Sales Dev Agent Spec

How do we think slowly so the agent gets smarter and the outreach gets cleaner each cycle?

Build Contract​

Principles​

What truths constrain how this agent operates?

The Job: When the kill date is 28 days away and there are zero paying customers, help the founder find qualified prospects, score them, compose personalized outreach, and sequence follow-ups — so human time goes to discovery calls, not research and admin.

Hidden objection: "AI outreach feels spammy." Agent composes — human reviews every message before send. Research is automated, relationships are not.

Why now: The product is ready. The demand generation isn't. Nine algorithms, four adapters, a CRM, an RFP tool — all built, none generating demand. The outreach IS the commissioning test.

The convergence: Eddy Whatt at Sneakers Media identified the same pattern from the buy side. Every lead gen provider delivers volume, not quality. Lead gen is a data flow problem dressed as a marketing problem:

OLD: Advertiser → Agency → Lead Gen Provider → Database → Leads (unqualified)
NEW: Advertiser → Agent → Pre-Qualified Leads (with conversation context)

The NZ angle: Small market (5M) makes lead gen quality critical. Relationship-based culture means bad leads damage agency reputation. Proving ground.

Performance​

How do we know it's working?

Priority Score​

Success Criteria​

Functional:

Outcome (Kill: 2026-03-24):

Team-level outcomes (calls booked, pipeline value, first customer) tracked in SPO.

Business Dev​

Platform​

What do we control directly?

Components​

This agent is one instrument in the SPO orchestra.

Build ratio: ~80% composition, ~20% new code.

Commissioning​

Protocols​

How does the system coordinate?

Agent/Human Split​

AGENT (AI-Led + AI-Only)              HUMAN (Human-Led + Human-Only)
────────────────────────              ────────────────────────────
Prospect research                     Discovery calls
Lead scoring + ranking                Relationship building
Outreach drafting                     Message review + send
Follow-up sequencing                  Deal negotiation
Activity logging                      Go/No-Go decisions
Channel optimization                  Trust building

WorkChart​

RESEARCH → SCORE → COMPOSE → SEQUENCE → MEASURE
    ↑                                        |
    └──── Feedback improves next cycle ──────┘

Build Sequence​

Horizon Model​

HORIZON 1: PROVE (now → 90 days)          ← THIS PRD
  Sales Dev Agent — one instrument
        │ feeds
        ▼
HORIZON 2: ORCHESTRATE (90d → 6 months)
  Sales Process Optimisation — the orchestra
        │ generates data for
        ▼
HORIZON 3: PROTOCOL (6m → 18 months)
  Trust Commerce — the protocol

Risks​

Kill signal: 50 messages, 0 replies, 30 days — ICP or message is wrong. Diagnose with explore-exploit data before iterating.

Players​

Who creates harmony?

Job 1: Find Prospects Worth Talking To​

Job 2: Say Something Worth Reading​

Job 3: Follow Up Without Forgetting​

Job 4: Learn What Converts​

Job 5: Pre-Qualify for Media Clients​

ICP: NZ Media Agency​

Archetype. Sneakers Media exemplifies the segment.

Psycho-logic:

The real problem: Not "higher quality leads" — leads their client's sales team doesn't complain about. The complaint loop costs more than the CPL.

Revenue model: Pilot CPL ($20-100/lead) → Platform competing with lead gen providers → Partnership (media + creative + AI/data).

Two Deployment Contexts​

Same WorkChart, different jobs. One instrument, multiple ventures. BOaaS proven.

Context​

• Sales CRM & RFP — The convert/retain half
• Sales Process Optimisation — The orchestra (Horizon 2)
• Trust Commerce — The protocol (Horizon 3)
• The Three Flows — Advertising, data, AI convergence
• Advertising Industry — The convergence thesis in context
• Sales Work Chart — Human/AI split
• Lead Qualification — Pre-qualification method

Questions​

If lead gen is a data flow problem dressed as a marketing problem, what breaks when the data flow is better than the marketing?

• When does the agent's qualifying conversation become the product — not a feature of the product?
• What data from Horizon 1 (this agent) must flow to Horizon 2 (SPO) for the orchestra to tune itself?
• If the NZ proving ground works, does it prove the model or prove the geography?