Workflows Versus Agents: Where to Automate and Where to Keep a Person in the Chair

Two different machines

A workflow is a recipe. The steps run in a fixed order, each step has a defined input and output, and the same request produces the same kind of result every time. A brief comes in, a draft gets generated using your brand knowledge, a checker compares claims against your approved register, and the result lands in an editor's queue. You can draw it on a whiteboard, and you can predict what it will do on a Tuesday afternoon.

An agent is closer to a capable assistant with a goal. You describe the outcome, give it tools and boundaries, and it chooses the steps, which means two runs of the same request can take different routes. That flexibility earns its keep on tasks where nobody can write the steps in advance, such as researching an unfamiliar market or investigating why a campaign underperformed. The same flexibility makes the results harder to predict and harder to audit.

Start with a workflow

My default is to build the workflow first, for a practical reason: predictability is what lets you trust, test, and improve a process. You can measure each step, see where it fails, and fix that step alone. When a workflow produces a weak result, you know which stage to open, whereas a weak result from an agent sends you hunting through its decisions.

Most marketing tasks that people want to automate are workflows in disguise. Resizing assets for channels, turning a webinar transcript into a first-draft summary, producing product description variants from a spec sheet, tagging and filing new assets, assembling a weekly performance snapshot: each follows a path you can write down. Writing the path down is the first test, since a task you can describe in steps already belongs to a workflow.

Five questions that decide

Run any candidate task through these before building anything.

• Does the task repeat on a schedule or on demand at a steady rate? Frequency pays for the build.
• Do the inputs arrive in a consistent shape? A consistent shape lets a fixed path work.
• Does a mistake show up quickly and cost little? Cheap, visible errors suit automation, while expensive, hidden errors want a person.
• Can you define done in one sentence a stranger could check? A clear finish line lets a machine stop where it should.
• Does the output depend on judgment about relationships, taste, or timing? Judgment sets the person's seat.

Three or more yeses on the first four make a strong workflow candidate. A yes on the fifth keeps a person involved however the first four land.

Run the arithmetic

Put numbers on the decision with your own data. Take the minutes a person spends on the task each time, multiply by the number of times per month, and multiply by the loaded hourly cost of the person who does it. That figure is the monthly cost of the task today. Now estimate the minutes of human review the automated version still needs, and add the monthly cost of the tools and the one-time hours to build it. If the task costs more per month than the automated version plus its review, the build pays back, and the gap tells you how many months it takes. When the figures come out close, the task usually isn't worth automating yet, and your hours belong somewhere else.

Where an agent earns a place

Agents suit work with an uncertain path and a safe boundary. Exploratory research, competitive scans, and first-pass analysis of a messy set of customer feedback all fit, because a person reads the result before anything depends on it. Give an agent read access to sources and a place to leave its work, and keep it away from anything that sends, publishes, spends, or deletes. The boundary does the governing: an agent that can only draft and file can only make drafting and filing mistakes.

As an agent proves itself on a task, you can often harden the path it keeps choosing into a workflow. The agent explores, you notice the route it takes every time, and you freeze that route into a recipe you can test. Exploration first and automation second makes a sound sequence.

Where a person stays in the chair

I keep people in the seat wherever the work depends on trust, taste, or consequences that are hard to reverse. That list includes the positioning of a brand, the tone of a message to an upset customer, a negotiation with a partner, the final choice between two creative directions, any statement about pricing or performance, and any publication under a real person's name. A model can offer options in all of these areas, and a person decides, because the person carries the relationship and the accountability.

This division also serves your team. People who hand off the assembly work gain time for the parts of the job that only they can do, and the best use of an AI budget is to buy those hours back.

A review point inside every automated path

Wherever you automate, place a human checkpoint at the end of the path, and pick its strength from the risk tier of the output. Log what reviewers change. If editors alter forty percent of the drafts a workflow produces, the workflow needs work, and if they change nearly nothing across a long stretch, you've earned the right to loosen the check on that task. Let the edit rate guide the dial.

Your next move

List the ten tasks your team repeats most, run each through the five questions, and circle the ones with three or more yeses. Choose the circled task with the clearest steps and the cheapest mistakes, write its path on one page, and build that workflow before you touch an agent. One working workflow with a measured edit rate teaches your team more than any plan for a dozen.