AI agents for marketing: what they do and where to use them 

"AI agent" has started to change its meaning over the last year. It gets used for a chat window, for a tool that writes and ships code, and for software that runs on a schedule and touches production systems. Those are three different things with distinct risk profiles, and treating them as a single category is why so many teams end up with an AI strategy that's really just a subscription.

So here's a more useful frame. There are three kinds of AI agents a marketer will actually encounter. They're good at different parts of the job, so let's unpack where each one belongs in your strategy.

Claude, ChatGPT, Gemini. You paste in context; they give you back thinking: campaign angles, subject line variants, a landing page rewrite, a summary of 400 survey responses. They're excellent at judgment-adjacent work, and they know nothing about your customers unless you tell them. This is where most marketing AI use still lives, and it's genuinely valuable.

Our AI prompt cookbook and 10 Claude prompts for marketers are both built around this tier.

Claude Code, Cursor, and the rest. Give one an API key and a clear brief, and it will stand up a working thing: a webhook listener, a data transform, a script that sends an email when a row changes. This tier is newer for marketers, and it's the one causing the most confusion, which we'll get to.

An agent that lives inside the system where your customer data, your segments, and your sending are already. It doesn't need you to paste in context because it can read your workspace directly. Customer.io's AI agent is in this tier, and so are the agentic pieces inside an automation, like LLM actions.

The difference that matters across all three is context and consequence. Knowing when to use which agent, and how, helps you ensure both the agent and the human-in-the-loop are set up for success.

An assistant is designed for low-stakes exploration, where speed and flexibility matter most. A platform-native agent combines deep context with the ability to act, making guardrails a feature of responsible execution. A coding agent occupies the space between them: it can build something real and valuable, while relying on human judgment to ensure that what it builds remains useful, maintainable, and resilient over time.

Strip away the category talk, and the work breaks into jobs. This reflects the framework we used when we asked whether AI agents can actually run marketing campaigns, and it holds up here.

Reading survey data, summarizing customer feedback, finding the pattern in a churn cohort, pulling the deliverability numbers you've been avoiding. Agents are strong here, and it's underused, mostly because analysis is the task people postpone rather than delegate.

Describing a group in plain language and getting a working segment back. "People who signed up in the last 30 days but haven't finished onboarding" becomes real conditions against your real attributes. This only works in the platform tier because an assistant doesn't know your schema.

Drafts, variants, translations, tone adjustments. The most mature use case by a wide margin. In our customer messaging research, 61% of marketers said they use AI for writing and drafting copy, the single most common use case.

This one splits based on the copy's purpose.

Concepting, positioning, and finding the range of possible angles work fine in an assistant, since none of that needs your customer data. But once you're writing an actual message that's going to ship, do it in the platform tier: the agent already knows your brand voice from your business profile, it can personalize against real attributes instead of leaving you placeholder brackets to fill in later, and the draft lands directly in the message rather than in a document you then have to copy across.

Turning an intent into a full automation: trigger, timing, branches, messages. Platform tier again, since the agent needs to know what channels and data you have.

This is the one people miss. An LLM action is a step inside a live workflow that calls a model at runtime for each individual person and stores the answer as an attribute you can personalize or branch on. Sentiment scoring, intent classification, and product recommendations are shaped by one person's actual behavior. We collected real examples and the prompts behind them, including a re-engagement email that sounds like the brand instead of like software.

Recurring checks that run without you: bounce rate watchdogs, weekly broadcast recaps, segment hygiene audits. In Customer.io, these are routines, and they're the least exciting item on this list and probably the highest return, because the audit you keep meaning to do gets done whether or not you remember it.

Notice that the middle four are hard to do well without access to customer data, which is the practical reason platform-native agents exist at all.

If you want to work from your own tooling rather than ours, that's supported too. Anything your team can build, launch, or analyze in Customer.io, your own agent can do through our MCP server or CLI. MCP fits chat-based tools like Claude and ChatGPT. The CLI fits terminal work and coding agents.

You can also use Customer.io's agent in Slack to seamlessly ask questions about your data and workspace.

Somewhere in the last year, coding agents got good enough that a reasonable person started asking: Why am I paying for a messaging platform when I could have an agent build this?

Here's the honest answer. If you need one low-volume email triggered by one event, with no compliance requirement and no second channel on the horizon, building could be fine. An agent will get you a trigger, a template, and a send call in an afternoon, and it'll work.

What that afternoon doesn't get you is the operational layer. And the distinction isn't about how good the code is, it's about what kind of problem you're solving. Some things you write once. Others you run continuously.

