.svg)
Braze Conversational Agents let customers complete tasks by chatting with an AI agent on WhatsApp, SMS, RCS or web chat. The feature is in beta. Design each conversational workflow like a journey, with one purpose, a defined audience and steps that end in a clear action. Test the logging outside preview, review conversation history, and take legal advice before production, as Braze advises.
Key Takeaways
- Conversational Agents let customers complete tasks by chatting with an AI agent on SMS, RCS, WhatsApp or web chat, and the feature is in beta.
- Each workflow has a purpose, channel settings, a target audience and natural-language steps that can use tools.
- For SMS and RCS, an existing keyword trigger takes priority, so the agent isn't invoked for matching replies.
- Custom attributes are set asynchronously, so an attribute read straight after it's set may be stale.
- Preview mode doesn't log custom attributes or events, so test logging separately.
- Braze advises consulting your legal team before deploying conversational AI to production.
What are Braze Conversational Agents?
Conversational Agents is a channel that lets users talk with an AI agent to complete tasks you define. A workflow might help a customer choose and buy items based on their preferences.
Braze labels it a beta and asks interested customers to contact their account manager. It also advises consulting your legal team about using conversational AI before you deploy to production.
Supported channels are SMS, RCS, WhatsApp and web. For SMS, RCS and WhatsApp you choose the subscription groups a workflow runs in, and web chat runs through a widget in the Braze Web SDK.
How is a conversational workflow built?
Each workflow has a purpose, channel settings, a target audience and natural-language steps. The purpose tells the agent when to use the workflow, and the audience is the set of segments a user must belong to.
Steps can include tools, which is where the agent acts on your data. The table lists them with the design question each one raises.
| Tool | What it does | Question to settle first |
|---|---|---|
| Search knowledge sources | Queries a catalog through a knowledge source | Which catalog fields may the agent see? |
| Set and get workflow attribute | Stores a value for later steps in the same workflow | Should the agent choose freely, or from allowed values? |
| Log custom event | Logs a Braze custom event, which can send a user into a Canvas | Which journey should this conversation trigger? |
| Set and get custom attribute | Writes or reads an attribute on the user profile | Is the agent allowed to change this attribute? |
| Call webhook | Sends a request to an external endpoint | What happens if the endpoint fails? |
Braze advises ending each step with an explicit action, so the agent knows how to phrase its next reply. Treat that as the rule for every step you write.
Which settings change how the agent behaves?
Brand guidelines and channel settings live on the Conversational Agents settings page. Web chat supports an opening message, and for the messaging channels you should review the keywords set for each subscription group.
Those keywords matter more than they look. For SMS and RCS, the agent isn't invoked if a reply matches an existing keyword trigger, so your existing keyword handling takes priority over the agent.
Check every keyword before launch. A keyword set long ago for a promotion can intercept replies that the agent was meant to handle.
Which behaviours should you design around?
Two documented behaviours catch teams out. Custom attributes are set asynchronously, so if the agent reads an attribute straight after setting it, the value may still be stale.
Design around that by using workflow attributes for values the conversation needs later. They're scoped to the workflow and avoid a round trip through the profile.
Webhook credentials need care too. Braze advises creating Connected Content credentials in the dashboard and referencing them in the webhook tool, rather than typing secrets into a step.
For web chat, the setup enables SDK authentication and turns on user-supplied JavaScript in the Web SDK initialisation. Ask your engineering and security teams to review both settings, since they change how the site and the SDK trust each other.
How should you test a workflow?
Preview lets you test one workflow, or the full experience for a web app or subscription group, as a specific user. But preview mode doesn't log custom attributes or custom events.
That means preview proves the conversation, not the data. Test the logging separately, with a test user in a non-production setup, before you rely on any Canvas the agent triggers.
Conversation history shows conversations with real users and from preview, including the tools the agent called. Review it daily in the first weeks after launch, because tool calls show exactly what the agent did on a customer's behalf.
