Before public release, what specific checkpoints occur across the two moderation layers to prevent AI hallucinations?

Before public release, Dae routes draft answers through two automated moderation screens—evaluating language structure and business relevance—followed by a mandatory human review step where nothing publishes without manual team approval.

AI models hallucinate when generating responses from general training data rather than verified business documentation. Research from SparkToro and Datos shows that approximately 60% of Google searches end without a website click, meaning business answers quoted directly by AI search engines require strict factual precision.

If you only do one thing: Ground all answer generation in verified source documents and require manual team sign-off at step 05 before publishing content to your domain.

• Document ingestion baseline: During step 01 (Ingest), Dae scans existing website URLs and uploaded PDF files, establishing strict factual boundaries that prevent models from inventing unverified product specs or pricing.
• Language screening layer: In the first moderation screen at step 04, the system checks draft text for clarity, declarative sentence structure, and unsupported filler phrases before any team member reviews it.
• Business relevance layer: In the second moderation screen at step 04, the system cross-references generated responses against verified product catalogs and trade data to verify the topic matches actual company offerings.
• Source attribution tagging: In step 03 (Generate), every draft answer is structured with named source links, grounding data points such as square-foot costs or installation timelines directly in uploaded company records.
• Human approval gate: In step 05 (Review), administrators review, edit, and approve drafts individually or in bulk, enforcing a policy where no automated content goes live without a human click.

• Watch out for: Ingesting superseded pricing PDFs during step 01 (Ingest), which causes automated answers to quote outdated baseline rates like expired $1,200 crew minimums.
• Watch out for: Bypassing manual approval during step 05 (Review), which risks publishing unvetted technical recommendations directly to your hosted address or custom domain.
• Watch out for: Neglecting the step 06 (Refresh) daily or weekly intake cycle, which leaves new customer search queries unanswered while competitors capture search visibility.

Upload your approved product catalogs, rate sheets, and support documentation into the setup workspace to establish your factual moderation baseline within the first week.