In this article
You know the win-back campaign you want to run. The audience is clear in your head: customers who finished onboarding, went quiet 45 days ago, and haven't heard from you in a month. Describing that campaign takes a minute, and building it takes most of an afternoon, because someone has to create the segment, draft two emails, apply the template, set the timing, and check the Liquid.
Your AI Agent in Customer.io builds from that description. Lifecycle and growth marketers hand it the campaigns, segments, and activation experiments they already have mapped out. Demand gen marketers use it to find warm contacts that haven't had a sales follow-up. Product marketers use it to build feature adoption sequences, and content marketers use it for subject line variants, segment-specific rewrites, and copy edits. In each case the Agent works inside Customer.io with your actual segments, attributes, events, and campaign history, so you direct the work and the Agent assembles it.
Marketers are already working this way. 73% report meaningful AI impact on their messaging strategy, 61% use AI to write and draft message copy, 45% use it for campaign optimization, and 37% use it for performance analysis, according to Customer.io's 2026 customer messaging guide. The question for most teams has moved from whether to use AI to how to scale it reliably.
Reliable results come from three things you control: the data the Agent can reason about, the permissions your admins configure, and the prompts your team writes. The examples below come from our own marketing team using the Agent in real workflows.
TLDR:
- The Customer.io AI Agent works inside your workspace and can create and edit campaigns, build email content in Design Studio, construct segments from plain language, analyze user profiles, and debug Liquid.
- 73% of marketers report meaningful AI impact on their messaging strategy, and Customer.io's own team uses the Agent for copy, email design updates, Liquid fixes, and user diagnostics.
- Strong prompts have four elements: context, a specific task, constraints, and a defined output format.
- Clean event names, complete attributes, and documented brand guidelines produce more precise Agent output, and a review step before anything goes live keeps first drafts from becoming mistakes.
What is the Customer.io AI Agent?
The Customer.io AI Agent is a built-in assistant that works inside your workspace and can do nearly everything you can do in Customer.io. You ask in plain language, and the Agent acts on your workspace data: segments, attributes, events, campaign history, and customer profiles. It also searches Customer.io documentation to help troubleshoot issues in your workspace. To use it, your workspace needs Customer.io AI enabled.
The Agent runs on the same customer engagement platform that 9,000+ customers use to send messages across email, push, SMS, in-app, and more, including 100B+ messages in 2025. Because the Agent sits inside that platform, it works with the data your journeys and campaigns already use, and your team doesn't need to export anything or move between tools.
What can the Customer.io AI Agent do?
The Agent can build and edit campaigns, create content in Design Studio, construct segments, analyze users and performance, and debug Liquid. In practice, that covers these tasks:
- Campaign building: Create multi-step sequences such as onboarding, win-back, and feature adoption campaigns from a plain-language brief, using your existing templates.
- Email content: Draft emails, write subject line variants, copy-edit drafts, and rewrite a message for a different audience segment.
- Segmentation: Build segments from a description, such as users who signed up in the last seven days and haven't completed a core action.
- User and campaign analysis: Pull a user's attributes, event history, and current campaigns into one overview, or walk through step-by-step performance for a campaign.
- Liquid: Write, debug, and check Liquid logic before it goes into a live message.
- Bulk operations: Apply the same change across multiple draft emails or campaigns in one request.
- Custom skills: Encode a task your team runs regularly so the Agent repeats it the same way each time, using custom skills.
How the Customer.io team uses the Agent to build campaigns
The Customer.io team uses the Agent for campaign copy, Liquid logic, email design updates, performance diagnostics, and bulk edits that used to require a platform specialist. These are real examples from our demand gen, partner marketing, engineering, and support teams.
Campaign content and Liquid
Our demand gen team needed a button that routed each customer to their own account manager's Calendly link. An apostrophe in one account manager's name broke the link. The Agent traced the problem and rewrote the condition with a contains function so special characters no longer throw an error. The team member estimated the manual route at 20 minutes of parsing, with a real chance of starting over with a different workaround.
