What AI Content Governance Actually Means
Content governance is not spell-checking. It is the enforcement of rules that protect the business brand voice rules, legal disclaimer requirements, product naming standards, pricing accuracy, regulatory compliance, and content freshness. AI governance means encoding those rules into a system that checks every page, every time, without fatigue or inconsistency.
Part 1 – The Governance Architecture
The system has four layers:
- Layer 1 – Content Sources: AEM Pages via Sitemap API, Content Fragments via CF API, pre-publish hook from Sidekick, Brand Rules stored as a Content Fragment
- Layer 2 – AI Scan Engine: App Builder action calls Claude API with brand rules as system prompt, returns structured violation JSON
- Layer 3 – Governance Workflow: Critical violations to AEM Inbox immediately, warning-level to weekly digest, audit log per page
- Layer 4 – Enforcement Layer: Pre-publish gate blocks critical violations, auto-unpublish for high-severity findings, rollback to last clean version
Part 2 – Brand Rules as a System Prompt
The AI governance engine works by encoding brand rules as a Claude API system prompt. Store this in an AEM Content Fragment so the brand team can update rules without any developer involvement. Key categories: product naming rules, legal requirements, voice and tone, content freshness.
Part 3 – The App Builder Scan Action
The scan action calls claude-sonnet-4-6 with the brand rules system prompt and page content, then parses structured JSON violations. Enable prompt caching for the brand rules system prompt, it is identical for every page scan, reducing cost and latency significantly.
For bulk scans, process pages in batches of 10 with a 200ms pause between batches to respect Claude API rate limits. Results are stored per page with violation history.
Part 4 – Violation Severity Tiers
Three severity tiers, each with a different response:
- Critical – Block and alert. Legal disclaimer missing, discontinued product promoted, pricing error. Page is auto-unpublished if already live. Brand team notified immediately via AEM Inbox.
- Warning – Flag for review. Stale statistics, borderline superlative claims, off-brand terminology. Page stays published. Flagged in weekly brand digest.
- Info – Recommend only. Tone suggestions, passive voice in CTAs. Surfaced in Sidekick governance panel as optional improvements. No workflow triggered.
Part 5 – AEM Sidekick Governance Panel
A custom Sidekick v6+ extension adds a real-time Brand Score panel (0–100). Authors run a governance check against the current draft before hitting Publish. Critical violations show with specific fix suggestions. The panel calls the App Builder action never the Claude API directly.
Part 6 – Nightly Bulk Scan
A scheduled App Builder action runs at 2am each night. It fetches the full sitemap, extracts page content, and runs governance scans in batches of ten. Critical violations trigger immediate Inbox notifications. Warning-level findings compile into a weekly digest. For sites with more than 1,000 pages, run over 3-4 hours rather than all at once.
Part 7 – Auto-Unpublish for Critical Violations
When the nightly scan finds a critical violation on an already-published page, the system auto-unpublishes via AEM Replication API and routes it for urgent review. The last clean version is stored for one-click rollback. The brand team receives a task in AEM Inbox with the violation detail and fix suggestion.
Measuring Governance Effectiveness
- Violations caught pre-publish vs post-publish – target 95%+ pre-publish within three months
- Mean time to resolution – target under 4 hours for critical, under 5 days for warnings
- Rule hit rate – high-hit rules signal a brand guidelines training gap
- False positive rate – above 10% means the rule prompt needs tightening
- Percentage of pages with zero violations – the clearest signal of programme health
