Twilio does not sell a product called "Twilio AI Agent." It sells the conversation layer that other people's agents run on: voice transport, orchestration across channels, memory, and the intelligence that reads what was said. If you are searching for a Twilio AI agent, you are almost certainly looking for Conversation Relay, and the useful question is what it does and does not cover.
The confusion is understandable. Twilio is the default answer for programmable voice and messaging, so "Twilio AI agent" reads like a product name. It is closer to a category. Below is what Twilio actually named, what each piece costs, and where the line sits between the conversation and the work.
What Twilio actually ships under "AI agent"
Twilio groups its agent-adjacent products under conversational AI, and each is separately named and separately metered. Nothing here is an agent you switch on. Each is a component you assemble.
- Conversation Relay: real-time voice AI. It converts speech to text, passes it to your large language model, converts the response back to speech, and manages turn-taking over a live call.
- Conversation Orchestrator: keeps one conversation coherent as it moves across channels, so a thread that starts in SMS and continues on voice is still the same thread.
- Conversation Memory: builds an identity-resolved profile by joining customer data with conversation history, so an interaction starts with context rather than a cold prompt.
- Conversational Intelligence: extracts structure from what was said, using Twilio-authored Operators or Custom Operators you define. Twilio's documentation and the Conversation Relay product page use "Conversational Intelligence"; the pricing page lists the same product as "Conversation Intelligence".
- Enterprise Knowledge: grounds replies in your own FAQs, policies, and product documentation. The Enterprise Knowledge API reached general availability at SIGNAL 2026.
Twilio also refers to Agent Connect, which lets AI agents take part in conversations alongside human agents. The framing across all of it is consistent: Twilio owns the pipe and the context, you own the reasoning.
What each piece costs
Twilio publishes per-unit prices for each product, so a cost model can be built from the pricing page rather than from a sales call. The trade is that a single conversation can touch four meters at once, so the per-minute headline is not the whole bill.
| Product | What it does | Published starting price |
|---|---|---|
| Conversation Relay | Real-time voice AI over a live call | From $0.07 per minute |
| Conversation Orchestrator | Keeps threads connected across channels | From $0.0002 per 1,000 characters ingested |
| Conversation Memory | Persistent, identity-resolved customer profile | From $0.0028 per 1,000 characters generated, plus $0.007 per memory recall |
| Conversational Intelligence | Extracts structure and context from conversations | From $0.005 per 1,000 characters for Twilio-authored Operators; $0.002 per 1,000 input and $0.018 per 1,000 output characters for Custom Operators |
| Enterprise Knowledge | Grounds answers in your own documents | $0.018 per GB-hour storage, $0.005 per retrieval |
Two costs sit outside that table, and both follow from how Twilio packages this. The first is telephony itself: the call still carries standard per-minute voice charges. The second is the model. Twilio orchestrates your LLM rather than bundling one, so tokens are billed by whoever provides the model. A realistic voice-agent estimate is Relay, plus telephony, plus tokens, plus whichever of Conversation Memory or Conversational Intelligence you enable.
What we found researching this post
We tried to research this the way a buyer would, and the first thing we hit was that the buyer's path is broken. Twilio's own product URL for AI assistants, twilio.com/en-us/products/ai-assistants, returned a 404 when we probed it on 7 September 2026, and the press release index at twilio.com/en-us/press/releases renders through JavaScript, so it came back empty to anything that is not a browser. The durable, citable sources turned out to be the pricing page and the SIGNAL 2026 launch post, not the product pages.
That matters beyond our inconvenience. It means a lot of what is written about "Twilio AI agents" is sourced from aggregators rather than from Twilio, and aggregators in this category invent product names with real confidence. Every name in this post is one Twilio uses itself. Where we could not confirm something from Twilio's own material, we left it out rather than filling the gap.
The second finding was in the pricing structure. Conversational Intelligence charges Custom Operators at $0.018 per 1,000 output characters against $0.002 per 1,000 input characters, a nine-to-one ratio. If you plan to write your own Operators, the design that keeps output terse is the design that keeps the bill sane. That is not obvious from the headline price, and it is the kind of detail that only shows up once you read the meters individually.
What changed at SIGNAL 2026
Twilio's May 2026 SIGNAL announcements were mostly about making voice AI survivable in regulated industries and buildable by non-specialists. The changes worth knowing:
- PCI-compliant workflows and HIPAA-eligible architectures, which is what unblocks payments, healthcare, and insurance use cases that previously could not put an LLM on the call.
- Conversation Relay Studio widget, a drag-and-drop path into Relay rather than pure code. It lowers the floor; it does not remove the need to design the conversation.
- Deepgram Flux support for smarter turn-taking, aimed at the interruption and talk-over problems that make voice agents feel wrong.