In what order should you launch Conversational Agents?
Launch in this order, so each risk is handled before customers reach the agent.
- Ask your Braze account manager about joining the beta.
- Take legal advice on using conversational AI, as Braze recommends.
- Write one workflow with a single purpose and a narrow target audience.
- Decide which attributes, events and webhooks the agent may use, and nothing more.
- Review keyword triggers for each subscription group the workflow uses.
- Preview the conversation, then test logging outside preview.
- Launch to a small segment and review conversation history daily.
- Widen the audience only after the history shows the agent behaving as expected.
What should your security and legal teams review?
Conversations carry customer messages, and the agent can read and write profile data. Ask your security team to review what conversation data is stored, and who can see conversation history.
For Agent Console, Braze's FAQ states that when the Braze Auto model is used, Google acts as a Braze sub-processor under the data processing terms between you and Braze. Confirm with your account team how that applies to Conversational Agents before rollout.
Consent belongs in the same review. Subscription groups decide where the agent can reach people, so check that every group a workflow uses matches the consent you hold.
Which mistakes should you avoid?
Name the mistakes so they're easy to catch. A catch-all workflow tries to handle every request, so the agent can't tell when it applies.
A keyword collision lets an old keyword trigger intercept replies meant for the agent. A preview-only test confirms the conversation but never checks the events and attributes it should log.
A stale read uses a custom attribute straight after the agent set it. An open-ended tool lets the agent change attributes or call webhooks nobody decided it should touch.
Where does this connect to the rest of the programme?
Agents log events that can start journeys, so event design in tracking user events in Braze applies directly. Conversations sit inside the wider design covered in multichannel engagement campaigns.
Catalog-backed answers depend on the content patterns in seven ways to deliver dynamic content and the personalisation in hyper-personalisation in Braze. Profile data the agent reads often comes from integrating CRM data into Braze.
As a Braze implementation partner, CustomerIK recommends starting with one narrow workflow and a small audience. Widen it only when the conversation history shows the agent doing exactly what you designed.
A bounded conversation to test before rollout
For a fictional retailer, give the agent one job: help a customer locate approved order-tracking information. Define the allowed data source, what identity verification happens outside the model, and when the conversation hands off to a person. Do not let a successful-sounding reply substitute for a completed task.
Test with fictional records in the supported environment and inspect both conversation history and downstream event logging. A preview passing is not proof that production logging works. Agree a failure owner and measure resolved tasks, incorrect answers and escalations alongside response time.
| Test case | Expected result |
|---|---|
| Order cannot be found | Explain the limitation and use the approved handoff |
| Customer asks for an unrelated account change | Stay within the workflow's scope |
| Data source is unavailable | Do not invent an order status; use the defined fallback |
The worked example and test cases are illustrative editorial guidance, not measured client outcomes or a claim that a live integration has been tested.
Frequently Asked Questions
1. Are Braze Conversational Agents generally available?
They're in beta. Contact your Braze account manager to ask about participating.
2. Which channels do Conversational Agents support?
SMS, RCS, WhatsApp and web. Web chat runs through a widget in the Braze Web SDK.
3. Why isn't the agent replying to some SMS messages?
For SMS and RCS, the agent isn't invoked if a reply matches an existing keyword trigger. Review the keywords for each subscription group.
4. Does preview test the events an agent logs?
It doesn't. Preview mode doesn't log custom attributes or custom events, so test logging separately.
5. What should we do before going to production?
Braze advises consulting your legal team about the implications of conversational AI before you deploy to production.
Sources
- Braze, Conversational Agents. Retrieved 18 Oct 2026.
- Braze, Agents FAQ. Retrieved 18 Oct 2026.
CustomerIK is a Braze implementation partner working across onboarding, technical integration, marketing operations and customer data management.
If you're planning a first conversational workflow and want the design and safeguards set before launch, let's talk.
.svg)
