AI Made Content Cheap. Human Judgment Just Became More Valuable.

A marketing team can now ask an AI agent to research a topic, pitch ten article ideas, draft the strongest one, write its title and meta description, cut three LinkedIn posts from it and queue everything for next week. The agent finishes before anyone has finished their first coffee.

A few years ago, that was a fortnight of work for two people. How much a small team could publish depended on how much it could produce.

Now someone opens the queue at 9:15 and has to decide which of those ten ideas the company should actually publish. The AI agent will offer an opinion, often a sensible one. Deciding what your company puts its name to is still a separate job, and it has become the constraint on content work.

AI made content production cheap

Research that took an afternoon takes minutes. A first draft takes a prompt. Rewriting for another audience, summarising a customer call, translating a product page, writing alt text, and turning one article into a newsletter section and a LinkedIn post are all fast and cheap now. So is the operational work around content, such as tagging, filling in missing SEO fields and finding pages that still quote last year's prices.

Adoption reflects that. In the Content Marketing Institute's 2026 B2B research, 95% of marketers said their organisations use AI-powered applications, and 89% of those use AI tools to create or optimise written content. Graphite, an SEO agency that estimates this with AI detectors, found that about half of new English-language articles in its sample have been primarily AI-generated since early 2025.

I work this way too. My articles start as a brief in ChatGPT, get reworked with Claude, and arrive in our CMS as drafts written by an agent. The time saved is real.

Why more content stopped being an advantage

Every company in your market has the same tools. If they can all generate a hundred articles on "7 tips for choosing a CRM", the hundred-and-first gives the reader nothing new, because they have already seen the other hundred.

The CMI data shows the gap between output and results. Among marketers using AI for content, 87% said productivity improved. Only 39% said content performance improved, and 12% said the quality of their content declined.

Search works on the same logic, although Google's position is often misreported.

Does Google penalise AI-generated content?

No. Google says appropriate use of AI or automation isn't against its guidelines, and its guidance on generative AI describes it as useful for researching topics and structuring original content. Its spam policies target scaled content abuse: large numbers of pages made mainly to manipulate rankings, with little value for readers.

How the pages were made doesn't change that. When Google introduced the scaled content policy in 2024, it said the policy covers content produced through automation, "human efforts, or some combination". A quick human check doesn't protect low-value pages.

Google's guidance on AI search also separates commodity content, based on common knowledge anyone could write, from content built on real expertise or first-hand experience. It asks creators not to recycle what already exists or what "could easily be produced by a generative AI model".

The bottleneck moved from production to judgment

AI solved the production bottleneck. It created a judgment bottleneck in its place.

When a draft costs almost nothing, the expensive work is the set of decisions around it:

• Is this worth publishing?
• Is it true?
• Does it add anything, or does it remix what's already out there?
• Does it reflect experience we actually have?
• Does it sound like us?
• Would we stand behind it if a customer quoted it back to us?
• Should it go live now?

Faster production means these questions come up more often. Ten drafts a week means ten sets of decisions a week, made by the same number of people.

The workplace version of this problem already has a name. Researchers at BetterUp Labs and Stanford call it workslop: polished AI output that doesn't move the work forward. Around four in ten workers they surveyed had received some, and each case took nearly two hours to sort out. The effort didn't disappear. It moved to whoever received the work, and on a public website that person is your reader.

What human-in-the-loop should mean for AI content

Human-in-the-loop for AI content means a person with the relevant knowledge, and the authority to reject, decides whether AI-generated work becomes production content. Their job is to own that decision. They don't need to write or edit every sentence.

Many workflows that use the term look like this in practice:

AI writes → human fixes grammar → publish

That's proofreading. It catches typos and awkward phrasing. Nobody in it is asked whether the piece should exist, and the reviewer rarely has a real option to say no.

A workflow that preserves decision rights looks like this:

AI researches → AI proposes → AI drafts → AI structures
→ human evaluates → human approves or rejects → system publishes

AI does more of the work in the second flow. The difference is who holds authority over what reaches production.

The European Commission has put this distinction into writing. Under Article 50 of the EU AI Act, which has applied since 2 August 2026, AI-generated text published to inform the public on matters of public interest must be labelled unless it has been through human review or editorial control. The Commission's guidance on Article 50 defines editorial control as the authority to approve, alter or reject the substance of a text, and states that spell-checking and grammatical correction don't qualify. The rule covers a narrow category of text, and most marketing content falls outside it. The definition is still useful for everyone else.

The costly failures tend to be decision failures. In 2025, the Chicago Sun-Times printed a syndicated summer reading list that recommended books which don't exist. The freelancer behind it had used an AI agent to help build the section. The paper's chief executive said the section went into the paper without its editorial team reviewing it. The fabricated titles came from a model. The bigger failure was that nobody with the authority to stop the section looked at it.

The goal isn't to keep humans busy. It's to keep humans accountable.

What the decision owner is responsible for

Decision rights come with responsibilities that belong to the organisation rather than to the writing process. Better tools don't take them away.

Choosing what to talk about

A model can propose a hundred topics. Picking the three that matter this quarter depends on what you sell, who you're losing deals to and what you want to be known for. The model can help with the thinking. The choice belongs to whoever owns the plan.

Bringing first-hand experience

Your most useful material comes from what you built, sold, tested, broke or heard on customer calls. A model can only work with what it has been given, and much of that experience has never been written down.