- Enterprise Knowledge API general availability, so grounding in your own policies is a supported path rather than a custom build.
Read together, these are maturity moves: compliance coverage, a lower-code entry path, and better turn-taking all target operating a voice agent rather than prototyping one. For a regulated buyer, the PCI and HIPAA items are the ones that change whether the project is possible at all.
The gap: work Twilio's conversation layer cannot see
Twilio's boundary is clean and worth stating plainly: if the work does not happen inside a customer conversation, Twilio is not the tool. Its products act at the moment a customer is on the line, and do nothing the rest of the week.
The work that falls outside includes:
- Recurring internal jobs on a schedule, like a Monday pipeline summary or a nightly data check, where no customer is present.
- Cross-tool work that spans your CRM, billing, and project tracker, where the useful output is a reconciliation rather than a reply.
- Back-office follow-through after the call ends: updating records, chasing the internal owner, checking that the promised thing actually happened.
- Anything where the trigger is time or a data change rather than an inbound conversation.
This is a real division of labor rather than a criticism. A voice agent that answers well and then drops the follow-through leaves the work half done, and that gap is usually closed by pairing the conversation layer with something that runs the recurring work behind it.
How to decide, by situation
- You need a phone line answered by AI. Conversation Relay is a strong fit, and because the per-minute price is published you can model the cost before committing. Budget for Relay plus telephony plus tokens.
- You are in a regulated industry. The 2026 PCI and HIPAA work is the reason to look again if you ruled Twilio out in 2025.
- You already run Twilio for messaging. Orchestrator and Memory are incremental adds on infrastructure you already operate, which is a much shorter path than adopting a new vendor.
- You have no developer. The Studio widget helps, but Twilio still assumes someone owns the model choice and the logic. A packaged agent platform will get you further faster.
- Your work is internal and recurring. Twilio is the wrong shape. Look at scheduled agents that read your tools directly, and see how agents differ from chatbots and assistants if the distinction is still fuzzy.
Where Gravity fits next to Twilio
These solve different halves of the same problem, and they compose more often than they compete. Twilio carries the customer conversation. Gravity runs the recurring work that the conversation creates, and the work that never involves a customer at all.
A support team using both would let Conversation Relay answer the call and take the request, then have a Gravity agent do the follow-through on a schedule: check the order system, update the ticket, flag anything still open after 48 hours, and post a summary to the team channel on Monday. Neither tool does the other's job well.
On cost, the models differ in kind. Twilio meters per minute and per character across several products at once. Gravity is a subscription: a free tier covering one agent at $0 per month, and paid plans from $20 per month that include $20 of usage, with the option to buy extra usage beyond your plan as you scale. If you are comparing total spend across categories, our guide to the cheapest AI agent platforms and the breakdown in AI agent pricing explained cover how per-unit and subscription models diverge at volume.
If the support desk is the real problem you are solving, AI customer service agents covers that use case end to end, and teams handling the follow-through usually pair it with Zendesk ticket triage. For the neighboring vendor explainers in this series, see Workday AI agents and Ramp's finance agents.
Frequently asked questions
Does Twilio have an AI agent?
Not as a single packaged product. Twilio sells the conversation layer that agents run on: Conversation Relay for voice, plus Orchestrator, Memory, Intelligence, and Enterprise Knowledge. You bring the model and the logic.
How much does Twilio Conversation Relay cost?
From $0.07 per minute, per Twilio's published pricing. That excludes standard telephony charges and your model provider's token costs, both of which apply on top.
What is the difference between Conversation Relay and Conversation Orchestrator?
Relay handles real-time voice on a single call. Orchestrator keeps a conversation coherent as it moves between channels, so an SMS thread that becomes a phone call is still one conversation.
Can Twilio agents do work outside a phone call?
Not on their own. Twilio's products act inside customer conversations. Recurring internal jobs, cross-tool reconciliation, and scheduled work need a separate agent platform.
Do I need a developer to build a Twilio AI agent?
Less than before. The Conversation Relay Studio widget added a drag-and-drop path in 2026, but you still choose the model, design the conversation, and own the business logic.
Is Twilio Conversation Relay HIPAA compliant?
Twilio announced HIPAA-eligible architectures and PCI-compliant workflows for Conversation Relay at SIGNAL 2026. Eligibility depends on your configuration and agreement with Twilio, so confirm scope with them for your build.
Sources
- Twilio pricing, per-unit prices for Conversation Relay, Orchestrator, Memory, Intelligence, and Enterprise Knowledge, checked 7 September 2026.
- Infrastructure for the agentic era: everything we launched at SIGNAL 2026, Twilio, 6 May 2026.
- Conversation Relay product page, Twilio, checked 7 September 2026.