How We Produce Our Content: Hybrid Division of Labour in Live Operations

One morning in early October I sat with my coffee in front of an article that had grown far too long. The topic was the OpenAI incidents and what companies should learn from them for running their own AI agents. So much had already happened that people ought to know. We wanted to summarise it, abstract from it and bring in our own theses at the same time. That no longer fitted into one text.

I asked whether this was too long for a single article. The AI agreed. In conversation we then worked out which story to use as the hook, how to break the topic down and whether it should become a series with an accompanying podcast. Once we had talked it through long enough, the texts fitted almost straight away, and I only had to change small things. That became the first part of a series on why AI safety is an operations problem.

This is the hybrid division of labour behind our Insights: a conversation between human and machine, embedded in fixed handoffs, routines and a named person who signs off. In this piece I lay the process open, with tools, numbers, a worked example and the mistakes we made along the way.

5–6 h → 1 h

effort per article, before and today

internal operations, 2026

5

handoffs between human and machine per article

internal operations, 2026

1–2 × 2

articles per week, each in two languages

publication plan, 2026

Contents

• Why we disclose the process
• Hybrid division of labour: five handoffs
• Worked example: how one article was made
• The routines: which cadence triggers what
• The gate: checking and judging
• What went wrong and what we changed
• The feedback loop: why the next run knows more
• When the consultant becomes a developer
• What transfers
• Conclusion
• Frequently asked questions
• Sources

Why we disclose the process

We disclose the process because we think it is the right thing to do. We want to lead by example and practise what we preach: use AI to gain quality and efficiency, stay aware of the consequences and develop processes we can keep improving.

Then there is the obligation. The AI Act transparency obligations have applied since 2 August 2026, and we have to label certain texts accordingly. For AI-generated text on matters of public interest, an exemption applies where the text underwent human review and a named person holds editorial responsibility. This page is the record of that process. What disclosure requires in detail is covered in a dedicated article.

Above all, we see the obligation as an opportunity: it gives us a reason to show openly how we work, and with it the strategy we follow ourselves and adapt and shape for clients.

Hybrid division of labour: five handoffs

The difference between an AI tool and AI operations is not who does what. It is where the baton changes hands, and how you can tell. Every handoff has a trigger, an artefact and a recipient. Without the artefact, all that is left is the hope that the other side knows what to do.

For presentations:SVGPNG

The nine steps behind those five changes, in order:

#	Who	Step	Artefact
1	Human and AI	Collect topic ideas: from client conversations and observations in the team, plus the list of ideas the AI tops up every week with thesis, occasion and sources	idea on the board
2	Human	Choose the topic, write a brief with core message, structure, sources and angle	brief in the ticket
3	AI	Research data and facts, develop the concept: thesis, outline, planned figures, fact check against the original source	concept comment
4	Human	Fact and logic check of the concept, sharpen, approve	approval in the ticket
5	AI	Check figures again at the primary source, write DE and EN, build diagrams, pass the quality gate	pull request
6	AI	Attach a PDF preview to the ticket, read link and task at the top	PDF preview
7	Human	Check tone, structure and stance, edit, approve	approval in the ticket
8	AI	Set the label, merge, nightly release, go-live on the publication date, live check	live article
9	Human	Feed what surfaced in review back into standards and project memory	updated standard

Tools, named: Claude for the list of ideas and steps 3, 5, 6 and 8, Asana for briefs, approvals and documentation, GitHub for pull requests and the nightly release, Astro as the framework, Coolify as the target. We name them deliberately: swappability is only verifiable when you know what is inside.

Where your company stands on AI Operations can be mapped in minutes with the free Self-Check.

Worked example: how one article was made

Every process sounds good in the abstract. So here is a real run from August: the article on measuring AI ROI, back then titled “81% cannot measure their AI ROI”, written on 3 August 2026, published on 17 August 2026. Back then there was no concept step yet, and I still pushed the changes myself. The run shows what worked, and the mistake we learned the most from.

The brief. Created on 7 July 2026 as an Asana ticket. It contained the core message, the H2 structure, three mandatory metrics, primary and secondary keyword, source pointers, internal links and the intended angle: “measurability as a budget survival question, do not repeat the gap narrative”. Time spent: around twenty minutes.

The run. On 3 August 2026 the routine checked the 14-day window and found two articles due. With nothing due, nothing would have happened.

Verification, where it got interesting. The brief cited “~5,000 companies” as the sample of the DIHK digitalisation survey. Checking the original source gave 4,686 participants, surveyed between 10 and 28 November 2025, report titled “Artificial Intelligence, Sovereignty and Resilience”. That correction worked. The brief’s two headline figures, 81 percent unable to quantify AI value and more than 75 percent without measurable return, were also considered confirmed after the check. They were not: they came from a secondary source that mixed values from several studies and did not appear in that form in the DIHK survey.

That is why we now check figures at the source twice: in the concept and when writing. Briefs age. Checking facts against the brief instead of the source perpetuates old numbers.

Production. German and English versions, one SVG diagram with language-specific labels, astro check clean, one commit. No push.