Someone on our engineering team used the Agent to copy-edit a newsletter for spelling, grammar, and word choice. Their note to the team: "I know this is like the simplest use-case of the agent, but it did a great job and was super quick. 10/10 will use again."
Email design and template updates
Someone on our partner marketing team needed to update the header, logo, and brand styles across nine draft emails in Design Studio for a partner contact. They asked the Agent to apply the design changes in bulk, referencing an existing campaign as the template, and the Agent worked through the emails without manual edits to each one.
Diagnostics and user analysis
Our demand gen team also asked the Agent why a journey attribute showed as populated in a preview when it hadn't been set for any users. The Agent identified the preview panel's default behavior as the cause and explained it clearly enough to skip a debugging rabbit hole.
Our support teams use the Agent to audit campaigns for customers: finding out why a segment condition isn't behaving as expected, tracing which campaign a user is in, or pulling a full view of a user's profile attributes and event history. One member of our support team summarized the pattern: "Claude can act like your analyst sidekick. It pulls together a holistic view of an individual user so you can see their journey in context." Asking the Agent directly replaces the work of navigating several views to reconstruct what happened to one user.
What do you need in place before using the Agent?
The Agent needs clean data, documented brand guidelines, and a deliberate permissions setup before your team relies on it for real work. The Agent reasons about what exists in your workspace, so a shaky data foundation shows up in its output.
Event tracking coverage: The Agent can build segments, identify patterns, and trigger campaigns based on events that are firing. If key product actions are missing or inconsistently tracked, the Agent can't reason about them reliably.
Attribute completeness: Personalization and precise segments depend on consistent, complete customer attributes. Null values, inconsistent formatting, and unclear naming (what does status_v2 mean?) leave the Agent without enough to act on confidently.
Brand guidelines in an accessible place: The Agent writes in your brand voice when you've documented it. A short description of your tone, a "never say" list, and a few of your best-performing emails go a long way. Include them directly in prompts or encode them in a custom skill your team reuses.
Segment definitions: "Our at-risk users" means something different to each person on a team. Documenting your key segment definitions gives every team member, and the Agent, consistent instructions.
How do you control what the Agent can change?
Admins control the Agent's access to live resources with the Allow agent to edit live data setting under Settings > AI. By default, the Agent can't edit live automations, one-time sends, or segments in use, and it can't delete profiles or trigger broadcasts, newsletters, and other messages.
If an account admin turns on live editing, a selector lets them choose how changes get applied. Switching to Auto means the Agent can edit live data without asking first, and Customer.io asks you to confirm that choice. The Customer.io AI toggle in the same settings page turns off every AI feature for everyone in the account.
Keep live editing off while your team builds prompting habits on drafts. When a specific use case calls for live edits, turn the setting on for that purpose and choose the approval option over Auto.
The best prompts for marketing AI agents
A strong Agent prompt has four elements: context, a task, constraints, and an output format. Prompt quality drives output quality more than any other variable.
- Context: Who is the audience, what product are they using, what lifecycle stage are they in, and what's the goal of this message? The Agent uses everything you give it.
- Task: What specifically should the Agent do? "Create a three-part email sequence for users who signed up in the last 14 days and haven't completed profile setup, using the standard onboarding template," gives the Agent a task it can act on.
- Constraints: Which campaigns are off-limits, which brand standards apply, and what's the character limit on subject lines?
- Output format: Do you want a draft email ready to review in Design Studio, a list of segment criteria, or a Liquid snippet? Naming the format means the Agent doesn't have to guess.
Compare two prompts for the same goal.
Weak prompt: "Can you help me build a win-back campaign?"
Strong prompt: "Create a two-email win-back sequence for users who haven't logged in for 45 days but previously completed onboarding. Email one should be friendly and low pressure, highlight recent product improvements, and include a single CTA to log back in. Email two, sent seven days later, should be more direct about their inactivity and include an easy unsubscribe option. Use the 'retention-text' template. Subject line character limit: 50."
The strong prompt takes about 60 more seconds to write and saves multiple rounds of revision.
