Monday morning at Gotcha, the fictional rental company with 50 vehicles from part 1. The weekend is over and the vans are coming back. A damaged sliding door, two customers who didn't fill the tank, a business customer who wants to keep his van another week. By noon the owner still hasn't had lunch.
His to-do list has said the same thing for three weeks: "Do something with the website." Since part 2 he has a company brain, a document with his facts, his promise and the words he never uses. But a document doesn't do anything. Someone has to open it, ask an AI assistant a question and do something with the answer. On Monday that's nobody, and on Tuesday too.
That's what an AI agent for marketing is for: a job that starts by itself every week, without anyone pressing a button. This is harness v1. One agent, one job, one fixed moment when you decide.
What an AI agent for marketing is
The difference from a chatbot isn't how clever it is, but who starts. A chatbot waits until you ask it something. An agent starts by itself, at a fixed time, with a fixed brief. It reads what it needs to read, does its work and leaves the result ready for you.
An agent doesn't have to be big. Most stories about AI agents are about systems that launch campaigns and split budgets on their own. That's not where you start. You start with one small, boring, recurring chore that currently doesn't get done.
An agent may well decide things, but we recommend starting carefully. Let it make a proposal first, and decide yourself what happens to it.
The six parts of an agent
Every agent we build has the same six parts. Leave one out and you'll notice within a few weeks.
- Schedule. When it runs. Every Monday at 7:00, for example. A fixed time makes it predictable, for you and for your colleagues.
- Instructions. What it has to do, written in plain language: the job, the order, the rules, and what to do when something is missing. This is the most important part and the one you'll refine most often.
- Input. What it reads first. Always the company brain from part 2, plus whatever the job needs: customer questions, a plan, last week's numbers.
- Output. Where it leaves its work. A draft in a fixed folder, never straight onto your website or your social media.
- Approval. One fixed moment when a person reads the draft and decides: good, change it, or drop it.
- Log. What it did, what it read and what didn't work. One rule belongs here that you mustn't forget: "found nothing" and "couldn't look" are two different things. An agent that reports the second as the first is hiding a problem.
What we built at RentVisie
Our first agent wrote a blog article. Every Monday morning it takes the next topic from the plan, reads our brain first, writes the article in Dutch and translates it into English, German and Spanish. Everything lands as a draft in one overview. The owner reads it, edits or approves, and only after that approval does an article go online. This article was made the same way. More jobs run alongside it by now.
Gotcha sets up harness v1
Gotcha doesn't start with an article. That's too big for a first agent. He picks a chore that currently never happens: working out what he should be writing about.
His agent has one job: every Monday, propose three topics based on the questions customers asked last week. During the week the planner adds those questions to a shared document, one line per question, without customers' names or contact details. Then Gotcha writes the instructions, in plain language:
First read Gotcha's company brain. Then read the customer questions from the past week. Pick the three questions that come up most often, or that are most about money. For each question, propose one blog topic, with a title, three sentences on what it should cover, and the customer question it answers. Only use facts from the brain. If there are fewer than three questions, say so, and don't make any up. Write the proposal in the Proposals folder, with the date in the name. Publish nothing.
Then he sets up the rest.
- Schedule: every Monday at 7:00, as a scheduled task in his AI assistant.
- Input: the brain and the document with customer questions.
- Output: one file a week in the Proposals folder.
- Approval: Tuesday at 8:30, over the first coffee, fifteen minutes. He picks one topic, or none.
- Log: at the bottom of each proposal the agent notes what it read and what was missing.
The first Monday, one of the agent's proposals was: "What do you pay if you don't bring the van back full?" That question came in three times. Exactly the topic that fits Gotcha's promise: knowing up front what you'll pay.
What one agent costs and what it earns
The technology is the smallest cost. Scheduled tasks are now part of the mainstream AI assistants. With Claude they run in Anthropic's cloud, even when your laptop is closed, and they're available from the Pro plan up (Anthropic's documentation, checked on 16 November 2026; the feature is still a research preview there). Claude Pro costs $20 a month billed monthly, or $17 a month on an annual plan (price on Anthropic's site, checked on 16 November 2026). ChatGPT has scheduled tasks too; what each plan allows is in OpenAI's help centre (November 2026).
The biggest cost is you. At Gotcha that's fifteen minutes a week to review, about an hour a month. The first time takes more: an evening to write the instructions and to refine them after the first two runs.
What it earns is a worked example. Say that without the agent the owner would spend an hour a week reading customer emails and thinking up topics, if he gets round to it at all. With the agent it's fifteen minutes. That saves three quarters of an hour a week, 36 hours a year over 48 working weeks. At €60 an hour that's €2,160, against $240 a year for the subscription billed monthly. If, as in the series More revenue from the same fleet, we use €100 per rental day as a hypothetical example, those hours equal almost 22 rental days.
The real money is somewhere else. A topic that answers a real customer question is an article the next customer with the same question will find. That only pays off once it actually gets written. That's why the agent is the first step, not the last.
How to start this month
- Pick one chore that currently doesn't get done and that comes back every week. Not the most important one, the most boring one.
- Write the instructions down the way you'd give them to a new employee, including what to do when something is missing. Start with "first read the company brain".
- Set the job up as a scheduled task in your AI assistant on a fixed day and time.
- Choose your approval moment and put it in your calendar as a fixed appointment with yourself.
- Read the log after two weeks and refine the instructions wherever the agent started guessing.
What you don't leave to AI (yet): publishing
An agent may propose, write, translate and summarise. At the start, it doesn't publish. Not on your website, not on LinkedIn, not in a newsletter.
A draft sitting in a folder costs nothing if it's wrong. A mistake that's online costs you a customer, an argument at the counter or a promise you have to keep. The difference between the two is one person who reads before anything goes live. That's why approval is a fixed part of every agent, not something you do when you have time.
At a later stage, that may change. Once an agent has delivered drafts for months that you approve without changes, you can let it publish part of its work itself. But that is a step you take based on what you have seen, not the step you start with.
What you have now
After this part you have, like Gotcha, a harness that runs without you starting it: a company brain, one agent with written instructions and one fixed moment when you decide.
In part 4 the agent doesn't just propose but writes too: articles, images and video that don't sound like AI, built on your input. All parts are on the page of the series Making money with AI.
Don't want to set this up yourself, but do want the work done every week? That's what RentVisie's digital marketing for rental businesses is for.
Frequently asked questions
What is an AI agent for marketing?
An AI agent for marketing is a job that starts by itself, at a fixed time and with fixed instructions, and delivers a draft: an article, a list of topics, a summary of the week. Unlike a chatbot, it doesn't wait for you to ask. Start with an agent that publishes nothing; a person approves the draft.
What does an AI agent cost per month?
A paid subscription to an AI assistant is enough. Claude Pro, for example, costs $20 a month billed monthly, or $17 a month on an annual plan (checked on 16 November 2026). The biggest cost is your own time: for one small job about fifteen minutes a week to review, and the first time an evening to write the instructions.
How do I build my first AI agent?
Pick one recurring chore that currently doesn't get done, write the instructions down the way you'd give them to a new employee, and set them up as a scheduled task in your AI assistant. Have the agent read your company brain first, leave its work as a draft in a fixed folder and keep a log. Choose one fixed moment a week to review.
What's the difference between an AI agent and a chatbot?
A chatbot responds when you ask it something. An agent starts by itself, at a fixed time, with a fixed brief, and leaves the result ready. So the work gets done even in the weeks when you have no time for it.
Book a call. We'll show you what your first agent looks like for your rental business, or we run it for you every week through digital marketing.
Book a call