Letting AI draft client emails without letting it send them
AI writing client emails is only safe when the brand rules, the preview and the send controls sit outside the model. Here's the loop.
6 min read
Why prompt instructions aren't controls
Telling a model "always use this sender address" is a preference, not a guarantee. Anything that must be correct every single time — sender identity, logo, signature, booking link, unsubscribe — belongs in the code that builds the email, where the model can't get it wrong.
The model's job is the words. Everything structural is decided server-side from context: which project, therefore which brand, therefore which sender and which links.
The loop
Five steps, in this order, every time.
- 01Draft — the model writes against real CRM data and the project context.
- 02Preview — the finished email renders at a private link, exactly as the recipient will see it. Never paste raw markup into a chat window.
- 03Approve — an explicit yes from a person. No implicit sends, ever.
- 04Confirm again for bulk — anything above a small threshold needs a second confirmation with the recipient count shown.
- 05Send — through the CRM, with caps, opt-out checks and pacing applied underneath.
The guardrails underneath
Daily send caps, do-not-contact and unsubscribe exclusions, deduplication so one person can't receive two variants of the same message in a week, and a log of what was sent to whom. These run regardless of what the model produced.
Make the audience visible before the send
Show the real recipient count, broken down by segment, at the moment of approval. A count that turns out to be ten times larger than expected is the single most common bulk-send accident, and showing it is the cheapest possible prevention.
Who this suits
It works best for a non-technical owner or director who knows exactly what they want to say and doesn't want to spend an afternoon formatting it. They ask, they read the preview, they say yes. The system carries the rest.