Strategic Briefing · 2026-06-29

Four business models,
one cost spine.

An option chain, not four bets. Who pays each spine cost — and whether it's CapEx or OpEx — is the only thing that really changes between them. Animated explainers unpack the non-obvious dynamics.

currency all figures AUD · order-of-magnitude status decision-grade once RPP_SEAT confirmed canon exchange PARKED · GREEN lane active
safe / GREEN lane / licence-clear caution / access ToU liability / RED neutral structure / data
§2 · The shared spine

Three costs recur everywhere

Every model carries the same three cost atoms. The models differ only in who pays them and when.

Explainer A
The CapEx ↔ OpEx swap
Cloud voice + inference (OpEx) grows with every minute used. The Framework Desktop is a flat upfront cost (CapEx). Past breakeven, the box is strictly cheaper — and it's already built.
$ cumulative usage / minutes → box (CapEx) — fixed cloud (OpEx) — per minute breakeven box wins beyond here
§3 · Master matrix

The four models, side by side

Baseline = "small shop" (~5 seats) scaling up. ⚠ = dominated by an unconfirmed parameter (RPP_SEAT, API_SEAT).

Cost line 1 · On-prem box2 · Hosted SaaS 3 · In-house4 · Franchise
Hardware CapExbox (1–2) → mainboard cluster~A$0box (own use)box/node — franchisee
Setup labourHigh (per site)LowLow (just you)Med
Fixed OpEx/mo~$0 + powerCLOUD_INFRACLOUD_INFRAcentral clearing only
Data base/moAPI_BASE ⚠API_BASE ⚠portal — see §4API_BASE ⚠
Data per seat/moAPI_SEAT ⚠API_SEAT ⚠RPP_SEAT ⚠API_SEAT ⚠
Inference + voice~$0 localper-min unless relay~$0 local~$0 local
Sales / closing labourbuyer'sbuyer'sHIGH — your closersfranchisee's
Redistribution licence?if leads leave siterequired (worst)not required ✅required (existential)
Marginal / +1 seatstep + data~flat + datanew closer $$~$0 to you
Marginal / +1 lead~$0centsdata + commission~cents
Revenue capturebox + licence/supportper-seat subfull commission/dealrake % + franchise fee
Cost-scaling shapestepwiselinearlinear in closerssub-linear
Explainer B
Four scaling shapes
As headcount grows, costs scale differently. In-house climbs fastest (each closer is linear cost); the franchise rides a sub-linear curve because franchisees fund their own CapEx and labour.
your cost headcount / scale → in-house (closers) SaaS (linear) on-prem (steps) franchise (sub-linear)
§4 · Model 3 deep-dive

Portal vs API — and the two gates

For in-house, the Cotality API may be unnecessary: enrich through the RP Data portal on a normal seat. You don't redistribute, so the enrichment method stays a blackbox. This drops the API line entirely — but it only opens one of two gates.

Explainer C
Two gates — only one opens
"Not selling it on" kills the redistribution gate completely. It does not touch the access-ToU gate — how you pull is governed regardless of what you do downstream.
✓ redistribution GONE — no resale access ToU STILL BINDING

Revised Model 3 cost line

LineAPI pathPortal path
UAT setup (testing only)A$2,500 +GSTA$0
Production minimum/moA$500–1,000A$0
Production volume (bulk-process book)⚠ ~A$250k (est., unquoted) for ~896kA$0 — only enrich leads you work
Per seat/moAPI_SEAT ⚠RPP_SEAT ⚠ (your heads)
Redistribution licencerequired at scalenot required ✅
Real saving—not the ~A$6k/yr entry cost — avoiding the ~A$250k bulk bill
Testing ≠ production. The A$2,500 + A$500/mo are testing-stage + minimum (confirmed, Cotality email). Production volume pricing was never quoted — Cotality prices per use-case/volume. The ~A$250k to bulk-process all ~896k is Dom's estimate, unconfirmed. You never pay it: the on-demand, human-triggered path (§13) enriches only the leads you work — that's its dollar justification.
Model-3-specific snag (independent of Cotality): owner-data-for-marketing may be capped by state land-registry codes regardless of contract. Model 3 is the one model that markets to those owners, so this bites hardest here and no Cotality arrangement clears it. Counsel line required before building owner outreach.
§5a · Access posture

Human operator, extended by tools

Chosen posture: HITL assisted automation (Camoufox) — a real human operator extended by tools of graduated agency. Defensibility is governed by one thing: how tightly intent couples to access, per unit of work.

