An editorial desk as a pipeline
MCC AI Weekly goes out every Thursday, in English, to people who decide about AI. Gathering sources, judging them, writing and sending run automatically in n8n. Two decisions stay in human hands — deliberately.
What it is
The newsletter is written for people who decide how AI gets used. Not “model X is now available”, but what a development means for an organisation, its contracts and its workforce planning. Every Insight ends with a “So what” — a recommendation that addresses a decision.
| Section | What goes in it | Origin |
|---|---|---|
| Insights | Studies, data, argued theses — findings that will still hold next month. | proposed automatically |
| Weekly Short News | Short reports from the week in AI, summarized at a glance. | proposed automatically |
| AI Player of the Week | One company or tool in brief — with a verdict on who it suits and who should skip it. | chosen by the editor |
| Tip of the Week | Something readers can put to use themselves. | chosen by the editor |
The week
The rhythm is the architecture. Steps marked in blue run without anyone touching them; the two in ochre are done by hand.
Running continuously
From source to issue
Several hundred finds a day become one issue with roughly a dozen items. The funnel is deliberately narrow.
Five specialist feeds are polled; some two dozen industry newsletters arrive in a dedicated mailbox. A filter rule separates source material from reader replies — only the source folder is read. Once a week, a dedicated workflow additionally searches selected company sites for new studies.
RSS · IMAP · web searchA time window, keywords and a cap per source cut the raw volume before any compute is spent. Around 230 feed entries leave single-digit numbers of candidates.
n8nA language model assesses every candidate against a written editorial charter: does this carry a finding that would make an executive decide differently? It returns a headline, a summary, a section and a score. Only what clears the threshold is kept.
Claude APIAccepted items land in a spreadsheet — one sheet per section. It is both the working stock and the control surface: an item can be earmarked for a later week instead of being spent too early.
Google SheetsOn Monday a form shows the strongest candidates per section, together with older items from the pool. Whatever is not picked stays — a thin week can fall back on the stock.
n8n formEach item becomes a finished section including its “So what”; a second pass writes the frame around the issue: title, preview line, introduction. Two editorial illustrations are also generated — a cover image for the issue and one for the week's most important story. The result is placed into a table-based e-mail template that holds up in older mail clients too.
Claude API · Image generation · HTML templateA test mail goes to the editor and the campaign is created as a draft — with no send time, so nothing can go out unintentionally. Only the release in the mailing system triggers the dispatch. Afterwards the issue is archived.
BrevoAn AI model generates exactly two illustrations per issue — a cover image for the whole issue and one for the week's most important story. Both follow a fixed visual language: calm, museum-like object compositions rather than robots, dashboards or generic AI aesthetics. If a generation fails, the issue still goes out — with the fixed cover image or with no image at all, never delayed.
The machinery
Every workflow has one trigger and one job. No workflow reaches into another's remit — that is what makes faults findable.
A tool for assembling processes between services: a trigger starts a chain of steps — a time of day, an incoming mail, a form. Each step reads, transforms or hands on data. What would otherwise have to be written and operated as a small program is here a visible, logged process.
n8n keeps the beat. It polls the sources, triggers the AI assessment, maintains the editorial pool, serves the selection form, builds the finished issue and records the dispatch afterwards. Every run is logged — if a step fails, it is clear which one.
The instance runs on our own server in Germany. Source texts, editorial data and the credentials for connected services therefore stay under our control and do not leave the environment.
| Workflow | Trigger | Job | Kind |
|---|---|---|---|
| Sign-up | webhook | Receives the form from the website, screens for bots and submits the address for confirmation. | automatic |
| Content ingest | daily 07:00 + mailbox |
Polls feeds, reads new source mails, pre-filters, has them assessed and writes to the pool. | automatic |
| Enrich suggestions | daily at 06:00 | Fills in title and summary for links added by hand — dropping in a URL is enough. | automatic |
| Web-search discovery | Monday 06:00 | Searches weekly for new studies from selected research houses and feeds suitable finds — flagged as researched — into the pool. | automatic |
| Monday invitation | Monday 08:00 | Reports the pool and links the selection form. Checks that link itself before sending. | automatic |
| Curation form | form request | Shows candidates per section and writes the selection back, together with the target week. | operated |
| Production | started by hand | Writes the copy and two editorial illustrations, builds the HTML, sends a test mail, creates the campaign draft, archives the issue. | started |
| Record the dispatch | daily at 12:00 | Recognises from the campaign's send date that the issue has gone out and marks the items as sent. | automatic |
Production checks every source URL before any compute is spent and stops if a link is not fit to send. The campaign is created without a send time, so a run cannot trigger anything. The Monday invitation checks its own form link and writes a warning instead of mailing a dead one. The dispatch record touches only the status column and writes only when something actually changes.
Data protection
The sending could have been built in-house. That would have been the single largest piece of work in the project — and the most error-prone, because legal obligations hang on it. Brevo brings them along.
A platform for e-mail marketing, formerly known as Sendinblue. It manages subscriber lists, sends campaigns to many addresses at once and handles everything attached to that: confirmation mails, unsubscriptions, undeliverable addresses, delivery and open statistics.
For all contact with subscribers. The sign-up on the website registers the address there, Brevo sends the confirmation mail and only adds the address after the click. The dispatch of the finished issue runs through it as well, including the unsubscribe link in every mail. Subscriber addresses live there and nowhere else — not in the editorial environment.
A confirmation procedure, a legally sound unsubscribe link, handling of bounces and proof of consent would together have been the largest single effort in the project — and the part where mistakes have legal consequences. The plan in use is free of charge and covers the need; there are no running fixed costs.
The operator is Brevo GmbH, based in Berlin, and processing takes place in the EU. That removes the discussion about transfers to third countries, which requires additional safeguards with US providers.
The agreement forms part of the terms of use and takes effect when they are accepted — no separate signature required, but documented and in place.
After sign-up a confirmation mail goes out; only whoever clicks the link in it is added. The confirmation is stored on the contact and is therefore the consent under Art. 6(1)(a) GDPR — evidenced, not asserted.
The unsubscribe link is a mandatory part of every mail and is inserted at dispatch. One click is enough; it takes effect immediately and without questions.
We collect the e-mail address and nothing else. No name, no company, no profiling. Bounces and unsubscriptions are handled by the mailing service itself.
Subscriber data sits exclusively with the mailing service. The editorial data — sources, assessments, drafts — is kept separately and holds no personal data about subscribers.
Reach measurement loads only after active consent through the cookie banner — before that, nothing is loaded and nothing is set. The sign-up form sends the address to our own endpoint, which checks it and passes it to the mailing service; a hidden field catches automated entries.
The limits of automation
Both remaining steps could be automated. They stay because these are the two points at which a person sees the issue before it goes out.
A model can judge whether an item fits the charter. Whether an issue works as a whole — whether it tells the same story three times, whether the mix is right — is an editorial decision.
The run costs compute and produces a campaign ready to send. It therefore begins with a click — and ends with a draft, not a dispatch. The release is a second, separate step.
Every step that does something outward — a mail, a campaign, an entry in the spreadsheet — reports its result back. If a part fails, the run stops with a message naming the place and the correction needed, rather than quietly leaving half a result behind.