The handoff back. Result comment in the ticket with the verified facts, the commit hash and the planned date. Ticket closed, review task created.

The review. Around one hour. I adjusted sharpening and sequence, plus the sample size, which had to read 4,686 instead of 5,000. I did not catch the wrong headline figures in the review. On 19 August we corrected the article, more on that in What went wrong.

Previously the same article would have cost five to six hours: research, writing, diagram, English version. The factor is five. It does not come from writing faster but from removing the work between decisions.

The routines: which cadence triggers what

A process without cadence is a statement of intent. Four routines hold this one together.

For presentations:SVGPNG

What the four routines do:

• Release, every night at 4:20 a.m. An automated run takes everything that has been approved and merged into the live version of the website, provided tests and checks pass. Early in the morning the website is rebuilt, and articles whose publication date has arrived appear.
• Pipeline, weekdays at 6:04 a.m. The AI works through the Asana board: it develops concepts for new briefs, writes articles whose concept has been approved, prepares approved texts for the release, checks whether scheduled articles are live and links to articles as soon as they appear. On Mondays it tops up the list of ideas.
• Publication, Mondays and Thursdays. Fixed publication days give readers and us a reliable rhythm. The date is stored in the article itself.
• Newsletter, Fridays. A script compiles a draft from the week’s new articles; a human reviews and sends it.

One detail matters more than it looks: approval and appearance are decoupled. That decoupling allows lead time and takes the time pressure out of the review.

The gate: checking and judging

In August I wrote here that factually, almost nothing is ever wrong. I would not put it that way today.

Mistakes happen. They happen when data or information is imprecise, when sources are believed too quickly or when the fact check does not take place. Most often they happen when there is not enough time to check things twice. We live in a tension between time-consuming meticulousness and economic and personal time pressure. We only find the middle ground if every mistake turns into an improvement in the process.

I have experienced both directions. The AI contradicted me when I had the wrong impression. And the other way round, I simply could not picture some figures it put in front of me. I asked, we researched further, and in the end it turned out that my intuition was right: the data had been misquoted or misinterpreted.

In our cases the AI does not invent data. It sometimes takes it out of context. We humans know that from ourselves too.

Sometimes it does not understand the context of a figure, sometimes it pulls the figure out to support the article’s argument. As with us humans, that has to be checked and corrected. For the figures that matter most to the argument, I therefore look at the source myself and check the data there.

The second thing I contribute is judgement. The machine cannot do it because it is subjective. No human can guide my pen either without me noticing that the text is not mine. I have to put my own hand to it so that it sounds the way I want to sound in that moment. Most often that is the introduction: that is where I make the words and the dramaturgy more human and add examples. I also change how topics and sections open, to make them easier to get into and to leave room for my visual language. I rearrange sentences and look for the tone that fits the moment. It is fine-tuning, often so situational that no prompt can generalise it. It is like with the colleagues I am lucky to mentor: I can explain a lot, derive it, give background and tell them in detail about examples from practice. Experiencing certain situations yourself and finding your own way of dealing with them is something else.

Legally that is clean. The exemption in Article 50 requires a deliberate substantive review with a real option to reject. With us that review happens twice for every article: as a fact and logic check of the concept and as a review of the finished text.

What went wrong and what we changed

An article about our own content pipeline that only shows successes would not be very credible. Two cases taught us the most.

The figure of 19 August. The two headline figures from the worked example went live as DIHK figures although they came from a secondary source. A LinkedIn post about it had to be deleted. It surfaced during our regular, recurring and automated tasks in which we research and check sources: the original source did not contain the figures in that form. On 19 August we switched the article to verified figures and added a correction notice. Since then we check every figure at the primary source, meaning the institute, the authority or the publisher of the study. Blogs, press portals and summaries do not count. There are two remedies for mistakes like this, and it takes both: human questioning with intuition, and fixed checking and control processes.

The feedback that got stuck. A colleague had given feedback on the article about shadow AI that was not worked in before it went live. The technology had worked, the process had not. Alongside the technology we therefore always have to question the process, especially when several people are involved. We changed three things:

1. Transparency. Do we know and understand the process? Can we trace it and explain it to others? That was the source of the error: the workflow was unclear and open to misunderstanding.
2. Starting over. We straightened the process out, piloted it, tested it and set it up again from scratch. With the new Asana board and its clear sections it is now much easier to see how the workflow runs, where an article stands and whose turn it is.
3. Tools. Briefs, article concepts and PDF previews that open easily and conveniently with one click in the ticket lower the error rate and with it the frustration.

The feedback loop: why the next run knows more

Step 9 is the least spectacular and the most important. What surfaces in review flows back into two places: the standards that govern all articles, and the project memory that the next run reads.

Three examples from recent weeks: the rule that every article needs its own diagram and that the metrics row does not replace it. The rule that every figure is checked at the primary source, because briefs age and secondary sources shift numbers. And the rule that every ticket starts with the read link and the task, so nobody has to search.

None of these rules existed at the start. Each came out of a correction. Without that loop the pipeline would be automation. With it, it is an operation that improves. We build the same difference for clients: we do not introduce a tool, we build processes into machines that learn and still remain under human control.