Prompts for building campaigns
Onboarding sequence
Create a three-part onboarding email sequence for users who signed up in the last seven days but haven't completed [core action]. Email one goes out immediately after signup, focuses on getting them to [core action], and explains the value in one sentence. Email two goes out on day three if they haven't completed it, addresses the most common obstacle, and provides a specific solution. Email three goes out on day six and is short, direct, and ends with an easy CTA. Use the [template name] template. Subject line limit: 50 characters.
Win-back sequence
Create a two-email win-back sequence for customers who haven't logged in for [X] days but previously completed onboarding. Email one is friendly and low pressure, highlights recent improvements, and carries no urgency. Email two goes out seven days later, is more direct, states a clear value proposition, and includes an easy unsubscribe option that shows we respect their decision. Use the retention text template. Keep each email under 150 words.
Post-purchase follow-up
Create a post-first-purchase email for customers who completed their first order in the last 48 hours. Thank them, set expectations for what happens next, and introduce one feature or resource they don't know about yet. Tone: warm, helpful, and free of promotion. Use the [template name] template.
Feature adoption campaign
Create a two-email feature education sequence for users who have been active for 30+ days but have never used [feature name]. Email one explains what the feature does and when it's useful, leading with the outcome and saving the feature name for later. Email two goes out five days later if they haven't engaged and shows a quick how-to or links to a short walkthrough. Suppress both emails if they use the feature after email one sends.
Prompts for building and refining segments
Activation-risk segment
Build a segment called "Activation risk: week one" that includes users who signed up in the last seven days, have logged in at least once, and have not completed [core action]. Exclude users who are already in an active onboarding campaign.
Churn-risk segment
Create a segment of users who were active monthly in [month] and [month] but had zero events in the last 30 days. Name it "Churn risk: recently inactive." Add a note to the segment description explaining the criteria.
High-intent, not yet converted
Find users who visited the pricing page in the last 14 days, have an engagement score above 5, and have not received a trial conversion email in the last 30 days. Call this segment "Warm: pricing intent."
Upsell candidates
Build a segment of customers on the [plan name] plan who have used [feature X] more than five times in the last 30 days and have never seen the upgrade flow. Name it "Feature power users: upsell ready."
Prompts for copy and messaging
Subject line variants
Write five subject line options for a win-back email going to customers inactive for 45 days. Vary the emotional angle: one curiosity-based, one value-focused, one direct and low-key, one that acknowledges the absence, and one that leads with a recent improvement. Character limit: 50. No clickbait.
Email rewrite for a different segment
Here's the copy from our standard onboarding email: [paste copy]. Rewrite it for customers in the enterprise segment with the same core message, a more professional tone, and outcomes-focused framing in place of startup-friendly language. Keep it under 200 words.
SMS messages for a time-sensitive campaign
Write three SMS message variants for a 48-hour flash sale ending [date]. Each should be under 160 characters. One leads with the discount, one leads with urgency, and one leads with the product benefit. Include a short URL placeholder.
Post-event follow-up
Write a follow-up email for customers who attended our [event name] webinar. Thank them for joining, summarize the one key takeaway in one sentence, and link to the recording. Include a soft CTA to book a demo. Keep it under 150 words and friendly, not formal.
Prompts for performance analysis
Campaign drop-off diagnosis
Look at the performance data for [campaign name]. Walk me through open rate, click rate, and conversion rate for each step. Where is the biggest drop-off, what are the most common causes, and what would you change first?
Segment performance comparison
Compare the open and click rates for [segment A] and [segment B] across the last three months. Which segment is more engaged, and is there a pattern by day of week or time of send?
User journey audit
Tell me about [user email or ID]: their profile attributes, recent events, the campaigns they're currently in, and any campaigns they've received in the last 30 days. Give me an overview of where they are in the lifecycle.
Subject line test analysis
Here are the results from our last A/B subject line test: [paste results]. Which variant performed better, why, and what should we test next based on this?
Prompts for testing and optimization
A/B test setup
We're running an A/B test on [campaign name]. The current subject line is [X]. Generate three alternative subject lines to test against it, each with a different hypothesis about why it will outperform: one shorter, one with personalization, and one that leads with a different emotional angle. Note the hypothesis for each.
