AI-Drafted Follow-Ups: What Actually Works When the Machine Writes Your Email

The first generation of AI email tools turned out a tidal wave of obviously-generated sludge. "I hope this email finds you well, I noticed your company is in the SaaS space and I think we could be a great fit…" at scale.

Open rates collapsed. Inbox providers started flagging the patterns. Recipients learned to skim for the tells in the first two lines and bin anything that looked machine-written.

The second generation is finally good. The difference isn't a smarter model, it's the context the model gets and the rails the user puts on the output.

Here's what changed, what the prompts look like, and what to insist on before you let any tool ship a message in your name.

What changed

Three things, in order of importance:

1. The model sees the relationship, not just the task

A first-generation tool prompted the LLM with "write a follow-up email to a SaaS company". A second-generation tool prompts it with the booking details, the guest's CRM record, the last six months of activities, the event-type title, and the trigger that fired the workflow.

The model doesn't have to guess. It writes the email that anyone with the same context window would write.

2. Templates are skeletons, not scripts

The first-generation pattern was "send this exact email, with the prospect's name substituted in." Recipients spot it in two seconds because they get five of them per day from five different senders.

The second-generation pattern is "here's the variable palette ({{guest_name}}, {{event_title}}, {{start_time}}, {{rating_url}}, {{cancel_url}}, {{reschedule_url}}, {{organizer_name}}), write the email and use them where natural."

The output varies per recipient because the model can decide whether to start with "Hi Jane" or "Jane, quick note" based on the rest of the context. Every email is a one-off.

3. The human signs off, every time

You can have the AI draft, but you don't let the AI ship without a glance. Real second-generation tools snapshot the previous draft so you can revert in one click, then ship.

Most of the time, the AI draft is 90% right. You change two words. Three minutes saved, eighteen times a day.

What a good prompt looks like

The single best discovery we made while building ThorStack's Draft with AI panel:

Don't tell the AI to "write an email". Tell it what the email should do.

Bad prompts:

"Write a thank-you email after a demo."

"Send a follow-up to a prospect who didn't show up."

Good prompts:

"Warm 3-sentence thank-you. Mention we'd love any feedback, keep it casual, sign off with my name only. Don't say 'I hope this finds you well'."

"Non-judgmental no-show follow-up. Lead with 'hope everything's OK'. One paragraph max. Include the reschedule link near the end as a soft offer, not a demand."

The model is genuinely good at executing intent. It's terrible at improvising tone.

What to insist on before letting a tool ship anything in your name

A checklist, distilled from a lot of late-night production debugging:

1. The user sees the draft before it sends. Always. No "AI autopilot" mode that ships without your eyes on it.

2. HTML escaping. If your template body says <script>alert(1)</script> because someone pasted it accidentally, the recipient should see <script> in the inbox, not actually execute it. Sounds obvious. A surprising number of tools get this wrong.

3. Variables fail loudly. If the prompt mentions {{customer_company}} and the variable doesn't exist, the recipient should see {{customer_company}}, not an empty space where the company name should be. Loud failures get caught in QA; silent failures get caught by the customer.

4. The draft is stamped. The reviewer should see "drafted by AI, you edited 4 lines", not just an unsigned email. Audit trail matters more than people realise once you've shipped a thousand of these.

5. The model knows the trigger. A post-meeting thank-you and a no-show follow-up share zero DNA. If the AI doesn't know which one it's writing, you get a thank-you sent to a no-show, which is hostile.

6. Undo is one click. Snapshot the prior draft, surface it as a button. The user is going to want it.

ThorStack's Draft with AI ships with all six. We hold the line because we got the first generation wrong too, and our customers' inboxes are the ones who pay if we get it wrong again.

The leverage math

A salesperson sends roughly 35 follow-up emails a week. Each takes 4 minutes when done from scratch. That's 2.3 hours a week, 115 hours a year, about 3 full working weeks lost to writing follow-up email.

If the AI draft is good enough that you spend 90 seconds per email instead of 4 minutes:

• Time saved per week: ~95 minutes
• Time saved per year: ~80 hours, two working weeks
• Quality improvement: typically higher, because the AI doesn't get tired at 4:47pm on a Friday

For a 5-person sales team, you're recovering 10 working weeks of output every year by letting the machine do the typing.

Where it goes from here

The next move, and we're already shipping this, is for the workflow itself to be conditional. "If the rating was ≥4, send the testimonial-request template. If ≤2, send the recovery template. If a comment was left, include a paraphrase of it in the response."

This stops being "AI email" and starts being "an actual sales-ops process that runs itself."

It's mundane and it's the future. We're betting on the mundane.

How to try it

Book a demo. Bring a prospect record from your current CRM, we'll wire a workflow against it live and let you see exactly what the draft looks like before it ships.