When the consultant becomes a developer

The most interesting effect appears in no flow chart.

As a consultant I used to create strategies, ideas, analyses and concepts, ninety percent of which ended up in documents. Implementation, whether organisational development, programming or action planning, needed many other people we depended on. Getting to the results we actually wanted took a lot of management effort, coordination and alignment. And afterwards, traction and effectiveness were lost every time.

Today I turn concepts into dummies and prototypes myself. That lets me clear up misunderstandings early and create shared pictures in people’s heads. I build first functional applications that let me test and demonstrate how something works. With them I check the requirements in reverse, so to speak, and document them for a later, professional implementation by experts. I can also pilot processes in advance, organisational and interpersonal ones alike: does an approach to collaboration or meetings help us, or does it lead us astray?

As managing director I now talk to my colleagues in the specialist teams much more on an equal footing. I get into topics more easily, prepare meetings better and no longer need someone from the team for every task, because I can handle much of it myself with the time and means I have. I steer this process myself too: I write briefs, review concepts and previews and decide on approvals.

What you must never forget is to check what comes out. The results are provisional until I have discussed them with an expert. As with the articles, misinterpretations and misunderstandings can always happen, and checking is how we prevent them. It is a question of mindset: finished quickly is far from ticked off, even when a complex task was done in no time.

When I show clients this way of working, most are enthusiastic and amazed. Many cannot imagine it in their own structures and environments, though. In the end the same topic always remains: control and security, over data and systems as much as over processes and results. Am I still in charge of what happens, even if I do not understand it? Can I influence it? Which security risks am I exposing myself to?

Technology has outpaced people’s capacity to grasp it, and their imagination along with it. We now have to get back on the same level so that these technologies are integrated sensibly. That is the most concrete thing to talk through in a diagnosis call.

What transfers

All of it transfers, the pipeline as much as the pattern behind it.

Hybrid division of labour means tasks go to teams of people and AI agents who solve them together. The human steers, the AI contributes. It is a creative assembly line that gives people superpowers, in expertise, quality and quantity: they get more work done at higher quality while always following the company’s rules. That changes processes, and it changes org charts.

For presentations:SVGPNG

Four questions turn any workflow into a hybrid division of labour:

1. Where are the handoff points? Every change of hands needs an artefact you can show.
2. Which cadence triggers what? Without cadence, nothing happens or everything happens at once.
3. Who decides, and how can you tell? An approval in the ticket, a label, a comment, not a feeling.
4. Where does the learning go? Without a feedback loop it stays automation.

Answer those four for a single workflow and you have the entry point. Why the workflow comes before the tool is covered in workflow-first.

Conclusion

That we raised our productivity by a factor of five with this process, while involving more colleagues and publishing more articles, is something we had not achieved in more than 15 years of company history. It was only possible because of the established process, a division of labour we designed as hybrid from the start: named handoffs, a fixed cadence, transparency, two human gates for checking and judgement, and a loop that returns what was learned. The mistakes of the past weeks showed me that this only holds if we question the process as much as the technology.

If managing directors take one thing away from this article, it should be this:

One thing I would never automate: conversations with people, whether colleagues or clients.

That is also my answer to the transparency question. A company that can build its process so it may be shown does not have a compliance problem. It has an argument. This page is both: the record and the evidence.

Frequently asked questions

Does the AI write our Insights on its own?

No. Both supply ideas. Topic selection, the brief, the fact and logic check of the concept and final approval rest with the human. Data and fact research, concept, drafting, diagrams and publishing sit with the AI. Both check figures: the AI at the primary source, and I also check the figures that carry the argument. Without human approval no article appears. The notice block under every piece shows the split for that specific article.

How much time does it actually save?

An article in two languages with a diagram used to cost five to six hours. Today around sixty minutes remain for the brief, the concept review and the text review. The saving does not come from writing faster but from removing the work between decisions.

What happens when the AI gets something wrong?

Mistakes happen, mostly because a figure was taken out of context or quoted imprecisely. That is why we check figures twice at the primary source, in the concept and when writing, and I look at the most important ones myself. One example: the brief cited a sample of roughly 5,000 companies; the verified figure was 4,686. The correction happened before publication. If something still slips through, like the two figures from a secondary source on 19 August, we correct it openly and tighten the rule.

Why does the machine not give the approval?

Because the approval is the decision. It marks the point where a human takes responsibility, in the process and in the sense of Article 50 of the AI Act. Automatic approval would remove the step that makes the process defensible.

Does this transfer to other areas?

Yes, if you answer four questions: where are the handoff points, which cadence triggers what, who decides and how you can tell, and where the learning goes. We work that through for one concrete workflow in a free diagnosis call.

Sources

• European Commission: Quick Facts, Transparency rules for AI systems, 2026
• DIHK: Digitalisation Survey 2026, Artificial Intelligence, Sovereignty and Resilience, 2026
• Astro: Content Collections documentation, 2026
• Coolify: documentation, 2026

All time figures and process details in this article come from our own operations, as of October 2026. The worked example shows the workflow as it ran in August.