Can AI agents actually run marketing campaigns? 

The answer depends on what you mean by "run," and most of the debate happens because nobody defines it.

Running a campaign isn't one job. It's at least six: deciding who to target, building the audience, writing the content, assembling the workflow, pressing send, and figuring out afterward whether any of it worked. Ask whether an agent can do all six autonomously, and you'll get a nervous "not yet." Ask which of the six an agent can own today, and the answer gets a lot more interesting.

Here's the honest version, based on what our AI Agent actually does inside Customer.io.

You describe the group you want in plain language, something like people who signed up in the last 30 days but haven't finished onboarding, and the agent translates that into a real segment against your actual attributes and events. No hunting through condition builders, no guessing which attribute holds the signup date. (If you've already been using AI to speed up your daily lifecycle workflows, this is the same idea with your live data behind it.)

Describe the outcome you want, and the agent proposes a full automation: trigger, timing, branching logic, and message content. It reviews the segments and engagement data you already have before it suggests anything, so what you get back is grounded in your workspace rather than a generic re-engagement template with the serial numbers filed off.

Ask which of your onboarding automations had the worst deliverability this month, or what the click-to-open rate is on your birthday flow, and you get an answer plus a recommendation for what to do about it. That's a task most lifecycle marketers postpone until someone asks in a meeting.

None of this is speculative. It's the daily use case, and it maps neatly to how marketers already spend their AI budget. In our 2026 customer messaging research, 61% of marketers said they use AI to write and draft copy, and 45% use it for campaign optimization. The agent covers both, with the meaningful difference being that it already knows your data.

Sending is where the design gets deliberate.

By default, the AI agent can't edit live automations, active segments, or one-time sends at all. It builds drafts. An account admin has to turn on Edit live data in your AI settings before it can touch anything running in production, and even then, you pick how changes get applied: Ask mode pauses for your approval on every change, and Auto mode routes each proposed change through a separate model that checks it against your stated intent, approves what matches, and holds back what doesn't.

The agent also operates strictly within your permissions. If your role can't create messages, neither can the agent working on your behalf. It can't reach your data warehouse, it can't browse the open web, and it can't read sensitive personal attributes.

Some people read that list and see limitations. We'd argue it's the reason the thing is usable in production. An agent that could silently push a broadcast to 400,000 people would be a tool nobody sane would enable. This is also the practical version of the governance question in our AI readiness framework: approval workflows aren't bureaucracy, they're what lets you move faster without holding your breath

If you want to see genuine unattended work, look at routines. These are recurring tasks the agent runs on a schedule with no prompting from you, delivering results to your inbox each time. A deliverability watchdog that tracks bounce rates and spam complaints. A weekly broadcast recap comparing recent sends. A segment hygiene check that flags empty or redundant conditions before they quietly break a campaign.

Routines run in safe mode, so the agent can read and analyze but can't change anything. Every run starts a fresh session, which means a routine firing at 6:00 am ET won't collide with whatever you were working on last night. If one fails three times running, it pauses itself rather than failing loudly on a loop.

This is the unglamorous version of autonomous marketing, and it's the version that pays. The audit you keep meaning to do gets done every week, whether or not you remember it.

The agent helps you build. LLM actions help the journey decide, one customer at a time.

An LLM action is a step inside a workflow that calls a model at runtime, using that specific person's data as context, and stores the result as a journey attribute. You can use that attribute to personalize the message downstream or to route the person into a different branch. Product recommendations shaped by individual behavior, sentiment scoring on a support interaction, and intent-based routing: none of that fits in a static template, and none of it needs a webhook to an external service anymore.

If you want to see what teams have built with them, we've collected real LLM action use cases and the prompts behind them, including a re-engagement email that sounds like the brand rather than software.

So the accurate picture is two different kinds of agentic work. One collaborates with you at build time. The other makes small decisions on your behalf thousands of times a day inside a running campaign. Both shipped together in our biggest release to date, alongside Goals, WhatsApp, and LINE.

