AI Agents for Marketing: The Use Cases That Actually Pay
The short version
An agent pays when a person can check its work faster than they could do the work. That one test explains most of what we've automated successfully and everything we've turned off. Catalog-scale data, recurring audits, and monitoring nobody has time for all clear it. Brand copy, one-off projects, and anything judged on taste do not, and they cost several times a normal AI bill to get wrong.
About three quarters of marketing organizations now use AI somewhere in their work. Only 13% of marketers use agentic AI, according to Salesforce's survey of 4,450 marketers. Most coverage treats that gap as a lag, something everyone else closes once they get comfortable.
Having spent the past year building agents that do real client work, I read it differently. Plenty of marketing work is better served by a good prompt, a saved workflow, or a person. The teams sitting at 13% may be further along at telling those apart.
This is the practical version of the question: which use cases have earned their keep for us, which ones looked promising and didn't, and what separates the two before you spend a quarter finding out. It's rung four of the Claude adoption ladder, written from inside a fleet that's been running long enough to have a few failures in it.
Most of what gets called an agent isn't one
Anthropic's engineering team drew the line in December 2024, and it still holds up. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths." Agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."
The difference matters commercially, because you're buying two different things. A workflow does the same steps every time, so you can predict what it costs and what it produces. An agent decides its own steps. That's what you want when the path can't be written down in advance, and it's also the reason an agent can do something you didn't plan for.
Their recommendation is worth repeating to anyone shopping for agent platforms right now: find the simplest solution that works, and add complexity only when the evidence demands it. Most marketing tasks that people describe as agent work are a fixed sequence someone hasn't written down yet. Write it down and you have a workflow, which costs less to run and doesn't drift.
The three questions worth asking first
Before we build one, the task has to survive three questions. They came out of things we got wrong.
Can a human verify the output faster than producing it? This is the question that decides it. An agent that drafts a page still needs someone to read the page, and reading it carefully takes most of the time writing it would have. An agent that checks 3,000 product records against a rule produces something a person can spot-check in five minutes. The first saves you little time. The second saves someone a day of work.
Does the task recur on a schedule? Setup is the real expense. Defining what good looks like, wiring permissions, building the review step, watching the first dozen runs. Spread that over a weekly job and it amortizes in a month. Spend it on something you'll run twice and you never get the setup time back.
Is being wrong survivable? It will be wrong sometimes, so the useful question is what happens then. Can you catch it before it reaches a customer, and can you undo it after? Everything our agents ship carries the commit that shipped it, so we can undo any change on request. Work that can't be set up that way stays a draft for a person to approve.
A task that fails any of these can still use AI, in a prompt or a workflow, with a person running it.
What has paid off
Four categories, roughly in order of how confident I am in them.
Data work at catalog scale
Product feeds are the clearest win we've had. A catalog with a few thousand SKUs has attribute gaps that nobody will ever work through manually, because the work is dull and the payoff per row is small. Across the feeds we manage, agent work has put enrichment onto several thousand products that were never going to get it by hand.
It clears all three tests. The rule is explicit, so checking the work is a query rather than a judgment call. It recurs every time the catalog changes, and a wrong attribute is fixable in the next sync.
Recurring audits with a mechanical answer
Anything where the question has a right answer and the volume is what makes it painful. Missing meta descriptions, broken internal links, pages that fell out of the sitemap, alt text gaps, schema that stopped validating. Our site agent runs this continuously, applies the small reversible fixes itself, and files everything else for a human.
What keeps it safe is the line between those two piles. Fixing a broken link is mechanical. Rewriting the sentence around it is a judgment call about voice, so that goes into a draft for a person to approve.
Monitoring nobody has time for
The case for a lot of agent work rests on what it replaces, and often what it replaces is nothing at all. Nobody was checking weekly whether a competitor's pricing page changed, or whether the brand shows up in AI answers for the queries that matter. Those checks get scheduled with good intentions and skipped for a quarter.
An agent doing that badly still beats an empty calendar slot, as long as you remember how low that bar is. I hold this work to a lower standard than client deliverables and a higher one than I did at first, because it is easy to keep paying for a monitoring report nobody reads.
First drafts inside a fixed template
This one pays when the output has a defined shape: a campaign brief with known fields, a report against the same metrics every month, a structured audit. An agent gets you to a solid starting draft, and the reviewer edits instead of staring at a blank page.
When the shape is open, the math flips. Somebody rewrites it, and you paid agent prices for a first draft that got thrown away. We use agents to draft against a template and people to decide what the template should be.
What has not paid
Final brand copy. An agent can produce competent copy. Whether it's the right copy is a taste judgment, and taste is what you can't write into a rubric for the machine to check itself against. We draft with AI constantly, and a person still makes the final call.
One-off projects. The setup cost never comes back. If you're going to run it once, run it as a conversation with a good prompt and a person steering.
Anything requiring context that lives in someone's head. If the reason a decision goes one way is a conversation with a client three months ago, the agent will produce something defensible and wrong. That is worse than an obvious failure, because a defensible wrong answer tends to get approved.
Work with no owner. Every agent we run has a person who reads its output and is accountable for it. The ones that drifted furthest were the ones where nobody in particular owned the output.
What agents cost
Agents cost more than chat, and multi-agent setups cost considerably more than that. Anthropic's January 2026 guidance puts multi-agent implementations at "3-10x more tokens than single-agent approaches for equivalent tasks." The overhead goes into duplicated context, coordination between agents, and summarizing results back up.
Their advice is to start with one agent, on the grounds that a well designed single agent with the right tools gets further than most teams expect. Our experience matches. Every time we've been tempted to split a job across specialists, the version that worked was one agent with better tools and a clearer brief.
Against that cost, put the labor the agent replaces, counted as hours somebody spent rather than hours you picture them spending. If nobody was doing the work, the comparison is against zero, and the question is whether the output is worth its bill on its own.
Two costs get left out of most business cases. The first is review time, which never goes to zero and grows with the number of agents you run, until it needs a layer that reads the agents' work for you. The second is the cost of a mistake reaching a customer, small in a product feed and large on a landing page, which should decide how much autonomy each agent gets.
How to find out cheaply
The way we test a candidate now takes about a week.
Run the task manually five times and time it, including review. Write down what "good" means, specifically enough that someone else could grade an output with it. If you can't write that down, the agent has nothing to aim at and you have your answer. Then build the narrowest version that does one slice, run it alongside the manual process, and compare review time rather than run time.
Run time is the number vendors demonstrate. Review time is the number that decides whether you keep it. The industry data points the same way: McKinsey's 2025 state of AI survey found 23% of organizations scaling an agentic system somewhere, with another 39% experimenting, and no more than 10% scaling agents in any individual business function. Most of the distance between experimenting and scaling comes down to this problem.
Where to start
Pick the dullest recurring task you have with a right answer and a rule you can state out loud. Feed hygiene, link checks, a monthly report against fixed metrics. Build one agent for it, keep a person on the output, and let it run for a month before you decide anything.
What you learn in that month is worth more than the automation. You find out how you specify good work, how long review takes, and where your data is too messy for any of this. Teams that start with the exciting use case learn the same lessons more expensively, and usually in front of a client.
We build and run this stack with brand marketing teams, and we're glad to give you a straight read on which of your workflows would survive the three questions. Start a conversation.