Explainer D
The coupling spectrum (L1 → L4)
ToU cares about access pattern, not whether a human is present. As autonomy climbs, one human intent triggers more access — the coupling stretches and "human present" becomes a figleaf. Defensible zone: L1–L2.
defensible — intent coupled per task L1 1 intent → 1 pull L2 bounded task L3 approve a batch ⚠ L4 autonomous harvest ❌ access per unit of human intent → multiplies
Camoufox carries two separable purposes. Access-enablement — the portal blocks vanilla automation even at human rates, so stealth just lets you function (defensible). Volume-concealment — spoofing to hide a pattern you suspect would be flagged (the liability). Lowers P(caught), raises severity-if-caught.
§5b · The contingent-liability row

A cost you pay only if

Every other matrix row is deterministic cash. This one is priced by expected value. The triggering event is discovery — Cotality detecting automated access and acting on it.

E[contingent cost] = P(discovery) × severity(discovery)

Without this row, portal beats API trivially (saves ~A$6k/yr). But API carries ~zero contingent liability. The honest comparison:

true cost(portal) = cash − ~A$6k/yr saved + E[contingent liability]
Explainer E
Two postures — same cash, inverted answer
Probability ring × severity bar = expected cost. CLEAN stays a rounding error under the $6k saved. STEALTH-HEAVY blows through it — and the API was cheaper all along. Numbers illustrate the mechanic, not estimates.
A$6k/yr saved (threshold) CLEAN ~2% sev ~A$20k A$0 E[]/yr STEALTH-HEAVY ~15% sev ~A$300k A$0 E[]/yr
PostureP(discovery/yr)Severity if caughtE[contingent]/yrVerdict vs API
CLEAN — L1–L2, no stealth, logged, human-rate~2%~A$20k (seat pause + rebuild)~A$400portal wins — save A$6k, risk A$400
STEALTH-HEAVY — L3–L4 + concealment, in negotiation~15%~A$300k (licence talks collapse)~A$45kAPI was far cheaper
Explainer F
Why this risk is nastier than normal
It's correlated with your own upside. The discovery event that kills the seat is the same event that poisons the licence Models 1/2/4 are built on. One bolt, two losses — and it fires precisely when the upside matters most. You can't diversify it away.
DISCOVERY seat terminated enrichment stops licence upside Models 1 / 2 / 4
Recommendation — run CLEAN. Stay L1–L2, use Camoufox only for access-enablement, and log every pull (operator, seat, timestamp, rate, purpose). That log is simultaneously your ToU defensibility and the first link in the provenance chain the exchange needs anyway. A business whose moat is provable consent + licence lineage must not have a deniable acquisition layer at its root.
§6 · Agentic price discovery

Model 4's margin engine

Not a tangent — it's what makes the franchise rake worth more than the on-prem licence. Per-lead pricing costs cents (≈$0 on the local box); its value compounds with network liquidity.

Explainer G
The liquidity flywheel
More nodes → tighter price discovery → more leads clear → bigger rake pool → more nodes. The honest dependency: it needs demand-side liquidity to mean anything. Build it after Model 3 gives the agent closed-deal signal.
more nodes tighter prices more clears bigger rake
§7 · Optimal path

3 → 1 → 4

The same sequence the canon already encodes (exchange PARKED, work the GREEN lane) — so it compounds rather than re-derives. SaaS (2) is conditional distribution only.

Explainer H
The sequence, and where the licence gate sits
Model 3 ships now, portal-only, zero licence dependency. The redistribution licence gate sits between 3 and 1 — the API cost switches on exactly when productisation gives you a reason to go land it.
3 in-house ship now licence 1 on-prem productise 4 franchise rake + flywheel
Act II · The full value chain

The matrix costs the spine. The chain has two more stages.

Property-data was the easy part. The real pipeline bolts social enrichment on the front and social engagement + ads on the back — both riskier and more regulated than the spine. And they form a loop, not a line.

Explainer I
The flywheel — and the narrowing funnel
Enrichment builds audiences → engagement reaches them → responses feed back as data → sharper scoring. The moat is a seed-audience asset. And you don't work every record: cost concentrates on fewer, higher-conviction leads as the funnel narrows.
own moat · 868k enrich (scorable) · ~30k engage (high-conviction) · ~5–10k close responses → data → better next audience
§9 · Social enrichment

Buy clean — when you're selling it on.

This is the CLEAN-lane rule, not a universal one. It applies when data is redistributed (Models 1/2/4). For the in-house lane, the calculus flips — see Cowboy mode (Act III), where scraping is the right call.

