How a campaign operations platform used AI as a first-pass review layer for content quality, destination safety, and manual moderation routing.
Context
A campaign operations platform needed a faster way to review user-submitted content without handing every decision to automation. The platform handled several related content types: events, campaigns, adverts, and short links used for external destinations.
The goal was not to let AI silently decide everything. The right approach was more controlled: use AI to clear low-risk submissions when confidence was high, then keep anything uncertain, incomplete, unsafe, or operationally blocked in a manual review queue.
Challenge
Content quality and destination safety are connected, but they are not the same problem. An advert or event might have acceptable text while still linking to a poor, unsafe, or unverified destination. Treating those checks as one simple approval flag would create avoidable risk.
There were also different readiness rules for each content type. Events needed coherent dates, venues, descriptions, calls to action, and links. Campaigns needed clear propositions and responsible targeting. Adverts needed valid creative, processed media, complete copy, and an active destination URL.
The system needed to move quickly for straightforward cases while keeping enough control for administrators to handle edge cases properly.
Approach
We built the review process around layered checks rather than a single AI verdict.
When a user submits an item, the API first validates required fields and workflow state. Eligible items move from draft or rejected into pending review. From there, an AI review job is queued through background infrastructure when configured, with local inline processing available for development.
Each review builds a structured payload from the fields that matter for that content type. Events include title, descriptions, schedule, venue details, calls to action, tags, platform assignments, and external links. Campaigns focus on proposition, description, campaign type, and targeting. Adverts include the current creative version, headline, body, call to action, destination URL, and route data.
AI approval is only allowed when the review passes and confidence is above the configured threshold. Failed reviews, low-confidence results, unavailable model responses, and incomplete linked checks all remain in pending review.
Short links sit behind a separate destination-safety layer. The system validates URL format, checks domain allow and block lists, and gathers page evidence where available, including HTTP status, content type, page title, OpenGraph data, meta description, and page text excerpts. AI then classifies the destination for safety risks such as adult content, violence, drug promotion, hate content, phishing, malware, fraud, and impersonation.
Safe destinations can become active. Unsafe destinations are disabled. If a disabled destination is already linked to approved content, enforcement can cascade by moving that content back out of approval and recording why.
Human review where it matters
The important product decision was that AI uncertainty does not automatically become rejection. Most failed or unclear outcomes stay in manual review, giving administrators a clear queue rather than hiding the decision inside a model response.
Manual reviewers can approve or reject campaigns, adverts, and events. Those decisions still emit activity logs, moderation completion events, and organisation notifications, so there is an audit trail for both automated and human decisions.
Outcome
The platform gained a review process that can handle routine approvals faster while keeping riskier cases visible to people. Content teams no longer need to inspect every submission manually, but they still keep control over ambiguous or policy-sensitive cases.
The wider benefit is clearer operational trust. Form validation catches missing data, media checks confirm creative readiness, short-link moderation protects external destinations, AI handles first-pass content quality, and human review remains available when judgement is needed.
Key takeaway
AI moderation works best as part of a controlled workflow, not as a single black-box decision. By separating content review, destination safety, media readiness, and human oversight, the platform could speed up approvals without losing accountability.