This website, the blog, the competitive research, the LinkedIn posts: most of the recurring marketing work at our company is done by AI agents. Not as a pilot, in production. What started as an experiment twelve months ago is now the standard operating mode for our outward presence. And in customer conversations, that practice turns out to be our strongest argument.
We are not making this public to show off. The question comes up in every other discovery call: “How do you actually use AI yourselves?”. Honestly, you would not see it from the outside, because AI agents work quietly. So here is a concrete tour of the stack, the workflows that genuinely run, and the traps we walked into before the system became reliable.
What “AI-native marketing” actually means for us
AI-native is an abused term. For us it has a specific meaning: a human decides what should be done. An agent decides how. A human reviews the result and approves. And a human handles every irreversible action, such as publishing or moving money.
This is not a fully autonomous workflow, and that is by design. Each pipeline contains deliberate approval points. The point is not to replace people, but to delegate the routine hours that a 10-person company simply cannot afford to spend manually.
In practice we use Claude as the underlying model, Cowork as a desktop agent for our marketing workflows, Claude Code for development, and the Model Context Protocol (MCP) for our integrations into GitLab, Mattermost and Search Console. On top of that sits our own Skills system, which encodes the domain knowledge: design standards, writing style, taxonomy, banned phrases. We deliberately keep the stack small, because every additional tool adds friction.
Workflow 1: blog articles from idea to cover image
Every Monday a scheduled task fires automatically. The pipeline picks the highest-priority topic from a backlog, writes the article in German and English following fixed skill instructions, runs an automated SEO check (title length, description length, heading structure, word count, internal links, banned marketing phrases), generates a cover image in a consistent visual style via the image API, posts a status update to internal Mattermost, and notifies the sales team via an inbox router.
What the agent does NOT do: it never pushes to Git. The final commit is owned by a human. That keeps the publishing step under control while the time-consuming preparation is already done by the time someone opens their laptop in the morning. We described comparable setups in our piece on personal AI automation in practice.
Workflow 2: competitive research as a scheduled task
A second agent runs monthly and scans a fixed list of competitor websites. Each run captures new blog posts, reference cases, service-page changes, SEO visibility (Domain Rating and new backlinks from Ahrefs) and SERP movements for the keywords that matter to us.
The output is a score-based report in a Mattermost channel with concrete recommendations for our own marketing. From that report, topic ideas flow straight into the blog backlog. Work that previously cost two days of manual research now takes about an hour and stays comparable across months, because the format is identical.
For DACH readers a quick note: by “Mittelstand” we mean small to mid-sized companies, typically 10 to 250 people, that drive a large share of the German economy and are our primary audience.
Workflow 3: SEO and performance audits
A weekly audit agent walks the live site. It checks Lighthouse scores, Core Web Vitals, accessibility against WCAG 2.1 AA, security headers, broken links and indexation status in Google Search Console. The report also lands in Mattermost, with regression findings clearly flagged and sorted by severity.
The interesting part: it has saved us several times from shipping performance regressions we would have missed manually. A missing image compression, an accidentally removed loading="lazy" attribute, a header that quietly fell back to a third-party CDN: these are exactly the things a human misses after hours of coding. An agent that shows up every week with the same checklist does not miss them.
The blueprint: MCP as the connecting layer
What technically connects the three workflows is the Model Context Protocol. Through MCP servers, every agent gets controlled access to external systems: GitLab for code and issues, Mattermost for communication, Search Console for indexing data, Ahrefs for SEO metrics. Instead of building a custom plugin per integration, all tools speak the same protocol.
We covered the technical background in more depth in our piece on the Model Context Protocol. For marketing practice the headline is enough: MCP is the reason we no longer need a custom connector for every new workflow. That dramatically lowers the barrier to handing yet another routine job to an agent.
What works, and what does not
Three areas run reliably: structured, recurring tasks with clear acceptance criteria (blog, audits, reports). SEO validation against fixed checklists. Image generation in a consistent style, once the prompts are dialed in.
Creative topic discovery works only mediocrely. Agents tend to repeat similar themes or drift into buzzwords. Sharp skill definitions with explicit ban lists help, where we hard-block the typical marketing phrases. External web research is also still a weak spot: cache effects and hallucinations are real problems, and every claim must be verified before it goes online.
What does not work is real strategic decision-making. Where does the company position itself? What story do we tell? Who is the audience? Those questions a human has to answer. An agent can only operate within the frame the human sets. If the frame is fuzzy, the output is fuzzy.
Twelve months of lessons learned
If you want to start with AI-native marketing, three things are worth keeping in mind.
Skills matter more than models. Which LLM runs in the background is less important than the quality of the instructions the agent receives. We spent weeks refining our internal Skills: what are our design standards, which phrases do we reject, what does our writing style look like? That work pays off tenfold because it is reusable across models.
Build approval points in. We let the agent write the article, generate the image, run the SEO check, but a human flips the live switch. That is the insurance against a bad model day and against hallucinations slipping through. We discuss a similar tool-selection question in our piece on AI tools for content creation.
Keep tooling minimal. More tools means more bugs, more license cost, more onboarding overhead. We get by with a handful of building blocks because we use the same ones often. Every new tool has to justify why it stays in the stack.
The biggest gain is not the time saved. It is the consistency. Our blog articles now hit a constant quality level that a rotating freelancer would never reach. Our competitive data is comparable across months. Our SEO checks run weekly, not “when someone has time”.
What you can take away for your own company
If you are considering AI agents in your own marketing or sales, do not start with tooling. Start with the question: which recurring process costs us the most time without requiring creative decisions? That is your first candidate.
The tools are secondary compared to the process question. We help mid-sized companies in the DACH region run that analysis structurally and set up a first pilot project that genuinely goes into production, rather than ending up in the PoC graveyard where most AI initiatives from 2024 and 2025 quietly disappeared.
You can read more about our approach on our AI consulting and AI strategy page. Or contact us directly, the first thirty minutes are free.