For the resale lane: the cheap scrape isn't cheap — its risk is hidden upstream. Proxycurl, the #1 LinkedIn data API at $10M ARR, was sued out of existence by LinkedIn/Microsoft in 2025 for running fake-account farms. If you're carrying lineage to a buyer, the cheap Apify actors are those farms one layer removed — provenance-tainted, able to vanish overnight. In-house, none of that bites (you don't redistribute) — that's the bifurcation.

Explainer J
Two provenance chains
A business whose moat is provable lineage can't afford a tainted link. The bought chain holds. The scraped chain has a link that snaps — and the market's best-funded version of it just disappeared.
BUY CLEAN source provider you ✓ holds SCRAPE source fake farm you ✕ Proxycurl $10M ARR — gone 2025

Best enrichment for the AU / SMSF book is cheap, clean, AU-native: ABN/ASIC business-owner signal (the standout), Roy Morgan/Mosaic, phone & email validation. Buy professional data from PDL/Apollo. Spike-test coverage first — consumer skew may make match rate 10–25%, which could make the whole layer low-ROI (the cheapest answer of all).

§10 · Engagement & the ad funnel

The funnel is gated before it starts

You cannot run finance ads in Australia without an AFSL (or declared exemption) — Meta and Google both verify financial-services advertisers. SMSF/property-investment counts. So the paid engine is gated on holding a licence, being an authorised rep, or a partner (Greentree?) who holds one.

Explainer K
The AFSL gate, then the CAC
No licence, no paid funnel — the gate stays locked. Past it, AU finance is one of the most expensive verticals: stacked up, a closed client costs roughly A$2,400 to acquire, viable only against SMSF lifetime value — and front-loaded, so it's working-capital heavy.
AFSL licence gate CPM ~A$25 · awareness (lookalike off moat) CPC A$10–50 · consideration CPL ~A$120 · lead CAC A$0

The moat is a seed-audience asset: Meta lookalikes want 5–10k quality records; you have 30× that. But uploading it needs a consent basis (hashing ≠ Privacy-Act exemption; the OAIC is watching) → build lookalikes off consented records only.

§11 · Self-hosted Convex

The box now holds the database — and that solves residency

Self-hosted Convex (open-source backend, Rust over SQLite/Postgres) runs the data plane on the box, next to the voice + inference stack. Box cloud cost ≈ A$0 + power — self-hosting zeros the Convex line; Clerk (free ≤10k MAU) is the only remaining cloud tie; the A$1.5–3k/mo applies to the cloud SaaS model only. But the dollar saving isn't the headline — an AU-located box means PII never leaves the country, and the reluctant US-East residency compromise simply evaporates.

Explainer L
Where the data lives
Cloud: PII leaves the country (mitigations carried, counsel-gated). Self-hosted on an AU box: it stays inside. Eliminated, not mitigated.
CLOUD (US-East) your box US cloud SELF-HOST (AU box) 🇦🇺 stays inside

Honest trade-offs: ops + single-point-of-failure → mandatory backups (snapshot → AU R2/NAS); Clerk is the last cloud tie (free tier covers small shops; full air-gap wants local auth later); cross-node federation stays bespoke (the clearing-saga atom).

The hardware: a stable box, volatile alternatives

The Framework Desktop (Ryzen AI Max+ 395, 128GB, ~A$3,300 landed) is the box — chosen because its fixed-config price is insulated from the DRAM spot market. It serves 1–2 people; scale headcount by clustering Framework Mainboards (~A$2,700 each) in a rack.

StageHardware~AUD
1–2 peopleFramework Desktop (stable, repairable appliance)~A$3,300
Scale headcountFramework Mainboards clustered in a rack~A$2,700/board
AlternativeMac Studio — when available (512GB M3 Ultra sold out; M4 Max caps 128GB)A$3,499+
The volatility isn't the Framework price — it's everything around it. The 2025–26 DRAM shock (DDR5 +~400%, Gartner +130% by EOY26) has made alternatives volatile and supply-constrained. The real risks: alternative availability, and Framework's own supply continuity — if it "has its Mac Mini moment" and sells out, we can't source units to scale. Hedge: the stack is hardware-agnostic (self-hosted Convex + Ollama), so the box is a swappable substrate — qualify backup vendors now. Optionality is the hedge, not price-watching.
§13 · Product positioning

The human-scale PI agent

A tool-calling Camoufox agent: your data-enrichment specialist that navigates the web like a human, at human pace, and enriches, engages, and tees up like your best salesman — across a database no human could hold in their head. The moat is operating legitimately while competitors get banned.

