Most marketing teams now run an AI agent somewhere in the workflow. Few have written down what that agent may decide on its own. We think every agent needs a one-page job description, the same way every new hire does.
We build an AI marketer, so we have a stake in this. Read what follows with that in mind.
Agents are everywhere. Agreement is not.
Genpact published a survey on 12 November 2026. It asked more than 2,000 executives at companies with revenues of $1 billion or more, across 17 industries, how they use agents. Genpact says 79% have agents in production, but only 20% use them broadly and consistently across functions.
Three caveats apply. The findings are self-reported. The release gives no regions or fieldwork dates, has no marketing-specific data, and comes from a company that sells agentic operations services.
Even so, the gap is easy to believe. Putting an agent into production is a purchase. Using it consistently takes decisions that someone has to make.
Autonomy is a design choice
The same release says AI leads or fully runs 40% of decisions. It also says humans still decide or audit in at least 87% of cases at every decision stage. Genpact adds that autonomy in practice “is not full autonomy”.
The release does not break the 87% down by stage, so treat it as a direction, not a benchmark.
Agents do not decide their own scope. A person sets it, or the tool’s defaults do.
The useful reading for marketers is simple. How much an agent does on its own is a choice. If nobody wrote the choice down, it was made by default.
Genpact also names two new human roles: exception specialists and agent trainers. It gives no numbers for either. Our reading is that both jobs start with knowing what the agent was meant to do.
What goes in the job description
Keep it to one page, in words a new colleague could follow. Five lines, plus one name.
- The task. One sentence on what the agent produces, for which channel and for whom.
- The reference. The brand material it must work from, named and dated. If it cannot see your brand guide, it will guess.
- Decisions it makes alone. Caption length, which approved image to pair, a posting slot inside an agreed window.
- Decisions it hands to a person. Claims, prices, anything about a customer, anything sensitive in the news that day, any new tone.
- Stop conditions. When it must stop and ask: a missing fact, two instructions that conflict, a banned term it cannot avoid.
The name is the owner: one person who reads the output and can say no.
Why a page and not just a prompt? A prompt sits inside one tool. A page can be read by the owner, by whoever built the workflow, and by a new colleague on their first day.
A worked example
Imagine an agent that drafts three LinkedIn posts a week for a B2B brand. The task is clear: three drafts from the week’s approved topics. The reference is the brand voice page and the ban list.
It decides alone on structure, the opening line and how many hashtags to use.
It hands to a person any number, any claim about results and any mention of a competitor. It stops and asks if a topic arrives without a source. The owner is the head of marketing.
Notice what the page does not say. It does not promise the agent will always be right, and it does not ask for perfection. It sets a boundary and names who checks.
That page is short enough to read in a few minutes. It also gives you something to check the output against, which a vague “keep it on brand” never does.
How Mavic handles the same problem
For the reference line, Mavic uses Brand DNA: a saved record of your voice, visuals, products, audiences, banned words and writing styles. Mavic reads it when planning any request, so you do not re-brief each time. We say it always works from your brand.
Tandem, the engine behind Mavic, splits a request across specialist agents, such as a planner, a brand writer and a brand designer. Each receives the fields it needs from the same Brand DNA. You can read how it works on the Tandem page.
For the hand-to-a-person line, Mavic shows you a plan and asks before it creates anything. Nothing is published until you approve it.
There are limits. The ban list is a strong nudge, not a hard filter, and banned terms can still appear, especially in longer documents. Your job description still has to say which decisions go to a person, and someone still has to check the work against it.
Start this week
Pick the agent or automated workflow that touches the most public content. Write its five lines and name its owner in half an hour.
Then run one week of its output past the owner, and mark every item that broke a line. Each break is either a gap in the page or an agent that stepped outside it. Fix the page first.
Repeat monthly. Look at what the owner rejected, and turn each repeated reason into a new line. The page gets sharper, and the agent gets less room to guess.
