What marketing job postings say about AI skills
Research shows employers want marketers who can turn processes into working AI automations.
By Margaret Lee, CMO, Devart and TMetric
Published on September 28, 2026
Share this article
Your next marketing job may ask you to do more with AI than write better prompts. Employers are increasingly looking for people who can build AI agents, automate workflows, and create systems that actually do the work.
Within a year, AI shifted from a topic marketers discussed to one that employers now include in job requirements. Indeed’s Hiring Lab found that the proportion of marketing job advertisements referring to AI rose from 8.4% to 14.9% in 2025. PwC’s 2026 AI Jobs Barometer, based on over a billion job ads, reports a 62% wage premium for AI skills.
We looked at 60 marketing job descriptions that are open, published on employers’ own careers pages, and confirmed as live. We chose them because they mention AI, so they represent the cutting edge rather than the entire market.
In 45 of the 60 cases, designing or shipping AI agents and automated workflows is a listed requirement. Only two mark it as a nice-to-have. One job posting states plainly: “AI is the way in which this role will be carried out. We are looking for a builder’s mindset: one that involves workflows and agents, not just prompts.”
It’s a ladder, and most people go straight to the top.
A typical error is starting at the wrong point — taking an agent-building course before you can describe on paper how one of your own processes works.
An agent only carries out the actions specified by the process you give it. If you give it a process that you’ve never written down, the result can appear correct, polished, aligned with the desired image, and confident, yet still be wrong in ways no one notices for a quarter.
That’s why it’s a ladder rather than a menu. You must go through the steps in order: first understand the work, then describe it, and finally automate it. Each step depends on the one before it. Here are all five.
| Step | Description | Outcome | | --- | --- | --- | | 1. Understand how it works | How a model produces an answer, and what an agent, a skill, and MCP actually are | You can explain it to your grandmother | | 2. Describe the process | Processes and standards written as text a machine can use | Someone else can run your process from the page | | 3. Automate what repeats | Prompts become skills, skills become agents | It runs without you remembering to start it | | 4. Build and publish | One app taken from idea to a live URL, by you | Something you made is public, and other people use it | | 5. Prove it still works | Targets and evals | You’d notice next month if it broke |
The tech stack you’ll actually touch
Instead of asking whether you’ve used AI, employers now specify particular products, and 36 out of the 60 mention at least one.
A model you use every day. Claude appears in 26 of the 60, compared with 19 mentioning OpenAI or ChatGPT. That may surprise anyone who assumed the market centered on ChatGPT.
One orchestration platform, mastered properly — n8n takes 15 hours, Zapier 12, and Make 8. The basic concept is the same in all of them: a trigger, various steps, conditions, and an output. Once you’ve become proficient with one, you can easily apply that knowledge to the others within an afternoon. Choose the one your company already pays for.
A method of accessing your own data. The Model Context Protocol is present in eight out of the 60, giving an agent access to your tools and records rather than forcing it to guess based on what it learned during training.
A coding assistant, a repository, and a place where the software can be deployed. One of each:
- Build with: Claude Code, Cursor, or Replit.
- Keep it in: GitHub.
- Upload it online using Vercel, Railway, Render, Supabase, or Firebase.
The postings refer to the same set of items, requesting “AI coding tools (Claude Code, Cursor, Replit) for the purpose of building custom solutions” and also mentioning “the use of GitHub.”
Consider this list perishable, since about half of the items will change within two years. That’s why you advance up the ladder instead of gathering the tools.
Step 1: Understand how it actually works
Not for the sake of the vocabulary. The mechanics are enough to stop it from surprising you.
The process of forming an answer
The model doesn’t retrieve a pre-stored answer and just give it back. Instead, it creates the reply bit by bit, each time predicting the most probable next bit of text based on all the information currently available.
The way it gets to your own materials
On its own, the model has only the information it picked up during training, which is why it can describe a feature you don’t sell. To base its output on your documents, each passage is transformed into an embedding — a long list of numbers that represent what the passage means, with the numbers arranged so that passages on similar topics are close to one another. These embeddings are stored in a vector database.
Three things AI gets wrong about your expectations
There are three things to understand about how AI works.
- It doesn’t retain memory from one run to the next unless you give it some. That’s why a long, meandering brief can overshadow the instruction you actually care about.
- It will always generate some output. Producing plausible text is the entire mechanism, not a defect.
- The same input can produce different outputs. The model selects among likely continuations rather than adhering to a fixed rule.
The components surrounding the model
An agent isn’t a chat window. It’s configured software that performs a single task repeatedly on a schedule, using a specific set of tools and writing its output to a particular location. A skill is the written set of instructions you give the agent so it performs the job your way.
You’ve completed this stage when you can explain everything to your grandmother in her own language, without needing a diagram.
Step 2: Describe the process before you automate it
Breaking down a process means taking one you originally handled on instinct and breaking it down into individual steps, each with its own inputs and decisions. It’s also the least expensive form of audit.
When you write down a process, things no one can justify, approvals that exist because someone once requested them, and handoffs that quietly result in a day being lost all become apparent on a single page. You can’t automate a process that you’ve never described, and you can’t give instructions to an agent for one that hasn’t been described.
Five elements to address
| Element | Description | | --- | --- | | Voice | Real passages you’d publish, each with a note on why it works. Examples teach a model far more than adjectives do. | | Product truth | A one-page product summary with a named owner and a review date. | | Claims | A list of what may be asserted about the product, split three ways. | | SEO and GEO rules | The shape of a page and how it earns a ranking and a citation. | | Design system | Components with their states, tokens, and rules of use, kept somewhere live. |
Step 3: Automate what repeats
The whole point of this ladder is to spot the tasks you do over and over and gradually move each one up to the next level. You start with a good prompt, refine it into a reusable skill, and eventually that skill becomes an agent that can run on its own.
Step 4: Build one app end-to-end and put it live
This is the step that marketers resist most strongly and the one that changes the most. Choose a small, real task that your team currently carries out by hand. Create it yourself using a coding assistant. Put it in a repository and deploy it so it has a URL others can access.
Step 5: Prove it still works next month
Record the outcomes that the automation is expected to achieve from a business point of view and, at 30, 60, and 90 days, check whether they’ve been met, are on track, or have been missed.
Don’t wait for the course
Don’t wait for the AI for marketing course to start. Learn through practice. Try solving your marketing tasks with AI on your own and learn by doing. Learn each day, slowly. It’s better to spend a small amount of time regularly than a large amount later. If you devote one hour each day to a real task, that time builds up over time. If you complete a course but never put it into practice, it doesn’t.
Select a process that you carry out every week, put it in writing, measure it honestly, automate one aspect of it, and store the outcome in a place where other people can access it. Repeat this procedure next month. A single completed workflow teaches you more than any course.