AI-enabled sales system
Scaling follow-up without scaling the setter team.
The business had a lead-volume problem disguised as a sales problem.
Leads arrived through marketplaces, paid social, websites, inbound calls and social channels throughout the day. A four-person offshore setter team worked a queue during standard US hours, which meant some leads waited hours, occasionally until the next day, before first contact.
Follow-up was inconsistent, CRM hygiene was poor, and there was no reliable way to ensure every lead received the same persistence or quality of handling.
What changed
I took ownership of the technical system and progressively rebuilt the lead journey around CRM orchestration and AI voice.
The final system connected 10 intake channels across 16 platforms, using GoHighLevel as the state layer, Make as the integration bus and Vapi for voice.
New leads could receive near-immediate contact, often preceded by SMS, then move through stage-specific follow-up sequences until they either booked, declined or reached another defined outcome.
AI handled the setting layer. Humans took over where the relationship became more valuable: the closer attended the booked video meeting and carried the prospect through the sale.
- 10 intake channels
- Normalisation / dedupe Rules
- GoHighLevel pipeline Rules
-
SMS/email RulesAI calling AIDirect booking Rules
- Stage-specific cadence Rules
- Availability / knowledge tools AI
- Appointment written Rules
- Closer assigned Rules
- Reminders / no-show recovery Rules
- Human sales process Human
At scale
- AI calls in a peak month
- ~50,000
- autonomously booked meetings per month
- ~400
- in sustained cash collection supported
- ~$1M/month
The system supported the business as marketing increased lead volume and the company scaled from roughly $200k/month to around $1M/month in cash collected.
The important change was not simply replacing phone calls with AI. It was creating a lead-management system where every prospect could receive consistent, persistent follow-up regardless of volume.
The challenge was not the AI call itself, but everything around it: state, booking, failures, documentation and human handoff. The system also reinforced that high-volume automation needs continuous monitoring and iteration, not a one-time build.
Read the full case studyLink acquisition operating system
Rebuilding a deteriorating manual operation from the process up.
Before the platform existed, the operation ran through two SOPs, Airtable, Pitchbox, four mailboxes and a six-person team.
The underlying business was healthy, but the operating model was deteriorating. Delivery speed had fallen quarter after quarter, and by launch there were 675 open overdue orders. Average delivery time had reached roughly 80–88 days.
The objective was explicit:
Fix delivery, increase throughput, reduce tooling and staffing overhead, and improve quality.
Before any code
A week was spent understanding and specifying the system before any building began.
That meant
- Reviewing the existing SOPs
- Interviewing the manager running the operation
- Mapping every stage and decision
- Defining the database structure
- Validating data sources
- Estimating API capacity and cost
- Deciding what should be deterministic, AI-assisted or human-reviewed
The operating model
The finished platform handles most of the workflow:
source vet contact interpret negotiate match approve deliver monitor
Deterministic rules make threshold and pricing decisions. Language models are used where interpretation is useful. Humans remain at financial commitments, ambiguous edge cases and quality gates.
That separation was deliberate: the LLM can read and classify words, but it does not invent prices or make uncontrolled business decisions.
- Source domain
-
- Free checks Rules
- Cheap data checks Rules
- Content / readability check AI
- Higher-cost quality checks Rules
Spend only on candidates that survive cheaper checks.
- Verified contact Rules
- Cold outreach Rules
- Reply classification AI
- Deterministic negotiation Rules
- Won publisher database Rules
- Order matching Rules
- Human approval Human
- Live link
- Monitoring Rules
The result
- domains assessed
- 25,883
- quality decisions recorded
- 163,566
- made without a person
- 87%
- links live in the first four months
- 915
- of inherited backlog delivered
- 75%
Before
6 people
- Manager Human
- Outreach Human
- 4 × fulfilment Human
- Airtable + Pitchbox + inboxes
2.5% live within 30 days
After
1 full-time operator + part-time manager
System performs
- Sourcing
- Vetting
- Outreach
- Negotiation
- Matching
- Tracking
Humans
- Approve money
- Resolve edge cases
- Manage payments
28.5% live within 30 days
The project reinforced three things: understand the process before automating it, design the economics into the architecture, and treat production feedback as part of the build rather than an exception to it.
Read the full case study