Scale X's August 5, 2026 report describes a browser game with more than 40,000 runs and 409,000 approve/deny decisions. It reports that the average player missed about one in three seeded threats. Participants faced artificial time pressure, and roughly a third of the commands were threats. These are the author's game statistics, not a representative study of law-firm reviewers or a production error rate. Original game report.
Our implementation takeaway is a recommendation: give a reviewer useful context and enforce permissions independently of the approval interface. The game does not tell a firm how well its own review process works.
Show what the person is approving
For a proposed message, show the recipient, content, relevant context and attachments. For a record change, show the target and proposed difference. Explain which source supports the action and what remains uncertain.
These are design recommendations. Their effectiveness depends on the workflow, implementation and people using them; this article does not establish a review success rate.
Reduce what can go wrong before review
Restrict access to the records and actions the workflow requires. Keep sensitive actions behind enforced permissions. A model instruction asking the system to be careful is different from a software control that refuses an unauthorized operation.
Define escalation with the firm
Agree which situations require a staff member or attorney, how the system stops and how responsibility transfers. Use representative and unusual cases to test the process. Do not rely solely on the model declaring its own confidence.
Verify the outcome
Record the proposed action, reviewer decision, execution result and any exceptions that the firm needs to reconstruct the event. Review failures and retest after material changes. Logging should respect the firm's access and retention requirements.
Request a Firm Intelligence Audit to discuss which workflows merit review.
Editorial correction: unsupported or unavailable-source claims were revised. Original publication date retained; this correction does not certify every other article in the archive.