Turn customer context into personalized messaging with AI

When customer messaging works, it feels timely, relevant, and specific to what someone is doing in your product. A new user gets the right onboarding nudge. A high-intent account gets routed toward the right follow-up. A disengaged user gets a different path before they disappear.

The hard part has always been making that happen at scale. You can write more branches, more template variants, and more routing rules, but that complexity piles up quickly. The more personalized the experience gets, the harder it becomes to maintain.

The AI node in Journeys gives you a better way to build adaptive messaging. It can classify users, score intent, generate notification copy, enrich profiles, and route each journey based on the context you already have. Instead of mapping every possible path by hand, you define what you want the model to return and use that structured output in the rest of the workflow.

The AI node rolled out to enterprise customers when we launched Journeys in March. Today, it's generally available to every Courier customer.

Make every journey respond to customer context

AI is useful in user messaging when it can do more than generate text. In Journeys, the AI node sits inside the workflow, so its output can decide what happens next: which branch a user takes, which notification they receive, what copy appears in the template, or what gets saved back to their profile.

Here's what that looks like in practice:

• Personalized notification copy at scale. Feed signup data, feature usage, and plan tier into the AI node. Get back a tailored subject line, body copy, and recommended next action for each recipient. One journey can replace dozens of template variants.
• Behavioral scoring and intent routing. Pass product signals (pages visited, features activated, support tickets filed) to the model. Get back a structured score and classification. Branch the journey so high-intent users and disengaged users land in different flows.
• Profile enrichment without a data pipeline. Classify users into personas or lifecycle stages based on activity. Persist the result to their profile so every future journey starts with richer context.
• Real-time information pulled mid-journey. Toggle web search for Anthropic models so the model can look up current data before responding. Generate notifications that reference recent company news, public pricing pages, or anything else outside your journey context.
• Adaptive tone and channel selection. Let the model decide how to deliver, not only what to say. Return a tone or channel_preference field and branch into different send paths based on the model's output.

Each of these would normally live in a separate service or require you to maintain conditional logic across dozens of templates. With the AI node, they're one step on the canvas.

All of these patterns use the same core loop: prompt, schema, structured output, downstream action.

How AI becomes part of the journey

You place an AI node anywhere on the journey canvas. When a run reaches it:

1. Courier assembles the prompt you've written, interpolating journey context (trigger fields, profile data, upstream fetch responses) into {{variable}} placeholders.
2. Courier sends the prompt and your output schema to the model you selected.
3. The model responds with structured JSON matching your schema.
4. Courier parses that response and merges the fields into the journey context.

From that point, every downstream node can reference the AI output the same way it references any other data in the journey. Use it in branch conditions, send templates, fetch URLs, or pass it into another AI node.

Frontier models, ready to call from any node

You pick the model per node, so different steps in the same journey can use different models. Use a smaller model for predictable structured tasks. Use a larger model when the prompt needs more context, more reasoning, or higher-quality generated copy.

OpenAI

Anthropic

Anthropic models also support web search, letting the model query the internet for real-time information before it responds.

The practical starting point: use Haiku or GPT-5.4 Mini for classification, extraction, and scoring. Move up to larger models when the output needs more reasoning or when copy quality matters more than throughput.

Structured answers you can trust and act on

Define exactly what the model returns

You define the shape of the model's response upfront. Add fields with names and types (string, number, boolean) in form mode, or paste a JSON Schema directly. The model returns structured JSON that your branch conditions, send templates, and downstream nodes use directly.

No free-form text to parse. You define the contract, the model fills it.

Prompt with full journey context

Your prompt has access to everything available in the journey at that point: trigger schema fields, user profile data, and responses from upstream fetch nodes. Type {{ in the prompt editor to autocomplete available fields. A token counter shows prompt size in real time.

Web search

Web search is available for Anthropic models. When toggled on, the model can query the internet before responding. Use it when your prompt needs current information that isn't in the journey context, like a company's latest funding round or pricing from a public pricing page.

Conditional execution

AI nodes support conditions. If the conditions aren't met when the run reaches the node, it's skipped and no credits are consumed. This lets you run AI selectively (only for certain segments, only on certain paths) and keep costs predictable.

See what AI will decide before users do

The configuration drawer has a Test button that runs the prompt against the selected model with sample variable values. You see the parsed JSON output immediately, so you can validate structure, check quality, and handle edge cases before the journey goes live.

Iterate on prompts and schemas without triggering real runs. When the output looks right, publish.

Know exactly what AI decided and why

Every AI node execution is inspectable in Run Inspection. Click the AI node step in any run to see:

• Which model ran and whether web search was enabled
• The fully resolved prompt (all variables substituted)
• The output schema that was sent
• The model's structured JSON response
• Input and output token counts

If the node fails (model error, timeout, malformed response), the step surfaces the error. The journey continues, but no AI output merges into context. You can add a branch condition downstream to handle the missing fields gracefully.

Start building

The AI node is available to all Courier customers today. If you've been using Journeys, you already have access.

1. Open any journey in the Journeys editor
2. Add an AI node to your canvas
3. Select a model, write your prompt, define your output schema, and test
4. Publish and go live

Or start with the AI node documentation for full configuration details and examples. For pricing, see AI node billing.

New to Journeys? The Journeys overview covers the visual workflow builder, triggers, branching, and all the other node types. Add AI nodes wherever you need intelligence in your flows.