Convo-Ease: A Policy-as-Prompt Gatekeeper for Real-Time Multimodal Moderation in Enterprise Communication

Abstract

Enterprise communication platforms carry text, image, and audio content across organizational contexts that public-facing moderation systems were never designed to serve. Existing moderation acts after content reaches recipients, depends on fixed harm taxonomies that cannot be updated without retraining, and rarely covers more than one or two modalities. Organizational settings require the opposite: validation before delivery, team-specific policies that change often, and coverage across every modality a message can arrive in. This paper contributes Convo-Ease, an applied moderation architecture defined by four design decisions and realized as a working system. A pre-delivery gatekeeper validates content structurally in the message-send path, so a message that fails evaluation is never shown to other members. A unified text abstraction reduces every modality to a single representation before evaluation, so one policy layer governs all three. A Policy-as-Prompt mechanism reads organizational rules as plain-language instructions at inference time, so a policy change is a configuration edit rather than a retraining job, and no model training is required at any point. A plugin-based backend separates each moderation task from the model that performs it, so a hosted application programming interface model and a local open-weight model are interchangeable per modality through configuration alone. We validate the central architectural claim on a held-out set of 300 human-labeled messages evaluated live under matched conditions on the same model with no parameter update between them: replacing a generic safety taxonomy with organization-specific natural-language rules raises recall from 46.0% to 76.7% and accuracy from 68.3% to 76.7%, a paired difference that is statistically significant (McNemar's exact p = 0.004). The same mechanism holds across fifteen distinct organizational rule sets without any per-rule retraining, and an independent enforcement-intensity control shows a measured, asymmetric effect on the operating point. The quantitative evaluation covers the text policy-evaluation core; the multimodal contribution is architectural and is not accompanied by end-to-end measurement on real image or audio media. Live policy evaluation against a hosted model sustains sub-second batched latency, within the budget required for real-time chat. Convo-Ease demonstrates that an adaptable, multimodal, pre-delivery moderation system can be assembled from interchangeable off-the-shelf models plus an architecture, with no training in the loop, and that this design occupies a niche no existing system fills.