Most conversations about an AI marketing agent begin with capability. It can write, it can analyse, it can schedule, it can optimise. All of that is true, and as a way of deciding what to put in your marketing function it is as useful as reading a candidate's list of skills. The question for a hire is what the job is. The same question works for an agent.
A weekly job description is a short document. It says what the agent does each week, with what, under which rules, who looks at the result and when it stops and asks. Writing one is the quickest way to find out whether you are ready for an agent, and the only way to keep it from drifting into a role nobody chose.
For each task the agent has, five lines.
The task. One sentence in the form a manager would use: "Draft the weekly newsletter from the week's published posts."
The inputs. What it reads: the content calendar, the published posts, the audience list, the brand voice rules. If an input is missing the task does not start.
The rule. What it must and must not do: the length, the tone, the claims that need a source, the words that are out. Rules are written as checks that can be run, not as adjectives.
The reviewer and the clock. Who looks at the output, within how long, and what they do with it: approve, edit, reject. A task with no reviewer is a task with no job description.
The escalation. The conditions under which the agent stops and hands over: a claim it cannot source, a number that moved outside its normal range, a request that touches something outside its list. The escalation is as important as the task, because it defines the edge of the role.
Six to ten tasks fit in a page. A typical week for a marketing agent: assemble the weekly performance report; draft the newsletter; propose the social posts for the following week from published content; clean and deduplicate new contacts; check every link in outgoing material; summarise what competitors published. Each is a row.
Task: assemble the weekly performance report. Inputs: last week's campaign figures from the ad accounts, pipeline created from the CRM, and the previous report for comparison. Rule: every figure carries its source and date range; a figure that differs from last week's by more than a set margin is flagged with the likely cause; no figure is rounded in a direction that flatters the result. Reviewer: the marketing manager, by ten o'clock on Monday, who approves, edits or rejects. Escalation: if any source is missing or a number cannot be reconciled with the CRM, the agent says so at the top and sends nothing.
That is eighty words and it settles most of the questions a new report would raise. Notice what is absent. It does not say what the agent is. It says what is expected of the report.
Every job description has a list of things the agent does not do alone. For marketing the list is short and stable.
Anything that spends money. Anything that publishes a claim about the company, a customer or a result that has not appeared before. Anything sent to an individual customer in an exception case, such as a complaint or a cancellation. Anything touching legal, contractual or regulated language. And any change to the rules of the job itself.
We went through the spend case in nobody should let an agent near a budget: the agent can propose a spend and a person signs it. The same pattern applies to the other four. The agent drafts and checks; a person signs. The list is the job description's most important page, because it is where accountability sits.
The question every manager asks of a new hire is how much supervision they need and when that will change. The same question needs a number for an agent, and the most useful one is simple: of the outputs a reviewer saw this month, what share did they accept without changes?
Count it per task. A newsletter draft accepted unchanged in nine of ten weeks is a task that has proved itself. One edited every time is a task whose rules are wrong, or whose inputs are missing, and the edits are the evidence of what to fix. A rising rate is a sign that the rules are working. A falling rate is the first sign something upstream changed.
It is not a perfect measure, because reviewers can get lazy and approve without reading. A second number guards against that: the rate at which reviewers catch a planted or sampled error. A reviewer who accepts everything, including a deliberate mistake, is not reviewing.
The acceptance rate gives you a way to loosen review without trusting the agent on faith. Set a rule in the job description: a task moves from full review to sampled review when its acceptance rate has been above a threshold for a stated number of weeks, and moves back if it falls. Ninety percent over eight weeks is a reasonable example, and the numbers are yours to choose.
Sampled review means the reviewer sees one in five outputs instead of all of them. The agent's work still appears in the log and the person can look at any of it. What changes is where the human time goes. A marketing manager who reviewed forty items a week reviews eight, and spends the difference on judgement the agent cannot provide.
The same logic runs backwards. A task whose rate falls below a floor goes back to full review automatically, with a note to the owner. Authority is lent, measured and recalled, which is how responsibility is handled for people.
In the first month nearly every task is on full review, and the reviewer's edits are the most valuable output. Each edit is classed: a factual correction, a tone change, a missing input or a rule that was too loose. The classes tell you what to change in the job description. After four weeks, the rules are better, the inputs are complete, and the acceptance rate on the tasks that suit an agent starts to climb. The tasks that do not suit one stay on full review, or go back to a person, and the job description says so plainly.
A written job description makes the human roles clearer too. The marketing manager's job becomes what the agent's is not: setting the rules, reviewing the exceptions, signing the spend and owning the weekly numbers. The tasks the agent does are visible as a list, so it is clear which parts of the function are automated and which are not.
It also makes it possible to compare the function with the alternatives. The all-in cost per task of the agent is a number, and we put one on a whole marketing function in what it costs to run a marketing function as a system. The job description is the list the cost is divided across.
We write the job description with a company's marketing owner, run the agent against it, measure acceptance by task and move review levels on the evidence, with spend and publication signed by your people. Discovery is no-risk: we draft the job description for the first six tasks and show what review each would need on day one, and you pay only if you go ahead. The build carries a guaranteed estimate, so if we estimate low, we absorb the difference.
An agent is not given a title and hoped for the best. It is given a job, measured on it and trusted a task at a time.
Watch the 6-minute guided demo, then book a walkthrough: we read your CRM, ad accounts and content and show what a governed system would do differently. You pay only if you proceed. Or see how we approach it.
More on the same problem:
"Something isn't working." "We're outgrowing our tools." "We're afraid to make the wrong move." No-Risk Discovery is a short, practical conversation that gets you clarity before you commit to anything big. We'll tell you if we're a fit. If we're not, we'll tell you that too.
Say who you are and ask. A person replies within one business day.