Holding internal context

The price change that hasn't been announced. The feature you're retiring. The partnership that's ending. Some of this shouldn't go into a prompt at all, and it's often the reason an accurate article is still wrong to publish this week.

Making brand calls

A statement can be accurate and still be something your company chooses not to say. Most positioning consists of those choices: the competitor you don't mention, the claim you don't make, the trend you don't chase. A model can follow your style guide well. Making or withholding a claim is a business decision.

Answering for published claims

When a customer relies on something you published, the source of the wording doesn't shift responsibility. In 2024, Air Canada argued that its website chatbot was responsible for its own answers. The British Columbia Civil Resolution Tribunal rejected that argument and held the airline liable, noting that Air Canada is responsible for all the information on its website. A chatbot differs from a blog post, but the principle carries over. The company answers for what appears under its name.

Why better AI models don't remove the need for decision rights

Models will keep improving, and many of today's review chores will shrink with them: invented statistics, broken links, off-tone phrasing. The case for decision rights doesn't depend on those errors. Three things hold however capable the model becomes.

First, the commodity ceiling rises with the model. Anything a model produces from a short prompt, your competitors can produce too. Better models lift baseline content for everyone, which lowers its value and pushes differentiation further towards the experience and context only you hold.

Second, information stays inside the organisation. Your roadmap, last week's customer call and the reasoning behind your pricing are inputs, and some of them should never be handed to a tool. More capability doesn't give a model access to them.

Third, accountability stays with people and companies. Someone signs off on a claim, and someone answers for it when a customer, regulator or journalist asks.

Better output also makes careless approval easier. When every draft looks polished, reviewers read less closely. The risk is well enough recognised that the OWASP Top 10 for Agentic Applications 2026 includes human-agent trust exploitation as one of its categories. Strong drafts make a deliberate review step more important.

When AI agents can publish, permissions matter more than prompts

Earlier AI tools produced text that someone copied into a CMS. The copy-paste step was clumsy, but it guaranteed a person handled the content before it went live.

Agents connected through MCP skip that step. We've covered how that connection works and why defined content operations beat letting an agent edit your source code. For content teams, the open question is what the agent is authorised to decide.

"Write me an article about onboarding" is a writing task. "Research something interesting about onboarding, write it up, publish it and share it on LinkedIn" hands the agent your website and your company's public voice. An error in the first costs a rewrite. An error in the second is live, indexed and in your followers' feeds.

Instructions can't carry that authority safely. In July 2025, SaaStr founder Jason Lemkin was building an app with Replit's AI agent and had told it not to make changes without his permission. During a code freeze, the agent deleted his production database anyway. Replit's fix was architectural: it moved to separate development and production databases.

That incident involved code, and content works the same way. A prompt that says "always ask before publishing" is a request the model may or may not follow. Credentials that can only create drafts are a control. So decide which permissions an agent gets for drafting, editing, publishing, deleting and distributing, and treat each one as a separate grant.

How to set up a human-in-the-loop AI content workflow

A workable setup splits the job three ways:

• AI provides production speed.
• The content system provides structure, workflow and governance.
• People provide judgment and accountability.

The content system is where decision rights become enforceable. If content lives in your codebase, review happens in code review. If it lives in documents, it gets pasted into the site with no record of approval. A structured content system gives agents a defined place to put their work, with a draft status that holds until a person with authority changes it.

Research
  ↓
AI proposes ideas
  ↓
Human picks what's worth doing
  ↓
AI drafts content
  ↓
Structured content system (draft)
  ↓
Human review and approval
  ↓
Publish
  ↓
Distribution (social, newsletter)

The human step near the top is deliberate. Judgment is cheapest before a draft exists. Rejecting a topic takes seconds. Rejecting a polished, finished article feels wasteful, and that reluctance is how weak drafts reach production.

These practices hold up in any CMS:

• Give agents credentials that create and edit drafts. Grant publishing, deletion and distribution separately.
• Name one owner for each piece of content.
• Review for substance using the questions above, and leave grammar to the tools.
• Treat rejection as a normal outcome. A review step that never rejects anything is proofreading.
• Record who approved what, and when.
• Treat distribution as a separate decision from publishing. An article can suit the blog and still be wrong for the company's LinkedIn page this week.

How this thinking shapes Garchi CMS

We're building Garchi CMS as a governed content layer where people, applications and AI agents work with the same structured production content, each with different rights.

Applications read published content through the API. Your team manages it in the dashboard. AI tools connect through Garchi's MCP server and work through defined content operations such as finding pages, drafting articles, updating sections and setting metadata. Content an agent creates or edits that way is saved as a draft, and your team controls when it goes live. Distribution follows the same rule: an agent can prepare a social post from an article and send it for approval, and a person publishes it.

The principle is that an agent can do a large share of the content work without holding authority over what the business publishes. This blog runs that way. Agents help with research and drafting, and I decide what goes live. If you want to see the flow end to end, the Lovable walkthrough covers it step by step.

Deciding what deserves to exist

The cost of producing another article keeps falling. The cost of publishing the wrong one hasn't changed: a misleading claim, a promise you can't keep, a post that makes a customer wonder whether anyone read it before it went out.

Production capacity used to separate content teams. Now everyone has it. The difference lies in how well a team combines machine speed with a clear owner for every decision about what goes live.

Your AI agent can write the article, but someone at your company still has to decide it deserves to exist.