Skip to content
Andrew Deighan

work / host-acquisition

01Host acquisition agent pipeline

AtlasOra

11 working agents and rule-based gates that researched every property manager on the Costa del Sol and drafted outreach for a person to approve.

Run this pipeline yourself

1The problem

AtlasOra is a two-sided vacation rental marketplace, so it needed homes before it could serve guests.

The homes on the Costa del Sol sit with property managers: 117 of them. Each one meant finding the decision-maker, researching the company and writing a personal message.

This followed in-person discovery with property managers across Malaga and Marbella.

2What I built

An orchestrated pipeline of 11 working agents and rule-based gates. Two further agents are registered but not yet built out, so they are not counted.

It worked through all 117 property managers: finding decision-makers, researching each company and drafting personalised outreach.

Models are tiered by task: Claude Haiku to sort, Sonnet to research, Opus to write.

Several agents use no model at all: prospect finding, sending, compliance and follow-up templates. That is deliberate.

From prospect to sent message, and who decides
From prospect to sent message, and who decides1. Find prospects (No model): Decision-makers at all 117 managers 2. Research each company (Sonnet) 3. Draft personalised outreach (Opus) 4. Tone check (Rules + model): Rules, then a model judgement 5. If it fails: Failed draft is rewritten: Up to three times 6. Compliance gate (No model): Throttling, GDPR lawful basis, consent Writes an audit record 7. A person approves (Person): Nothing is sent without approval 8. Send (No model) 9. A person handles replies (Person): The reply agent never replies by itselfFind prospectsDecision-makers at all 117 managersno modelResearch each companysonnetDraft personalised outreachopusTone checkRules, then a model judgementrules + modelFailed draft is rewrittenUp to three timesfallbackCompliance gateThrottling, GDPR lawful basis, consentWrites an audit recordno modelA person approvesNothing is sent without approvalpersonSendno modelA person handles repliesThe reply agent never replies by itselfperson
stepa personfailure pathright-hand tag: model or rule used

3How the AI is controlled

It may
  • Find decision-makers, research companies and draft personalised outreach. h2
  • Rewrite a draft that fails the tone check, up to three times. h6
It may not
  • Send anything without a person approving it. h8
  • Reply to a property manager by itself. The agent that handles replies never does. h8
  • Send before the compliance gate has checked throttling, GDPR lawful basis and consent. h7
Who approves
  • A person approves every message before it leaves. h8
When it fails
  • Automatic fallback to a smaller model, with an alert when fallback passes 10% in an hour. h9
  • Retries with backoff, a daily spend cap, and kill switches by agent, channel and region. h9
  • Every send writes an audit record. h7

4The result

Run log10 rows · source: fact file
MeasureValueFact
Properties agreed to listAgreements to list, not live listings~1,000h3
Time to those agreements< 3 weeksh3
Reply rate, email1%h4
Reply rate, WhatsAppAfter one measured channel change25%h4
Property managers worked throughEvery manager on the Costa del Sol117 / 117h2
Managers agreeing to list~20h3
Working agentsTwo more registered, not built out, not counted11h1
Sends without a person approvingnoneh8
Tone-check rewrites per draft, max3h6
Fallback alert threshold> 10% / hourh9

These are agreements to list, not live listings.

Email outreach measured a 1% reply rate. I moved the channel to WhatsApp and replies rose to 25%.