AI drafts now only mention meetings that actually exist

When your AI drafts a reply to a prospect, the details in that reply now have to be real. A meeting time, a meeting link, or any other specific claim gets pulled from the actual conversation, your knowledge base, or a real booking — never generated on the fly because it sounds plausible.

What changed

Before this update, a drafted reply could in principle include a meeting link or a time slot that read as completely normal but didn't correspond to anything that actually existed. The underlying model was doing what generative models do when a detail is missing from context: filling the gap with something plausible instead of leaving it out or checking. Most of the time that went unnoticed, because a plausible-sounding link or slot looks exactly like a real one until someone clicks it or shows up at the wrong time. That's what made it risky — the failure mode was invisible right up until a prospect acted on it.

That path is now closed. Any reply that references a meeting has to trace back to something real: the actual link from a real booking, the actual slot that was actually confirmed, or actual details already sitting in the thread. If a meeting hasn't been booked yet, or the thread doesn't contain a real link, the draft won't fabricate one to fill the space. It either omits the detail or reflects only what's genuinely known at that point in the conversation.

This is now visible directly in the product, not just in backend behavior. In the Inbox, when the AI drafts a reply that confirms or references a meeting, the draft carries the actual booked slot and the actual meeting link — not a placeholder, not a rounded-off approximation, and not a plausible guess standing in for the real thing.

How to use it

1. Let a conversation progress to the point where a meeting actually gets booked — through your booking flow, or through a time explicitly confirmed in the thread.
2. Open the Inbox and locate the AI-drafted reply tied to that conversation.
3. Check the meeting reference inside the draft. The link and the time should match, exactly, what was actually booked — no rounding, no invented backup link, no filled-in placeholder text.
4. Continue your normal review pass before sending. You should notice this specific failure mode — a hallucinated link or a meeting time that was never booked — showing up far less often, if at all, since the underlying data the draft pulls from is now real by construction rather than generated to look real.
5. If a meeting hasn't been booked yet in that thread, watch what the draft does instead: it won't manufacture a link or a slot to sound complete. It sticks to what's actually confirmed so far, which is the behavior to expect going forward.

Why it matters

A cold outreach campaign runs entirely on trust that's built one reply at a time, and nothing spends that trust faster than a rep sending a prospect a meeting link that doesn't resolve, or confirming a time slot that was never actually booked on either end. That kind of mistake doesn't just look sloppy — it lands at the exact moment a prospect is deciding whether to take the meeting seriously, and it can cost the deal on the spot, with no good way to recover the credibility afterward.

Before this change, that risk existed quietly under the surface: a generated reply could include a detail that was wrong in a way neither the rep nor a quick skim would necessarily catch, because a fabricated link or time looks identical to a real one until it's tested. That's a hard thing to guard against manually, review after review, especially at the volume outreach teams run replies through.

With this change, you can lean on AI-drafted replies with real confidence instead of treating every meeting reference as something to double-check by hand. You're not trading speed for risk anymore — the drafts are grounded in what's actually true about the conversation and the booking, so what you send matches what you actually promised. Your review step still matters as a general practice, but the specific category of error where the AI simply invented a fact is gone, which means you can move faster on the replies that matter most: the ones confirming a meeting is really happening.