Journey optimization
Look at the step-by-step performance of [journey name]. Identify the step with the lowest click-to-open rate, explain the most common reason for the drop-off at that step, and suggest two specific copy or timing changes to test.
Cadence review
Our [campaign name] sends on days 1, 4, 7, and 14. Open rates drop significantly after day 7. What alternative timing would you test first, and why?
Prompts for operations, growth, and sales teams
Operations and admin
Find all campaigns that haven't sent a message in the last 60 days and are still in draft status. List them with their last-edited date and the team member who created them. Flag any that appear to be duplicates of active campaigns.
Growth teams running activation experiments
Create a three-step activation experiment for users who complete [trigger event] but don't return within 48 hours. Step one is an in-app nudge on their next login. Step two is an email on day three. Step three is an email on day seven with a different angle. Build the segment logic and draft the messaging for each step.
Teams supporting a sales motion
Build a segment of accounts where the primary contact has opened at least two emails in the last 30 days but the account has no logged sales activity in the same period. Name it "Warm contact: no sales follow-up." Add a note with the segment logic.
What should you build with the Agent first?
Start with high-frequency, low-stakes tasks that are tedious to do manually, because those are where the Agent's value shows up fastest and where your team builds prompting instincts that transfer to harder work.
Good starting points:
- Subject line variants: Give the Agent a campaign brief and ask for five options with a rationale for each. The task is fast, reviewable, and low-stakes.
- Segment building: Ask the Agent to build a segment in plain language, then review the logic before activating. The exercise shows how the Agent interprets your event and attribute structure.
- Copy editing: Paste in a draft and ask the Agent to improve clarity, fix grammar, or rewrite a section for a different audience. The scope is controlled, and the result is easy to evaluate.
- User profile lookup: Ask the Agent about a specific user's attributes, event history, and current campaigns. This is one of the most useful diagnostic uses, and it's read-only.
Save these for later: bulk operations on live campaigns, complex multi-step sequences before you've reviewed how the Agent structures journeys, and any task where a mistake is expensive to reverse, such as sent emails or bulk segment updates.
How do you scale Agent use across your team?
Scaling Agent use comes down to shared prompts, custom skills, and clear rules about what the Agent is allowed to do. Once individuals are getting value, these three practices keep everyone from rebuilding the same workflows.
Shared prompt templates: Keep a shared document of proven prompts for your most common tasks, such as win-back emails, activation sequences, segment logic, and performance diagnostics. New team members start from a working baseline.
Custom skills for recurring workflows: If your team runs the same type of task regularly, such as weekly newsletters, campaign briefs, or post-event emails, encode it as a custom skill (https://docs.customer.io/) so everyone gets consistent output.
Clear guidelines on permissions and review: Set the Agent's permissions, communicate them to the team, and establish a norm of review before any Agent output goes live. A short internal guide, even a single Notion page, keeps the first mistake from repeating.
Ready to see what the Agent can build for your workspace? Read more about your very own AI Agent.
FAQs
What data does the Agent need to work well? The Agent needs clean, consistently named events and complete customer attributes. It reasons about what exists in your workspace, so a stronger data foundation produces more precise output.
Can I use the Agent without a developer? Yes. Day-to-day use is prompt-based, and building campaigns, creating emails, and analyzing performance require no technical skills. Initial configuration, such as connecting integrations and setting permissions, can require admin access.
Is the Agent trained on my data? The Agent uses your workspace data, including segments, attributes, campaign history, and customer profiles, during a conversation to produce relevant output. That data informs the conversation and doesn't train the model.
How do I give the Agent my brand voice? Include a short brand context description in your prompt, or build a custom skill that encodes your voice, tone guidelines, and example copy. More specific context produces more consistent output.
What happens if the Agent gets something wrong? For draft content, review and edit in Design Studio or ask the Agent to revise. For live resources, keep the Allow Agent to edit live data setting off or on the approval option so changes are reviewed first. If a change goes through that you didn't intend, ask the Agent to revert it or correct it manually.
