Situation

On our platform for high-conflict co-parent communication, every message is legally significant, often emotionally loaded, and sometimes headed to a courtroom. Human reviewers checked each one against strict communication guidelines before it reached the recipient — no hostility, no manipulation, no off-topic escalation.

Complication

The judgment calls were the easy part for experienced reviewers — the problem was throughput. Every message got equal scrutiny regardless of how likely it was to be a problem, so reviewers spent as much time confirming routine, benign messages as they did catching the genuinely concerning ones. A false negative could let hostile language through to someone in a vulnerable position; a false positive could delay a message that was actually fine. That risk profile ruled out handing moderation decisions to a model outright.

Resolution

We integrated an AI tool that reads each message before a human does and flags the ones that look likely to violate guidelines — hostile tone, veiled threats, subject changes into contested topics like custody or finances. It doesn’t approve or block anything. It sorts the queue.

That distinction mattered for two reasons:

• Accountability stays human. Every message that reaches a recipient has been approved by a person, so there’s a clear record of who made the call — important when transcripts end up as court documentation.
• Reviewers spend attention where it counts. Flagged messages get closer scrutiny; routine ones move through faster. The review is still complete, just no longer evenly distributed across messages that don’t need it.

Results

Reviewers cover the same message volume in less time, with more attention landing on the messages that actually need it, contributing to the platform’s overall 80% reduction in time spent on manual tasks. Consistency also improved — a flagged pattern gets caught the same way every time, rather than depending on which reviewer is on shift or how many messages they’ve already read that day.

Back to Case Studies