Follow-up Sequences

A customer asked something and went quiet. A follow-up is a message sent into that silence, which makes it the easiest thing in the product to get embarrassingly wrong. Most of the machinery here exists to decide not to send one.

How steps are scheduled#

  • When a conversation needs follow-ups, every step is created at once, each with its own scheduled time and its own timed job.
  • Offsets are sorted and de-duplicated first, so a messy configuration still produces a sane order.
  • Email offsets are in days, chat channels use hours. A day-based schedule is ignored on a chat channel and the other way round. The units aren't interchangeable.
  • Each step fires at its own moment rather than being picked up by a periodic sweep, so a 6-hour nudge lands at 6 hours.

Content isn't written at schedule time. It's generated when the step fires, from the conversation as it stands then - which is why a follow-up reflects knowledge you added after the original reply.

Meta's window at schedule time#

On WhatsApp and Instagram, free-form messages only work inside 24 hours of the customer's last message. Splice enforces that when scheduling, not when sending:

  • The window end is computed from the customer's last inbound message.
  • Any free-form step that would land after it is never created, rather than being created and later failing.
  • On WhatsApp, your configured template steps are then appended after the free-form ones, with their own day-based delays - these are valid outside the window.
  • Template steps can carry a condition, so a step that only makes sense for contacts without an email on file is skipped for everyone else.

No templates means silence after 24 hours

With no in-window hours configured and no template steps, nothing is scheduled at all. This is logged rather than silently ignored, but the practical effect is a WhatsApp conversation that simply stops.

What happens when a step fires#

A scheduled step is a plan, not a promise. When its moment arrives, it's re-validated against the world as it is now - because a lot can change between scheduling and firing.

Everything that cancels a step#

Condition at fire timeWhy it cancels
The agent is inactiveA paused agent shouldn't speak. The step is cancelled, not deferred.
The lead is Active Client, Closed or LostChasing a customer you already serve, or someone who said no, is worse than staying quiet.
The conversation no longer needs oneThe customer replied, or it was resolved in the meantime.
The channel can't carry itFollow-ups exist for email, WhatsApp and Instagram. A live channel gets nothing.
Automation was paused on the conversationPer-conversation settings win over the agent's defaults.

Cancelled steps don't come back

A cancelled step is finished. If the conversation later goes quiet again, a fresh sequence is scheduled from that point - you don't get a stale message from three weeks ago arriving out of context.

Meeting-aware suppression#

The last check is the subtlest. A follow-up that ignores what was agreed on a call reads as nobody having listened, so the drafted message is compared against the last meeting's outcome before it goes:

SituationOutcomeWhat you see
The draft contradicts a decision from the callEscalateNeeds review - contradicts call.
Someone on your team owes the next step, and the draft doesn't acknowledge itSkipPaused - waiting on your next step.
The agent owes the next step, but the draft doesn't reference itHold for contextHeld rather than sent, with the reason.
A meeting happened but the transcript hasn't arrivedAsk for contextPaste call notes to re-score.
Nothing conflictsSendNormal auto-send rules apply.

Each of these lowers the confidence of that specific draft rather than blocking it globally, so the existing threshold does the actual holding.

Sending or drafting#

A follow-up that survives every check is treated like any other reply: high enough confidence plus the channel toggle means it sends, otherwise it waits in the inbox as a draft. There is no separate follow-up auto-send setting to forget about.

Designing a sequence#

Cadence

  • Two or three email steps. The third is where you offer to close the loop, and there shouldn't be a fourth.
  • One WhatsApp nudge inside the window. Chat is intimate, and three messages into silence reads as harassment.
  • Configure templates if WhatsApp matters to you - that's the only thing that works after 24 hours.

Content

  • Give each step a different job in the agent's instructions, so check in, then add something useful, then offer to close it out, rather than three variations of 'just following up'.
  • Let it reference the specific thing discussed. The model has the conversation and the knowledge base to hand.
  • Trust the cancellations. Most of the value here is in the messages that never send.

For the settings, see Follow-up Automation.