ChangelogBook a demoSign up

Agents

Audience: Business users (marketers, lifecycle managers, retention strategists)
Prerequisites:

  1. Configuration →
  2. Prepare data for AID →
  3. Configure a destination for AID (SFMC, Braze, Iterable)

By the end of this article, you’ll be able to:

  • Understand what an agent is and how it powers a campaign
  • Navigate the Agents interface in Hightouch
  • Create an agent and configure its core building blocks: audience, actions, outcomes, and guardrails
  • Manage and monitor an agent over time

What is an agent?

An agent is the container for your campaign strategy in AI Decisioning. It defines:

  • Audience – Who to target
  • Actions – What messages to send
  • Goals/Outcomes – What success looks like
  • Guardrails - When and how often messages can be sent, and which users are eligible

Once an agent is live, AID uses reinforcement learning to optimize message selection and delivery timing—personalizing the experience for each user and improving continuously based on performance data.

Understand the building blocks

Audience: Who should this campaign reach?

Each agent starts with an audience—defined using your existing models in Customer Studio.

This is usually a broad segment of users with a shared lifecycle or intent, like:

  • “New users who signed up in the last 30 days”
  • “High-LTV customers who haven’t purchased in 60+ days”
  • “App users who viewed a product but didn’t buy”

You can optionally add a holdout group (e.g., 20%) to measure how much lift AID is driving compared to no intervention.

Customer Studio Audiences.

Actions: What messages can AID send?

An action is a specific message (email, push, SMS, etc.) that AID can send to eligible users.

Each action includes:

  • A template from your engagement platform (Braze, Iterable, SFMC, etc.)
  • Channel and content type
  • Variables to personalize content (e.g., subject lines, images, offers)
  • Optional: tags, collections, and eligibility filters

Example

In a win-back campaign, you might configure:

  • An email offering a 10% discount
  • An SMS with loyalty points
  • A push notification with a new product alert

AID will learn which message works best for each user.

Actions Page.

Outcomes: What defines success?

Outcomes are the conversion events AID uses to measure performance and learn.

Each agent requires at least one outcome, such as:

  • Purchase
  • App install
  • Referral completed
  • Email click

You can define multiple outcomes and assign them different levels of importance:

  • Best – the primary goal (e.g., purchase)
  • Good / Very Good – secondary goals (e.g., clicks or add-to-cart)
  • Bad / Worst – negative outcomes (e.g., unsubscribes)

You can also weight outcomes (e.g., $100 purchase > $10 purchase) and map campaign attribution fields to tie results to specific messages.

AID Outcomes

Agents page

You’ll find the Agents section in the left sidebar under AI Decisioning > Agents.

From here, you can:

  • View all existing agents and their status (Running, Paused, Pending)
  • See target audience breakdowns (treatment vs. holdout)
  • Click into any agent to:
    • Initialize agent
    • Edit the target audience
    • Add or update actions
    • Add or update outcomes
    • Specify scheduling
  • Click Add agent to create a new motion

Step 1: Create a new agent

Add agent

Click Add agent and configure the following:

Agent name

Use a goal-based name, e.g.:

  • “First purchase nudges”
  • “Win-back for lapsed users”
  • “Upsell: second-to-third purchase”

Target audience

Choose from audiences created in Customer Studio. These are typically lifecycle-based segments (e.g., new sign-ups, high-LTV customers, churn risks).

If you do not have any audiences yet, create an audience first.

Holdout group (optional)

Apply a percentage of users (e.g., 20%) who won’t receive any messages. This helps you measure the true incremental lift of AID.

Click Add agent to finalize creation.

Agent window

Step 2: Configure your agent

After creating your agent, open it and go to the Configuration tab—this is where the core campaign logic is defined.

Agent configuration

1. Add actions (the messages AID can choose from)

Actions are messages (email, SMS, push) that AID can send to users. These are created using your connected ESP, like Iterable, Braze, or SFMC.

To set up an action:

  1. Click Add actions
    Start from your agent’s Configuration tab.
  2. Select a channel
    Choose the type of message (e.g. email, SMS, push) and the engagement platform you’re using (e.g. Braze, SFMC, Iterable).
  3. Choose a template
    Pick one or more message templates from your engagement platform. These serve as the creative foundation for the action.
  4. Name your action
    Give it a descriptive name that clearly identifies its purpose (e.g., “Reactivation offer – SMS” or “First Purchase Email A”).

Action configuration

2. Add outcomes (what success looks like)

Outcomes define what the agent should optimize for.

  1. Choose an event model (e.g., purchase, add_to_cart, click)

  2. Classify it:

    • Best → primary goal (e.g. purchase)

    • Good → helpful signals (e.g. clicks)

    • Worst → negative outcomes (e.g. unsubscribe)

  3. Apply weights (optional) to factor value (e.g. $100 > $10 purchase)

  4. Map attribution using campaign IDs

If you’re optimizing for totally different business goals, use separate agents.

