1. Core Haven Labs
2. Blog
3. Why an AI reply never goes live unchecked

Why an AI reply to a review never goes live unchecked

30 August 2026 · 4 min read · Safety

A reply on Google appears under your business name. If something is wrong, that is your problem, not the AI's. That is why Auto-Reputatie is built around one rule: when in doubt, nothing goes out.

The problem with “just an AI”

A language model writes a friendly reply effortlessly. That is not the hard part. The hard part is what happens with the review you would rather think about for a moment: a guest writing that his daughter felt unwell after the meal, someone mentioning a lawyer, a complaint about discrimination. An automatic reply that puts “thank you for your feedback” under that is worse than no reply.

A half-right reply under your business name is worse than no reply.

Step 1: before the AI, the safety check

Every new review is first read by a check that writes nothing and only judges: is this about something sensitive? The system recognises eleven risk categories, including injury, health, food poisoning, discrimination, harassment, a legal claim, privacy and minors.

If a review falls into such a category, it is held back. Even at five stars. The owner gets a notification and a draft, and decides what happens. Automatic publishing can never bypass this rule; it is fixed in the system, not in a setting.

Step 2: the AI, with your rules

Only then does the AI write a proposal: in the tone the owner chose once (formal, friendly or enthusiastic), in the language of the review, and with only the contact details the owner entered themselves. The model always receives the review text as untrusted text: whatever a guest puts in a review cannot change how the system behaves.

Step 3: after the AI, the check on the reply

The proposal is checked before it goes anywhere:

• Does it contain a phone number, email address or website the owner did not enter themselves? Then it does not go online.
• Does it offer a discount, admit fault or repeat the complaint? Then it stays a draft.
• Does the language match the review? If not, it is done again.

Whatever fails the check is not published. It becomes a draft the owner can edit, approve or discard.

And after that: watch first, then let go

New businesses start in preview mode: every draft is visible, nothing is published. You turn on automatic publishing per star rating, only once you trust it. One switch turns it off again, any time. That makes the AI a colleague you train, not a system that happens to you.

Want to know how this works in practice? See Auto-Reputatie or read how we put two AI providers back to back.