AI Agents for Marketing: Where They Break

"AI agent" gets marketed as a system that can run a marketing function end to end with nobody watching. In practice, agents break at three specific, well-documented points, and knowing where lets you build the right check instead of hoping the model gets smarter.

Where it actually breaks

Unverified output reaching an audience. This is the failure mode that does the most damage per incident. An agent can draft a ranking claim, a statistic, or a customer-facing number that sounds correct and is not, and if nothing checks that claim against real data before it publishes, the error reaches an audience at whatever speed the agent runs, not the speed a human would have caught it. This is why production agent setups route anything customer-facing through a separate verification pass and a human approval step rather than letting the drafting step also be the publishing step, the same reasoning behind sorting AI tools by what they actually do instead of what they claim in AI SEO Tools, Honestly Tested.

Reliability across multi-step tasks. Marketing work is rarely one step. Pulling data, deciding what it means, drafting a fix, and formatting it for review is already four steps chained together, and agent reliability degrades as that chain gets longer. Carnegie Mellon's TheAgentCompany benchmark tested AI agents on realistic office tasks, browsing, coding, running software, communicating with a simulated team, and found that Gemini 2.5 Pro, the best performer on the benchmark, completed only about 30 percent of tasks to full completion, with most other tested models scoring lower (TheAgentCompany paper; reported by The Register). That is a ceiling on how much of a marketing workflow can run without a checkpoint, not a floor that will simply rise with the next model release.

Bad input data the agent cannot catch on its own. An agent that pulls from a stale keyword export, a broken tracking pixel, or an outdated CRM field will act on that number with the same confidence it would show for a correct one, because nothing in the task tells it the input is wrong. The agent inherits whatever quality the data source already has. This is also why Gartner's research on the category is blunt about outcomes: the firm predicts over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls as the drivers, not a lack of model capability (Gartner).

The fix is process, not a smarter model

None of the three break points above get solved by waiting for a better language model. They get solved by adding a check: a verification pass that compares every claim in a draft against a real data source before publishing is allowed, and a human approval gate on anything that reaches a live audience or spends money. That is also the operating model behind this zine itself, every article here is drafted by a scheduled agent and sits in a review queue until a person approves it, which is the same gate this article is describing. For the broader case on why that split between autonomous drafting and human approval matters, see What Is Generative Engine Optimization? If you want a real, current snapshot of where an agent-run process would actually find gaps in a specific site, that is what YaBaDo's GEO audit produces.

Sources: TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, AI agents get office tasks wrong around 70% of the time (The Register), Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027