AI Follow-Up Systems: sequencing without the spam.

State of inbox deliverability in 2026

Google and Microsoft both shipped classifiers in 2025 that specifically target template-with-token outbound. The visible signal: open rates dropped industry-wide by 15–25%; reply rates dropped less, but bounce rates climbed.

The implication: send fewer, better messages. The volume-first playbook of 2022 doesn't work anymore. The teams that haven't adjusted are watching their numbers degrade quarter over quarter and blaming 'inbox fatigue,' when really the classifiers have just gotten better.

Training a voice model on your own threads

Generic 'professional B2B' output is recognizable now. Inbox classifiers and humans both spot it. The fix is training the drafting layer on your team's actual closed-won email threads — the ones where a real person wrote a message that got a real reply that became a real deal.

Practically: feed 50–200 example threads into the system, label which messages got responses, and the model learns the cadence, vocabulary, and structural patterns your specific team uses. The output stops looking like 'AI cold email' and starts looking like your team.

Sequence structure that still works

The format that's holding up in 2026:

1. Touch 1 — short, specific, useful. One concrete reference to something verifiable about the lead, one specific question. Under 80 words.
2. Touch 2 — value, not nudge. Send something useful (a relevant insight, a teardown, a benchmark) with no ask. 3–5 days later.
3. Touch 3 — direct ask. 'Worth 15 minutes to talk about X?' Honest, no padding. 4–6 days after touch 2.
4. Touch 4 — breakup. 'Sounds like the timing's off — should I close the loop, or check back in Q3?' 5–7 days later.

Four touches. That's it. The 9- and 12-touch sequences of the old playbook are now actively harmful — both to your sender reputation and to your brand.

Multi-channel: email + LinkedIn + SMS

Multi-channel done right beats single-channel done well. Done wrong, it's just spam in three places.

The pattern that works: email is the workhorse, LinkedIn is the warming layer (light, valuable connection requests with no immediate ask), SMS is for late-stage scheduling friction. Each channel reinforces the others without repeating the same message in different formats — which is the lazy way and the worst way.

The review workflow

Every send goes through human approval. Our customers review drafts in batches — typically 50–200 at a time, three times a week. Average review time per draft is 8–14 seconds. Bulk-approve the obvious ones, edit the ones that need work, kill the ones that don't.

This is the single most important design decision. Auto-send is what got the AI cold email category in trouble. Review is what makes the same technique sustainable.

Deliverability fundamentals

• Subdomain isolation. Send from mail.yourdomain.com, not your primary yourdomain.com. Protects the corporate domain.
• SPF, DKIM, DMARC. All three configured, DMARC in 'reject' mode for unauthenticated mail.
• Inbox warming. Two to four weeks before volume, automated low-volume warming through reputable warming networks.
• Volume limits. Per-sender daily caps that ramp gradually. We start at 25/day and ramp to 80–120/day over four weeks.
• List hygiene. Pre-send verification on every address. Bounces over 2% mean stopping and investigating, not pushing through.

Concrete sequence patterns we ship

We've written about specific sequence variants on the blog — including six variants that outperformed our control by 1.8–3.4× this quarter. The full message text is published there. The short version: be shorter, be more specific, be less enthusiastic, ask one thing.