How I run a marketing team with AI agents

I was hired at Codility to rebuild marketing around AI agents. It is how I pitched the job when I was interviewing, and it is how my team works now. This is the setup, the rules that keep it honest, and what I would do in your first 30 days.

The brief was bigger than the team

The job covers positioning, launches, content, the website, competitive work, sales enablement, customer email and the story our leadership tells. Hiring a specialist for each of those jobs was never my plan. The plan from day 1 was a small team of marketers who work with AI agents, with the agents doing the production work.

One workspace, with a full history

Everything the team makes lives in one folder: briefs, positioning, case studies, emails, landing pages, analysis, and the rules we write by. The folder is a git repository, the tool developers use to track code, so every change has an author, a date and a reason. Any agent can read the whole of it.

We work in Claude Code and the OpenAI Codex CLI. Both run in that folder, and both connect to the systems marketing depends on: Salesforce, Tableau, Pendo, GA4 and Slack. That means an agent can check a claim against the CRM before the claim goes anywhere.

Write a job down once, then run it every time

The biggest change is that repeat work becomes a written skill. A skill is a file that tells an agent how a job is done, what good looks like and which checks to run. I have written more than 15. A few of them:

• A landing page skill builds a page from a brief, checks accessibility, search basics and speed, and gives me a preview to approve. A new page now goes live in 2 minutes, down from 2 hours.
• A launch video skill writes the script, builds the video in code with Remotion, adds an AI voiceover, and checks the result for flicker and loudness before I watch it.
• A monthly product update skill turns release notes into the customer email, the release page, a heads-up for sales and LinkedIn posts.
• A competitive skill refreshes what we know about each competitor every week.

A skill improves every time I correct it. The correction goes back into the file, so the next run starts from a better place. After a few rounds, the skill holds most of what I would have said in review.

Checks before anything ships

Agents write fast and sound sure of themselves. So nothing leaves the workspace without 4 checks.

1. Facts against the source. Every number traces back to a file or a system. If an agent cannot find the source, the number does not go out.
2. Buyer language. We built a database of how buyers actually talk, from 12 months of sales calls, customer calls and customer emails. Copy is checked against it: would a buyer say this, and would they believe it?
3. Plain language. Every line is read as if to a colleague who was not in the room. If they would stop and ask what it means, it gets rewritten.
4. A person signs off. Emails end as drafts that I send myself. Pages go live only after I approve the preview.

The day a senior leader called it AI slop

In October 2026, a senior leader read one of our trackers and called it AI slop. It was a fair call: the tracker was full of shorthand that made sense to me and to the agents, and to nobody else.

That same day I made plain language the first rule for everything the team writes, and I turned it into a check that every agent runs. The lesson stays with me: AI makes words cheap to produce, so the cost moves to the reader. The reader has to become the test.

What other teams get from it

Working this way, marketing also builds tools for the rest of the company:

• An account-health score that customer success uses to decide which renewals come first. It has flagged customers who wrote that they would not renew before any notice was logged.
• An account-level revenue model that leadership uses for pricing and packaging decisions.
• A review every 2 weeks that matches each product release to lost deals, so account executives know which deals to reopen.
• A competitive review that shows where we win and lose against each competitor, checked against win and loss records.

None of these was in a marketing job description. They exist because building them became cheap enough to try.

Where people stay in charge

Agents do the production work. People decide what the work is for: which story to tell, what to cut, which customer to call and what something should cost.

A few rules are fixed. Agents never send an email, publish a page or contact a customer on their own. Before an agent runs a large batch of work, it says how many items it will touch and what that will cost, and it runs a small pilot first.

How to start in 30 days

1. Put everything in one place with a full history: briefs, copy, rules and analysis.
2. Write your rules down where agents read them: your voice, the words you never use, and what you can and cannot claim.
3. Pick the 3 jobs you repeat most and write a skill for each.
4. Connect the systems your claims come from: the CRM, analytics and product data.
5. Add checks before anything ships: the source, the buyer’s language, plain language and a person’s sign-off.
6. Every time you correct the output, correct the skill too.
7. Build one tool for another team and watch what they use.

If you are building something similar, I would like to compare notes. Write to me at chris@grecochris.com.