AI Agents in Marketing: Workflow or Agent? A Decision Guide

Vendors now call almost every automation an agent. The label sells well, and it hides a choice that costs real money: whether a marketing task needs a model that picks its own next move, or a fixed sequence with a model inside a few of its steps. Get that choice wrong and you pay for variance you did not need.

Two readers should find this useful: the leader fielding vendor pitches for "AI agents", and the operator who has to deliver something dependable. What follows is a short test for any task, a sort of everyday marketing jobs by outcome, the cost evidence from Anthropic's own engineers and the safeguards each type needs.

The vocabulary builders actually use

Anthropic's engineering team separates two things. A workflow orchestrates models and tools through predefined code paths. An agent is a system in which the model dynamically directs its own processes and tool usage. The same post advises the simplest solution possible and adds complexity only when it pays. Agents buy flexibility, and the price is predictability.

Here is a quick way to puncture a vendor claim. Ask what picks the step that comes after this one. A rule somebody wrote means a workflow, whatever the brochure says, and workflows may call a model at several points for drafting or classifying without becoming agents.

What an agent costs that a workflow does not

The cost argument is stronger than the risk argument, and it comes from the builders. Anthropic described its multi-agent research system and reported that agents use about four times the tokens of a chat interaction, and multi-agent systems about fifteen times. The same write-up says such systems suit tasks valuable enough to justify that cost, breadth-first work that splits into independent directions and information too large for one context window. It lists poor fits too: work where every agent needs the same context, or where many dependencies connect the parts.

It also records failures worth reading before you buy. Early versions spawned too many sub-agents for simple queries, duplicated work when instructions were vague, and favoured search-optimised content farms over authoritative sources until given explicit guidance. Errors compounded over long runs, and non-deterministic behaviour made debugging hard. A marketing team inherits the same traits without the engineering budget to tame them.

Four tests for any task

Can the steps be written down before work begins? If a flowchart is possible, the job is a workflow. The Friday roundup of new articles follows the same path every week.

How regular is what goes in? A feed from one system each week is regular, while open-ended research on unfamiliar sites is not.

What does a mistake cost, and how quickly would a person notice? A draft takes minutes to review, but a hundred silent edits on a live site take far longer. Anything public or hard to reverse belongs in a workflow with an approval gate.

Does the job justify extra spend? With the multipliers above, an agent must change the result enough to repay its cost.

Tally the answers truthfully and keep them. Two or more pointing to a workflow settles it. If only flexibility and value favour an agent, run a read-only pilot. Those notes later serve as your approval record.

A worked score: the newsletter assembly passes the steps test, the input test and the mistake test, and fails only the "worth extra spend" test in the agent's favour, so it stays a workflow. Market research on a new segment fails the steps test and the input test, passes the mistake test only if read-only, and may justify the spend because the answer informs a larger decision. Different scores, different builds, and none of them needed a debate about definitions.

Typical jobs and where each lands

Fixed-path jobs suit workflows: turning a long article into social posts, assembling the weekly newsletter, drafting from a brief with sources, and routing inbound questions, where a model can classify but the categories stay predictable. Publishing to the live site and emailing prospects also belong in workflows, with an approval gate, because both are public and carry brand and legal exposure. The one clear agent candidate is research into a new buyer segment, run read-only, since the sources vary and nothing gets written. For the build, the five-step workflow and the automation starter guide help.

Our call: default to the workflow and make the agent earn a seat

The industry mood runs the other way. Surveys show interest in agentic tools climbing, and SAS and Coleman Parkes found in 2025 that 51% of marketing decision-makers planned to invest in agentic AI within a year, while 21% were testing it in live environments. Interest is not evidence of fit.

Our position is that a marketing team should treat the agent as the exception that needs a written case: which step cannot be listed, what it costs in tokens and review time, and what it may touch. Most teams will find the flexible part of a task is one step, and a single model-driven step inside a fixed workflow covers it. The price of this stance is some manual design work up front and the occasional case the workflow mishandles. A team whose core work is open-ended research at volume is the exception, since the breadth-first pattern Anthropic describes is exactly the job. For publishing, outreach and reporting, default to the fixed path.

Safeguards by type

OWASP traces most harm from model-driven systems to three roots: too much functionality, too many permissions and too much autonomy. Among its preventive steps are limiting the extensions a model can call, giving downstream systems minimum permissions and requiring human approval for high-impact actions. Applied here, a workflow needs per-step tests, a log of every run and an approval before anything public. An agent also needs a scoped identity, read access first, a spending cap and a stop switch, plus a weekly review of what it did. Keep the least-privilege checklist to hand before any agent receives write access.

Two illustrations, details invented

Release notes to product pages. Each sprint someone checks which features changed and updates the matching copy. The path never varies: read the notes, find the pages, propose edits, route to the page owner. That is a workflow. An agent would add cost and variance to a job needing neither, and the only model call is the proposed rewording.

Buyer questions in an unfamiliar vertical. A strategist wants to know what purchasers in a new industry ask before choosing a supplier. Where answers live depends on what surfaces: forums, analyst pages, regulator sites. An agent that searches, opens pages and adapts fits, as long as it reads only and returns sources for a person to check. If it could also edit your site, it would hold write access the task never needed.

Warning signs that you picked the wrong type

An agent that keeps accumulating rules is telling you each rule was a step you could have listed, so convert it to a workflow. A workflow whose branches multiply faster than you can test them may be cheaper to maintain with one model-driven step. And logs that nobody reads mean an agent is a risk with a pleasant name.

The smallest dose of autonomy that works

Rather than asking whether an agent can do the job, ask how little autonomy gets it done dependably. For most B2B content teams the answer is a fixed workflow with a model at the drafting and checking stages and a person at publication. The broader view sits in the agentic marketing guide, and the people side in editorial autonomy without losing control.