The model isn't your constraint. Your metadata is.

The agent reads your attribute names, event names, and descriptions the same way a new hire would. If you have an attribute called cname with no description, the agent has no idea whether that's a company name, a customer name, or a DNS record. It'll guess, and it'll guess wrong sometimes.

Two things fix most of this:

• Fill in your business context. Workspace settings include a business profile covering your industry, audience, and tone. Every AI feature in the platform reads from it, so a thin profile produces thin output.
• Describe your data. Go through your Data Index and add descriptions to the attributes and events you actually use in targeting. You can generate the descriptions with AI and edit from there, which takes an afternoon and improves everything downstream.

Teams that prioritize this step get the most value out of the agent, so don't skip it!

It can build the segment, assemble the workflow, draft the messages, and tell you what happened afterward. It can monitor your deliverability every week without being asked. It can make per-person decisions inside a live journey.

What it won't do is decide on its own that this is the week you discount, or that a particular cohort deserves a different offer. That's still strategy, and strategy is still yours. Knowing which calls to hand over and which to keep is quickly becoming the core skill, which is the argument we make at length in how to build an AI-first marketing team.

Which is the trade most of us would take anyway? In our research, 72% of marketers said AI already saves them 20% or more of their time, and more than half said they're shipping campaigns faster because of it. The time that comes back doesn't come back as free time. It comes back as more testing, better segmentation, and campaigns that get thought about properly instead of being built at 4:00 pm on a Friday.

Pick one real project, spend twenty minutes with the agent on it, and set up a single routine to watch something you've been ignoring. If you'd rather warm up outside the platform first, our lifecycle marketer's AI prompt cookbook is a good place to steal a few ideas. Either way, you'll know within a week whether the answer is yes for your team.

Can an AI agent send messages to my customers without my approval?

Not by default. The agent can't edit live automations, active segments, or one-time sends unless an account admin turns on Edit live data in AI settings. Even then, Ask mode is the default and pauses for your approval on every change to live data. Sending otherwise follows your normal approval and sending flows.

What's the difference between the AI Agent and LLM actions?

The agent is a conversational collaborator you work with directly: ask it to build a campaign, analyze performance, or draft copy. LLM actions are a step inside a journey that fires at runtime for each individual customer, generating content or making routing decisions automatically. The agent helps you build, and LLM actions help the journey decide.

What data can the agent see?

It uses your business context, the page you're on, and workspace data like profile, event, and object attributes, plus automation and broadcast metrics. It's limited by your role and permissions, so it never has access to more than you do, and it can't read sensitive personal attributes.

Is my customer data used to train AI models?

No. Data sent to hosted models isn't used for training. Our AI features work with the names and descriptions of your data rather than the values, so a first_name attribute is visible to the agent as a field, not as the person's actual name.

Can the agent connect to my data warehouse or browse the web?

No on both. It can't query external SQL databases or data stores, and it can't visit URLs or call third-party APIs. It can help you configure integrations and connections to those services, and it can search Customer.io documentation.

How many routines can I run?

Routine limits are per user, per workspace. Essentials plans get 1 active routine per workspace with a weekly minimum interval. Premium and Enterprise get 5 active routines with a daily minimum. Every plan can create up to 10 total routines and keep the extras paused. You need Editor or Admin permissions to create one.

What are AI credits, and do I need them?

AI credits are the unit of consumption for LLM actions, not for the agent itself. Paying accounts receive a one-time grant to experiment with, and additional credits are available for purchase. Usage varies with the model you pick and the length of your prompt, context, and output. Check the AI credits documentation for current amounts and pricing.

Can I use my own AI agent instead of yours?

Yes. Whatever your team can build, launch, or analyze in Customer.io, your own agent can too, through our MCP server, the CLI, or our APIs. The MCP server suits chat-based tools like Claude and ChatGPT; the CLI suits terminal work and coding agents.