What to automate—and what to keep human
Good candidates: turning bullet notes into a clear memo, restructuring a messy outline, shortening a long update for a different audience, checking consistency of public facts across documents. Poor candidates without special controls: disciplinary letters, safeguarding material, confidential personnel narratives, or unfinished board strategy that should not leak.
A reliable pattern is human facts → AI structure → human edit → human send. Reverse that order for consequential messages and you will eventually publish something clever and wrong.
Prompt patterns that preserve institutional voice
Give the model audience, purpose, length and must-include facts. When the system allows, supply a short sample of past writing. Ask for options (formal / warmer / shorter) rather than one “perfect” answer. Require assumptions to be listed so editors can catch invented detail.
Build a small library of approved prompts for common tasks—parish notices, chancery circular scaffolds, school parent updates—so quality does not depend on one clever staff member with a secret prompt notebook.
Risk controls for everyday drafting
- Strip unnecessary names and identifiers from prompts
- Use organisational accounts, not personal free chatbots, for work content
- Keep a human responsible for every outbound message
- Log which tool is approved for which class of document
- Train new staff with bad/good examples from your context
For privacy-sensitive threads, return to faith-based data privacy before expanding AI use.
How to know it is working
Track time-to-first-draft, revision rounds and factual corrections. If AI saves ten minutes but adds two errors per letter, you have not won. Celebrate editors who catch mistakes—that is quality culture, not obstruction. Over a quarter, you should see fewer blank-page stalls and no rise in embarrassing corrections after send.