Every CMS is adding AI. The pattern is remarkably consistent: a panel in the editor, a Generate button, and a body of text that appears in the field you were about to write in. Then a Publish button, which you press.
That arrangement is described everywhere as "human in the loop." It is more accurately human as a rubber stamp, and the difference is not philosophical — it shows up in what actually gets published.
The problem with review-by-default
The trouble with reviewing AI output at the moment of generation is that you are reviewing it in the worst possible state of mind. You asked for it. You are expecting it. It reads fluently, because fluency is the one thing these systems are reliably excellent at. And the alternative to accepting it is writing the thing yourself, which is the work you were trying to avoid.
Nobody rejects the third paragraph on a Friday afternoon. The result is a slow drift: content that is plausible, on-brand, and not quite anything anyone verified.
Oversight that happens at the moment of maximum motivation to approve is not oversight. It is a formality with a button.
What we did instead
In Birkly, an agent writing through MCP does not write to published content. It creates a draft, and that draft lands in an approval queue.
The queue is the same one that holds submissions from public forms. This was not a shortcut — it is the actual design claim. A contribution from an AI agent and a contribution from a stranger filling in a form on your website are the same category of thing: content that arrived from outside your editorial process and has not yet been accepted into it. Treating them identically means the review step is a normal part of the working day rather than a special ceremony that only fires when AI is involved.
The consequences are concrete. You review it later, in a queue, alongside other unreviewed things, without the generation context making you feel invested. You see it as a diff against nothing rather than as text that has already appeared in your editor. And nothing an agent produces is ever live before a human with the right role has said so.
What an agent can and cannot do
Being specific here matters, because "AI-native CMS" is a phrase currently doing a lot of unearned work.
Over MCP, an agent connected to a Birkly instance can read broadly: list collections, read schemas, read entries, read the site's project files, read the published Agent Guide. Reading is where most of the value is, and it is safe.
Writing is where the gate sits. Creating and updating entries, creating sub-entries, changing collection schemas — these are approval-gated. The agent proposes; a human with permission disposes. Destructive operations and anything touching roles, users or settings sit behind stricter permissions again, because the blast radius of a confidently wrong agent is not the same as the blast radius of a confidently wrong sentence.
The roles system is doing real work here. An agent is not a superuser with a friendly interface. It is an identity with a role, and it can do exactly what that role permits and nothing else — which is the same sentence we would write about a freelancer with an admin login.
The objection: this makes AI slower
It does, and we want to be honest about that rather than pretending the trade-off is free.
If your goal is volume — forty product descriptions by Thursday — a review queue is friction, and a CMS that writes straight to published will get you there faster. We are not going to argue you out of that use case. We are going to point out that the review step is the entire difference between a CMS you can trust with your public content and a very fast way to publish things nobody read.
There is also a compounding argument. Content that goes out unreviewed becomes content that agents later read, cite, and build on. If you are publishing to a web where AI systems are increasingly the readers, the accuracy of what you publish stops being an internal quality matter and becomes an input to what everyone else's assistant says about you. That is the case the Agent Guide is built on, and it only works if the pipeline into your content has a human in it.
Human in the loop as a workflow, not a slogan
The test we hold ourselves to is simple: can a person publish AI-generated content without a deliberate act of review? If the answer is yes, the safeguard is decorative regardless of how it is described in the marketing.
In Birkly the answer is no. Not because we have made AI hard to use, but because we put the agent where an outside contributor goes — which is where, on an honest reading of what these systems are, it belongs.
Birkly will be free and open source at release. Join the waitlist to try the workflow yourself.