Define outcomes

3. Define guardrails for your actions

Click into any action and navigate to the Guardrails tab. Here, you can set controls to keep delivery safe and relevant:

  • Send limits – e.g., max 3x/week per user
  • Timing windows – e.g., send only M–F between 9am–6pm
  • Eligibility filters – e.g., only send if user hasn’t purchased in 14 days

Guardrails

4. Preview your content

Click into any action and navigate to the Content tab. Here, you can see a live preview of your message with sample user data and variable combinations.

Preview content

5. Optional: Advanced configuration and Cortex

The Advanced configuration tab gives you tools to improve campaign performance, reporting, and personalization over time. These settings aren’t required to launch—but using them unlocks richer insights and helps AID learn faster.

You can:

  • Add tags to describe content themes, tone, or product focus

  • Attach collections to recommend personalized products or content

  • Add campaign IDs for tracking in your downstream ESP

  • Use Cortex to analyze message tone, complexity, or structure

Advanced configuration

Cortex

Tags: Help AID learn what your content is about

Tags are labels that describe your content’s tone, focus, format, or goal. They help AID group similar content and understand which kinds of messages work best for different people.

Why use tags?

  • See clear comparisons in Insights (e.g., "Which tone performs best for Gen Z?")

  • Share learnings between different actions or agents

  • Help AID personalize faster and smarter with historical context

What should you tag?

Ask yourself:

  • What is this content trying to do?goal = reactivation, goal = first_purchase

  • What product, feature, or value prop does it highlight?product = shirts, product = pants

  • What’s the tone?tone = serious, tone = playful

  • What’s the visual style?image = people, image = scenery

Tag each variant/template individually—not just the Action—to enable the most accurate learning and insights.

Example

If you tag three email templates with different tones (tone = serious, tone = playful, tone = educational), AID can learn which type of content is preferred by user—and prioritize similar content next time when you add new iterations into the system.

Collections: Dynamically personalize what each user sees

Collections let you pull in personalized content—like product grids, articles, or promotions—based on user behavior and preferences.

Think of it as letting AID choose which item to show, not just how to show it.

Why use collections?

  • Tailor content to each user in real time

  • Test different product sets or themes without hardcoding

  • Make content modular and scalable across use cases

What to include in a collection

Collections are built from catalogs you’ve synced into Hightouch. For example:

  • Products (e.g., “Spring Accessories over $20”)

  • Content (e.g., “Health articles tagged as beginner”)

  • Promotions (e.g., “Active discounts for loyalty members”)

You can filter collections by fields like category, season, stock status, or price. You can also add eligibility filters so only certain users receive certain items.

Use collections when you want AID to recommend instead of hardcoding—especially in upsell, cross-sell, win-back, and content engagement campaigns.

Tips

  • Use 3–10 items per collection to give AID enough choice

  • Add filters like inventory_status = In Stock to avoid showing unavailable items

  • Preview your collection in the Content tab to see what users will get

Campaign IDs: Connect AID to your downstream analytics

Campaign IDs let you tie each action back to your ESP (like SFMC, Braze, or Iterable) for detailed reporting and attribution.

Why use campaign IDs?

  • Connect AID actions to ESP analytics dashboards

  • Attribute performance (opens, clicks, conversions) by variant

  • Ensure consistency across systems for testing or compliance

When to use

If your ESP uses campaign identifiers (like Journey IDs or Triggered Send names), add them here to:

  • Auto-tag performance in Braze or Iterable

  • Track SFMC sends in Journey Builder or Triggered Send logs

Cortex (optional): Analyze content tone and complexity

Cortex is an AI-powered tool that analyzes the tone, style, and structure of your content. It’s especially useful for teams managing large libraries of variants.

Why use Cortex?

  • Detect tone mismatches early (e.g., “this version is too formal for Gen Z”)

  • Compare content complexity across templates

  • Get suggestions for variation (e.g., split a long-form email into short-form variants)

Even if you don't act on Cortex data right away, it gives your team a head start in QA, creative testing, and scaling messaging libraries.

What Cortex can analyze

  • Readability and linguistic complexity

  • Sentiment and tone categories

  • Word count, structure, and formatting patterns

Step 3: Manage and maintain your agent

Once your agent is running, use the Overview tab to monitor performance:

  • Sends and action distribution

  • Outcome conversion rate

  • Treatment vs. holdout lift and statistical significance

  • Variant-level engagement metrics

You can:

  • Pause or resume the agent

  • Add or update actions

  • Adjust guardrails as your strategy evolves

Overview

What’s next

→ Inspector: View individual user actions and QA delivery

Ready to get started?

Jump right in or a book a demo. Your first destination is always free.

Book a demoSign upBook a demo

Need help?

Our team is relentlessly focused on your success. Don't hesitate to reach out!

Feature requests?

We'd love to hear your suggestions for integrations and other features.

Privacy PolicyTerms of Service

Last updated: Jun 9, 2025

On this page

Was this page helpful?