent loads a short summary of recent history: what they asked about, what was booked, what is still open. Loading the whole history every time makes responses slower and more expensive. The record should also hold the customer's language, so someone who called in Spanish gets their follow-up text in Spanish (more in multilingual voice AI).

One set of rules

Booking rules, pricing language, escalation triggers and hours should be defined once and used by both channels. Two separately configured bots drift apart, and customers notice when the answer depends on how they asked.

One set of actions

Both channels should call the same tools: check availability, book, reschedule, cancel, create a CRM note, send a payment link. If those actions write to your real systems, the safeguards from our guide to AI agents with write access apply to both channels equally. A text saying "cancel my appointment" should go through the same confirmation step as a spoken request.

Writing for text is not writing for voice

A common shortcut is to reuse the voice agent's prompts for text. It reads badly. Spoken responses are built for the ear: short sentences, confirmations repeated back, no visual structure. Text replies should be shorter still, front-load the answer, and avoid sending three messages where one will do.

A few rules that hold up in practice:

The voice side has its own craft, covered in voice agent conversation design.

The texting rules that break most setups

Texting from your number is a separate project from answering calls on it, and it is where many otherwise good deployments fail quietly.

Registration before sending

In the US, business texting from a regular local number goes through A2P 10DLC registration: the business and each messaging use case are registered with the carriers, usually through your messaging provider. Toll-free numbers have their own verification process. Unregistered or poorly described traffic is often filtered, and the failure mode is silent. Your system logs the message as sent, and the customer never receives it. Registration also takes time to approve, so start it early, not the week before launch. Our AI receptionist phone setup guide covers enabling texting on a number you already own.

Consent and opt-outs

Texting rules depend on what you are sending and where. Transactional messages the customer asked for, like a confirmation of a booking they just made, are treated differently from marketing messages. Reply keywords like STOP must be honored across every system that can text that person, not just the one that received the word. If your SMS agent, your reminder tool and your marketing platform each keep their own opt-out list, someone eventually gets a text after opting out. Keep one list, and check it before every outbound message. Rules vary by country and sometimes by state, so get legal advice for the markets you serve. The voice side of consent is covered in voice AI call recording and consent.

Even registered traffic can be filtered if it looks like spam, such as public link shorteners or large identical bursts. Use your own domain for links, keep content consistent with the registered use case, and monitor delivery receipts, not just send counts.

Measuring a combined system

The useful measures follow the customer, not the channel:

Our guide to AI voice agent monitoring covers the review loop, and AI automation ROI shows how to count the return without double counting a booking that touched both channels.

Where we fit

Null Studio builds voice and SMS agents as one system rather than stitching a voice bot to a texting tool. We built the voice agents, sales pipelines and automations behind CallGuard AI, which answers conversations, books appointments and captures revenue around the clock, and voice agents and automations for CallSetter AI, which answers every call in under 60 seconds, qualifies leads and books appointments. For Fortell we build the AI voice and SMS systems used for intake in more than 100 languages. A combined deployment includes the shared customer record, both channel agents, the booking and CRM actions, 10DLC or toll-free registration, one opt-out list, and monitoring. If you need it built around a workflow no off-the-shelf product covers, that is our custom AI agent development work.

Buyer checklist

Before you sign with anyone, ask for these:

  1. A live call that ends with a text containing the address, time and a working link, arriving before the call ends.
  2. A text thread that turns into a call where the agent already knows what was said by text.
  3. A reschedule by text for an appointment booked by phone, with the change showing in your real calendar.
  4. Proof that STOP on one channel stops all automated texts from every system you use.
  5. A clear answer on who owns the 10DLC or toll-free registration and what use cases are registered.
  6. Delivery receipts, not just send logs, in the reporting.
  7. Proof that the same rules apply on both channels: ask the same pricing question by call and by text and compare the answers.
  8. A clear answer to who reviews transcripts and threads every week and fixes what they find.

Customers already move between calling and texting without thinking about it. A combined agent simply stops your business from forgetting what they said when they switch.


Null Studio designs, builds and runs voice and SMS AI agents as one system: shared customer history, booking and CRM actions, carrier registration, opt-out handling and monitoring. Book a demo and we'll show a call and a text thread working off the same conversation.

FAQ

What is a combined voice and SMS AI agent?

An AI system that answers and makes calls and sends and replies to texts as one conversation, with one memory per customer, one set of business rules and one record in your CRM, so a customer who calls and then texts never has to repeat themselves.

When should an AI agent text instead of call?

Text anything the customer needs to keep or tap: addresses, times, confirmations, intake forms, booking and payment links, and reminders. Use a call for back-and-forth questions, urgent issues and upset customers, and offer a call when a text thread stalls or gets complicated.

Do I need A2P 10DLC registration for an AI agent that texts?

In the US, business texting from a regular local number generally requires A2P 10DLC registration through your messaging provider, and toll-free numbers need their own verification. Unregistered traffic is often filtered silently, so start registration early and monitor delivery receipts.

How should opt-outs work when several tools send texts?

Keep one opt-out list shared by every system that can text a customer, and check it before every outbound message. A STOP reply on one tool must stop automated texts from all of them. Rules vary by country and state, so get legal advice for your markets.

Want it built, not just explained?

We design, build and run these systems end-to-end — shipped in days, not months.

Book a demo →

Keep reading