Content Operations Automation: What We Automated On Our Own Desk

Content operations automation pays off at the queue, not at the calendar. The version that works is a board where every piece of content for every brand sits as one card, machine checks run before a card is allowed in, and a person still decides what ships. We run one across our own brands as of September 2026.

This post covers what we automated on it, the duplicate publish that changed its design, and the two features on it that are still empty.

It starts where the last one stopped. Our planning agent drafts marketing tasks from what is actually selling, then hands them over. That post ended at the moment a draft card reaches the board.

Marketing got the AI first and the plumbing last

The US Census Bureau’s Center for Economic Studies published working paper CES-WP-26-25 in April 2026, drawing on the Business Trends and Outlook Survey for a November 2025 to January 2026 reference period. Sales and marketing is the most common place firms put AI, at 52% of AI-using firms, ahead of strategy and business development at 45% and IT at 41%. The same paper reports that 57% of AI-using firms have AI in three or fewer business functions.

So adoption is both real and narrow. The drafting got automated. The queue the drafts land in usually did not.

Content Marketing Institute and MarketingProfs published their B2B trends research on 8 October 2025, from 1,015 B2B marketers surveyed between June and August 2025. Asked where 2026 investment goes, respondents put AI-powered tools first at 45% and people, meaning salaries and training, last of ten options at 9%. In the same study 95% of organisations use AI applications and 12% rate themselves highly effective with them.

The CMO Survey’s 35th edition, published 31 March 2026 from 308 marketing leaders, reads the same way: training is down to 3.8% of marketing spend against a pre-pandemic high of 5.8%, and no marketing capability scored above 5 on its 7-point scale.

More output capacity, the same number of people, no better process. That gap shows up in the queue first.

One board per brand, three columns, and somewhere to put dead work

Each piece of content is one card: brand, title, status, scheduled date, format, target platforms, notes, and an optional assignee. The board runs three columns, planning, in progress and done, plus a fourth terminal state for work that was cancelled or replaced.

That fourth state is the boring part that keeps the rest usable. Killed campaigns and duplicate ideas have somewhere to go that is not the done column and not the backlog.

A kanban view and a calendar view sit over the same cards, filtered by brand, so each brand reads as its own lane instead of one merged pile. Same records, two ways to look at them. Nobody maintains a second calendar.

Two ways a card arrives, and one of them carries a flag

Work enters the queue two ways. A person adds a card, or the automated planning step inserts one that is already marked as AI-drafted. Both land in the first column.

The automated path is a tomte, our word for one production agent with one defined job. Its job ends when the card is in the queue. It does not assign the work, it does not move the card forward, and it does not touch cards it did not create.

The honest weakness: the AI-drafted flag is stored on the card, not shown at a glance in the column. You open the card to see which is which. That is a gap we know about rather than a design decision.

The check that runs before a draft reaches the queue

There is a dry-run mode that runs the content quality guardrails against a batch of candidate cards and inserts nothing. It answers one question, would these pass, before anything reaches the board.

A duplicate check compares each new draft against what is already published. Exact repeats are blocked outright. Softer topical overlap is flagged, not blocked, and a person decides. Machines match strings well and judge whether two angles on one subject are really the same piece badly, so the split follows the machine’s actual competence.

That review layer is now the bottleneck across the industry, not the drafting. Jasper, with the research firm Benchmarkit, published The State of AI in Marketing 2026 on 28 January 2026 from 1,400 marketing professionals: blockers from legal, compliance and brand review rose 3.4 times year on year and now outrank budget as the top barrier to scaling AI. Fielding dates are not disclosed on either release page, so read it as vendor research with a third-party partner. The direction matches what we see. Production got cheap. Review did not.

The duplicate publish that changed the design

Two automated runs each decided the same card was theirs, and one piece went out twice.

The fix was to make every status change compare and swap. A caller has to state the status it expects the card to still be in, and the update fails if the card has moved since. Two runs can race for the same card and exactly one wins. The loser gets an error instead of a second publish.

This is the same principle as the approval gates we install in production: the agent does not get trusted to be careful, the data layer refuses the bad write.

When an automated run dies holding a card

A background sweep reclaims cards that have been stuck in progress past a time threshold and puts them back in play. A crashed run cannot park a piece of content forever, and nobody has to notice manually that a card went quiet.

Every automation that claims work needs a reclaim path, and it is the part that usually gets left out of the first build. Without it the failure is silent: the card looks like it is being worked on by someone.

Where a board beats the tools most teams already have

Capability	Shared calendar or sheet	General project tool	Content desk board
An agent can create work directly	No	Only with custom wiring	Yes, flagged as drafted
Duplicate check before work enters	No	No	Yes, block or flag
Recovery when an automated run crashes	No	No	Yes, timed reclaim
One lane per brand over one record set	Separate files per brand	Separate projects per brand	One filter
Publishes to the site	No	No	Yes

Most teams are not missing a tool. They are missing the place where an agent is allowed to put work down and a person is guaranteed to pick it up.

Approvals and analytics: the two we have not built

The topic this post was scheduled against promised queue, approvals and performance in one place. Two of those three are real as of September 2026.

There is no approval workflow. The navigation entry exists and is disabled. Roles are modelled on user profiles, and the enforcement behind them is not wired up yet. Marking a card done is a human action in the interface, which is the crude version of an approval step: nothing publishes itself.

There is no per-post analytics view either. What exists is a scheduled health check that counts cards created, completed and overdue over a rolling window and raises a flag when the desk looks stalled, plus a content mix breakdown by platform and by brand in the list view. That is operational health, not performance.

Writing this post as though all three shipped would have taken ten minutes and cost more than it saved. The board is good at intake and bad at review, and that is the accurate order to fix things in anyway: a review step over a queue nobody trusts is theatre.

What a build like this runs

The publishing half of this is the content engine, $4,900 to set up over two to three weeks, with an optional $990 a month if you want us operating it after handover. It runs on your accounts and your keys, and you keep it if you cancel.

The same pipeline publishes daily into our own 363-product B2B catalogue, which began producing quality inbound leads in its first month, as of July 2026.

Want the same on your brands? Start async. Send what you publish now, how many brands you run it for, and where it has to end up. You get a written reply within one business day, and no meeting.