Where the mistake lands
Ask for “change the Friday hours to 9–17” and you might get Saturday instead. Ask for a price update and a decimal point can move. These are ordinary misunderstandings - the kind a new colleague makes in their first week.
The cost of that mistake is wildly different depending on where it surfaces. In a preview it costs you one sentence: “no, Friday, not Saturday.” On a live page it costs you a customer standing in front of a closed door, a price you now have to honour, or a screenshot in a group chat. Search engines and AI assistants may have already read the wrong version and cached it.
So the question worth asking a provider is not “how often is your AI wrong?” It is “what happens when it is?”
Approval is a boundary, not a formality
In a Conversational CMS the chat is where you request a change. It is not where the change goes live. The system prepares it, shows you what it did, and stops. Publishing is a separate, deliberate act - and it belongs to a person.
That boundary is also where accountability sits. If the system could publish on its own, nobody could say who decided that the page should read what it reads. With an approval step the answer is simple: the person who clicked approve, after seeing exactly what changed.
This is the line between a conversational system and an agentic one. Agents are given goals and act on them, with a human reviewing by exception. That model has real uses. It also quietly moves the decision away from you, and for a business website that trade is rarely worth it.
What it changes day to day
You see the change before anyone else does, on a real preview of your own pages, not a description of what was supposedly done. If it is wrong, you say so and get another version - no ticket, no deployment window.
Nothing is published while you sleep, and nothing is published because a model was confident. Your public site keeps serving the last version you approved until you approve the next one.
That is also the honest limit of this model: approval does not make AI accurate. It makes AI’s mistakes cheap, visible and reversible before they reach anyone who matters.
We could drop the approval step and publish faster. We keep it because a website is a promise to customers, and a promise should be made by a person - even when a machine writes the sentence.
If you want the wider map - how a Conversational CMS compares with a traditional CMS, an AI website builder, headless and agentic systems - we wrote it down in What is a Conversational CMS?