IP reputation, domain warming, bounce and complaint monitoring, ISP relationships. It degrades if nobody's watching, and it doesn't degrade in a way that shows up as an error message.

Opt-in and opt-out tracked per channel and per jurisdiction, with an audit trail, as the rules keep moving. This is the single hardest thing to retrofit, because by the time you need it corrected, you already have years of records that were captured incorrectly.

SMS alone means carrier registration and routing rules that differ by country before you've sent anything. WhatsApp adds Business API requirements, template approval, and a session window that doesn't exist elsewhere. Push and in-app have their own permission models. And once you have two channels, they don't just sit side by side; they have to know about each other, or you send the same person the same thing twice.

Customer.io customers sent more than 100 billion messages in 2025, across 9,000+ companies in over 120 countries, at 99.8% infrastructure uptime. A system designed for a few hundred sends per day usually needs a rebuild to handle that, and that rebuild tends to happen mid-growth under pressure rather than on a calm timeline.

Then there's the certification question. SOC 2 Type II, ISO 27001, HIPAA-readiness, GDPR compliance with EU data residency. Those aren't features you add; they're multi-year programs, and a self-built system starts from zero on all of them.

Worth noting: Lovable and Cursor both run their messaging on Customer.io, alongside companies like Angi and Notion. Teams whose entire product is AI-assisted building still chose not to build this particular thing.

The build-versus-buy framing assumes a fork in the road. In practice, most teams doing this well are using both, on different parts of the job.

They draft and test copy with an assistant. They prototype campaign logic in a chat window. They use a coding agent for the integration nobody wants to hand-write. Then they run all of it on infrastructure that handles delivery, consent, and orchestration underneath. Our own AI agent, CLI, and MCP integration exist to make that pairing native rather than a workaround.

That's the pattern to copy. Use agents for the parts that are logic and language. Keep the parts that are operations on something built to operate.

None of the three tiers decides that this is the week you discount, or that a particular cohort has earned a different offer, or that the campaign you're about to ship is beneath your brand. Agents execute against intent. Supplying the intent is the job, and it's getting more valuable rather than less. We made that case at length in showing how to build an AI-first marketing team.

The other thing that stays yours is data hygiene, and it's the most common reason agent projects disappoint. A platform-native agent reads your attribute and event names the way a new hire would. An attribute called cname with no description is a coin flip between company name, customer name, and a DNS record. Fill in your business context and describe the fields you actually target. Teams that skip it conclude the agent isn't very good. Teams that do it report the opposite. Our AI readiness framework walks through the rest of the diagnostic.

Pick one job from the list of six, not an entire strategy.

If you've never used an agent on real marketing work, start with analysis, because it's the lowest-stakes and the most immediately convincing. Ask for the worst-performing automation in your workspace and why.

If you're already using assistants daily, move one job down into the platform tier. Build a segment by describing it. Set up a single routine to watch something you've been ignoring.

If you're weighing whether to build, don't argue about it abstractly. Write down which of the four operational items above you'd own: deliverability, consent, multi-channel coordination, and certification. If the answer to any of them is "we'd figure it out later," you've answered the question.

Broadly, software that takes an instruction and carries out multi-step work toward it rather than just returning text. In practice, marketers deal with three types: general-purpose assistants you paste context into, coding agents that build software, and platform-native agents that operate inside your marketing tools with access to your customer data.

The line is action and context. An assistant responds within a conversation and knows only what you tell it. An agent takes steps against real systems, so it can read your workspace, build a segment, draft a campaign, or run on a schedule without being prompted. The consequence of a mistake is correspondingly higher, which is why agents who touch production need approval controls.

In Customer.io, the agent builds drafts and can edit live automations, active segments, or one-time sends once an admin enables Edit live data in AI settings. Users will have the option to be in Ask or Auto mode.

The agent also can't exceed your own role permissions, access your data warehouse, or browse the web, so you can feel safe knowing the agent is acting within set parameters, even on Auto mode.

If it's one low-volume email, one trigger, no compliance requirement, and no second channel, building is reasonable. The moment any one of those changes, you're looking at a different project, not a bigger version of the same one. The tripwires to watch for are adding a channel, needing an audited certification, and realizing your data model wasn't built for the volume you've grown into.

Both, and most teams getting real value are doing exactly that. Our MCP server connects chat-based tools directly to your workspace, and the CLI covers terminal and coding-agent workflows. Anything your team can do in the platform, your own agent can do through those.

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

Three places, consistently. Poor data descriptions that make the agent guess what your fields mean. Missing business context, which produces generic output. And being handed a strategy instead of execution, which is where you get campaigns that are fluent and pointless. The first two are fixable in an afternoon.