How AI follow-up drafts actually work — no marketing fluff

Every CRM ships an “AI” feature this year. Most of the marketing copy is a haze of capability claims and zero substance about what’s actually happening under the hood. That’s a problem because the questions buyers actually have — what gets read, what gets stored, who sees the drafts, what happens when the model is wrong — are exactly the questions the brochures dodge. Here’s how Vanta CRM’s AI follow-up drafts work, with no marketing fluff.

What context the model reads

When you click “draft follow-up” on a deal, the model gets a carefully scoped bundle of context:

• The prior email thread on this deal. The most recent ten messages, including subject lines and timestamps.
• The lead and contact record. Name, title, company, and any custom fields you marked as “include in AI context.”
• Public notes on the deal. Notes you’ve added, including call summaries and meeting recaps.
• The current pipeline stage. So the draft matches where the deal actually is — a discovery follow-up reads differently from a proposal follow-up.
• Your tone calibration. A short style description you set once per user, so the drafts sound like you and not like a chatbot.

What the model does not read

This is the part vendors usually skip. The model never sees:

• Any other customer’s data. Tenant isolation applies to AI calls the same way it applies to database queries. The request bundle is scoped to your workspace at construction time. More on this in tenant isolation explained.
• Private notes you’ve flagged. Any note marked private is excluded from the prompt.
• Custom fields you haven’t opted in. Custom fields default to not included; you have to explicitly mark a field as AI-readable.
• Attachments. File contents are never sent to the model.

The draft never sends itself

Every AI draft lands in an approval queue. The flow is:

1. Model produces a draft.
2. The draft appears in your queue with the proposed subject, body, and recipient.
3. You read it, edit it, or discard it.
4. Only after you click “send” does anything leave your inbox.

There is no auto-send mode. We’ve been asked for one and we’ve said no for the same reason airlines don’t let passengers fly the plane: the cost of a single bad send is much higher than the inconvenience of the click.

How tone calibration actually works

Tone is the part most AI features get wrong. The default voice of a large language model is somewhere between “LinkedIn motivation post” and “hotel concierge,” which is the wrong voice for almost every sales context.

Calibration in Vanta CRM is a one-time setup per user. You paste two or three of your real past emails into a settings field, and we extract a short style description — sentence length, formality, common sign-offs, whether you use contractions. That style description is prepended to every draft prompt. It’s not magic, but it’s the difference between drafts that sound like you and drafts that sound like a robot dressed up as you.

What happens when the model is wrong

The model gets things wrong. Sometimes it invents a feature that doesn’t exist, sometimes it misreads the tone of the prior thread, sometimes it just produces a bland email. Three things to know:

• You see it before anyone else does. The approval queue is the safety net. A bad draft costs you the four seconds it takes to delete it.
• You can give feedback in-line. Editing the draft before sending teaches the calibration over time. Discarding with a reason does the same.
• You can turn it off per deal. Some deals are too sensitive for an assist. Mark the deal “manual only” and the AI draft button disappears.

If you’re nervous about AI in sales, you’re right to be

Most AI sales features assume the salesperson is the bottleneck and the model is the solution. That’s usually backwards. The bottleneck is judgment, and the model is a useful first draft. Treated as a draft, AI saves you fifteen minutes per follow-up. Treated as the final word, it produces the kind of generic outreach that trained everyone to ignore generic outreach.

See it in action on the product page, or read about our security defaults if you want to understand the guardrails before you turn it on.