The 47-Step AI Outbound System We Use to Book Meetings

Most “AI outbound” is a ChatGPT prompt bolted onto a bulk sending tool. That setup does not book many meetings. Our system connects 47 steps across data, enrichment, copy, sending, and reply handling. AI handles the per-lead work at each step. Here is the build we run for clients.

The short version

An AI outbound system is a pipeline, not a tool. It moves a raw company list to a booked call through 47 steps grouped into 6 stages:

• Stage 1, Targeting: define the list before you touch a tool. 6 steps.
• Stage 2, Data: find the right companies and people. 8 steps.
• Stage 3, Enrichment: turn rows into reasons to reach out. 11 steps.
• Stage 4, Copy: write per-lead messages that earn a reply. 7 steps.
• Stage 5, Sending: deliver across email and LinkedIn without burning domains. 9 steps.
• Stage 6, Reply and book: handle responses and get the call on the calendar. 6 steps.

No single step makes the system work. The value comes from connecting them so a lead never waits on a human. AI handles enrichment, hooks, copy variants, and the first reply triage. You make the judgment calls.

The 47 steps exist because “paste into ChatGPT and blast” is not a system.

Stage 1: Targeting (steps 1 to 6)

You cannot automate a bad list into a good outcome. This stage requires human judgment and gives you the highest return on your time.

1. Define the offer in one sentence: who, what result, by when.
2. Pick one ICP, not three. One title, one company size band, one trigger.
3. Write the disqualifiers. Who do you not want, and why.
4. Choose the trigger that signals timing: new hire, funding, job posting, tech change.
5. Set the channel mix. Email-led, LinkedIn-led, or both in one sequence.
6. Set the target: meetings per month, working backward to touches per week.

Skip this and the next 41 steps run faster toward nothing.

Stage 2: Data (steps 7 to 14)

Now you build the raw list. Apollo and similar sources give you the rows. The job here is volume with a filter, not volume alone.

7. Pull the company universe from Apollo by size, geo, and industry.
8. Layer the trigger search: hiring, funding, or tech signal.
9. De-dupe against your CRM so you never re-hit a live deal.
10. Suppress past opt-outs across every channel.
11. Pull the right contacts per account, capped to avoid spamming one company.
12. Validate email at the source to cut bounces before they cost you.
13. Resolve LinkedIn URLs with identity checks so you message the right person.
14. Score the list and cut the bottom third before it ever moves forward.

A clean 1,000-row list beats a dirty 10,000-row list every time. We cut hard here.

Stage 3: Enrichment (steps 15 to 25)

Most stacks stop here, and that costs them meetings. Enrichment turns a database row into a reason to contact someone. We run it through our Clay enrichment service, with AI evaluating each lead separately.

15. Pull company description and recent news per account.
16. Pull the prospect’s recent LinkedIn activity, if any.
17. Detect the trigger event and tag it as a structured field.
18. Identify the likely pain tied to that trigger.
19. Find a specific, checkable detail per lead. Not “I see you work at X.”
20. Generate the hook: one sentence that proves you did the work.
21. Score hook quality and flag weak ones for a human pass.
22. Translate the hook to the lead’s language: TR, EN, NL, DE, AR.
23. Localize the angle, not just the words. A Dutch buyer reads differently than a German one.
24. Write the enriched fields back as clean custom fields.
25. Hold any lead with no defensible hook out of the send queue.

Step 25 is the discipline. No hook, no send. A generic message to a great lead still burns the lead.

Stage 4: Copy (steps 26 to 32)

Now the message. AI drafts, your framework constrains it, and you keep variants alive so the system learns.

26. Generate the opener from the enriched hook, per lead.
27. Generate the offer line: the result, not the service.
28. Generate the ask: one low-friction call to action.
29. Build 2 to 3 follow-ups that add a new angle, never “just bumping this.”
30. Produce an A and a B variant per step for testing.
31. Run a spam-trigger and length check before anything queues.
32. Hold high-value named accounts for a hand-written pass.

Use AI for the long tail. Write the priority accounts by hand. Both can run in the same system.

Stage 5: Sending (steps 33 to 41)

Delivery is an infrastructure problem, not a copy problem. The best message in a burned domain books zero calls.

33. Warm domains and inboxes before volume, not during.
34. Cap per-inbox daily volume and rotate across many inboxes.
35. Run email through Sendkit or Smartlead, LinkedIn through GetSales or Heyreach.
36. Keep LinkedIn to 15 to 25 requests per profile per day on a warmed account.
37. Sequence the channels: one touch should know about the other.
38. Throttle by reply signal, not by a fixed daily quota.
39. Monitor bounce and spam rate per inbox, pull any inbox that drifts.
40. Honor suppression and opt-out across every channel in real time.
41. Keep a GDPR-defensible lawful basis and a clean opt-out path on every send.

Add inboxes and accounts when you need more volume. Do not push one inbox harder. That is the most damaging sending mistake we see.

Stage 6: Reply and book (steps 42 to 47)

The system is not done when the message sends. It is done when the call is on the calendar.

42. Triage every reply with AI into interested, question, objection, or noise.
43. Filter the noise: out-of-office, “wrong person,” referrals.
44. Route real positives to a human within minutes, not hours.
45. Draft a reply per category so the human edits instead of writes.
46. Push the booking link or propose times, then confirm the calendar hold.
47. Write the outcome back to the CRM so the next cycle starts smarter.

Agencies often lose good opportunities at the reply stage. Leave a positive reply untouched for a day and it may go nowhere.

What this looks like in tools

The 47 steps run on about 6 tools. The steps do the work. The tools move data between them.

Stage	What it does	Tools we run
Targeting	Define list, ICP, trigger	Manual, no tool
Data	Find companies and people	Apollo, Clay
Enrichment	Turn rows into reasons	Clay plus AI
Copy	Write per-lead messages	AI, human pass on named accounts
Sending	Deliver across channels	Sendkit, Smartlead, GetSales, Heyreach
Reply and book	Triage, route, book	AI triage plus human close

If you measure this system on activity, it looks busy. We measure it on booked meetings. A 40% connection-accept rate that produces no calls is a vanity number, and we do not bill for vanity.

Why 47 steps beats one big prompt

A single AI prompt does not track list quality, suppression rules, domain health, or reply routing. It can write a good message and still send it to the wrong place. In this system, each step is small, checkable, and connected to the next. When reply rate drops, you do not rebuild everything. You find the broken step, such as a stale trigger, a drifting inbox, or a weak hook score, then fix it.

Clients can take ownership for the same reason. The done-with-you outbound model works because each step is documented instead of sitting in one operator’s head. You pay for booked meetings while we run the system. When you are ready, you bring the same 47 steps in-house.

Build the system, not the prompt. The calendar fills when every step does its one job.