"Human pace" ≠ "human." Platforms prohibit automation, not speed; a slow bot is still a bot; an autonomous agent giving financial guidance is still unlicensed advice. The defence is the human in the loop on consequential actions — pace is a consequence of that, not a substitute.
Explainer M
Augmentation, not autonomy
Make one operator do the work of ten — defensible, durable, never banned. Don't replace the operator with ten throttled bots — that's the Proxycurl pattern, slower, and it gets shut off.
AUGMENTATION 1 1 operator × 10 reach SLOW AUTOPILOT BANNED 10 bots → shut off

Mechanics (confirmed direction)

The differentiator, and it's true: "while everyone else's automation gets shut off."
Act III · The cowboy in-house stack

Two thought modes, one firewall

The "buy clean" discipline is a redistribution rule. For the in-house lane — where data stays in the building and we sell our services, not the data — the calculus flips: scrape, Apify, browser automation, all in play. The flip trigger is one question: does the data leave to a third party?

Explainer N
Cowboy vs Clean — the lane firewall
In-house draws from everything (cheap, scrappy, proprietary) and stays internal. Resale draws only from licensed sources. A provenance tag at ingestion is the wall — cowboy data can never cross into anything sold.
🤠 COWBOY · in-house Apifyscrape browser-useFirecrawl internal sell services FIREWALL · provenance tag ✕ 🎩 CLEAN · resale licensed sell data / exchange PDL/Apollo
Two guardrails don't bend even in cowboy: public-vs-authenticated (account-bans don't care about intent → logged-in surfaces stay human-paced) and the AU Privacy Act (public business data = green; personal-individual data still fenced).
§8b · The moat, reconsidered

Static list → living system

Actors don't just enrich — they source, on a schedule, on triggers. "We have 896k records" is a commodity; "we know who became a motivated SMSF-property prospect this week" is the edge. Apify + Convex reactivity turns the moat from a snapshot into a stream.

Explainer O
Snapshot → stream
A bought list decays. A living system grows (continuous sourcing), refreshes (scheduled re-runs), and fires on events (sold a property · new ABN · posted intent). Trigger-based outreach converts multiples higher.
STATIC LIST frozen · decays LIVING SYSTEM source refresh events fire: sold · new ABN · intent
§14 · Architecture

Convex × combinators — mostly already built

Four substrate layers, with stolen-Clay recipes composed on top. The substrate exists: Convex is live, the combinators are shipped, Apify/LLM are on hand. The work is composition — plus one load-bearing artifact.

Explainer P
The layers, and the spine through them
Acquisition · combinators (semantic grammar) · Convex (data+functions) · LLM (judgment) — recipes on top. One result-type cell threads every layer: {value, source, lane, confidence, freshness, cost, provenance}. Get it right and the firewall, waterfall, budget and audit all read off it.
stolen-Clay recipes (the IP) Acquisition — Apify · Firecrawl · browser-use · Bright Data Parser combinators — waterfall = `<|>` · sequence · guard Convex — reactive store · scheduler · actions · audit LLM — judgment node (research · conviction) result cell
Discipline: the architecture fitting means less code — compose 2–3 recipes against the existing substrate, design the result-type contract + the tree-interpreter, and ship. Not a general combinator platform. Tell you've drifted: improving the grammar instead of enriching leads.
§14b · What Apify does for us

Actors across the value chain

Not a scraper — programmatic access to any web surface, as agent tools. In cowboy mode the public-data taps (🟢) scale freely; authenticated/personal (🟡) stays human-paced + privacy-bounded.

Explainer Q
Wire in this order
MCP first (makes everything agent-callable), then sourcing fills the funnel, triggers make it living, intent adds warmth, validation stops wasted spend. All 🟢 except the last.
P0 · Apify MCP — the agent's hands (enabler) P1 · Google Maps — source SMSF-trustee candidates P2 · Real-estate listings — transaction triggers P3 · Reddit / forum — warm intent P4 · Contact verify — validate before spend 🟡
StagePlayLaneValue
SourceQLD/NSW business owners (Google Maps)🟢SMSF-trustee candidates
SourceWho's transacting now (RE portals)🟢property-active prospects (trigger)
SourceForum intent (Reddit / PropertyChat)🟢warm, in-market
EnrichBusiness-site read (Firecrawl on-box)🟢qualification signal
EnrichContact find + verify🟡don't pay for dead contacts
MonitorNew listings · DA approvals · new-ABN🟢event triggers (living system)
AgentMCP: any actor as an on-demand tool—per-lead 360° research
BuildCrawlee custom AU surfaces🟢proprietary data = IP
Start here: P0 (MCP) → P1 (Maps) → P2 (RE triggers) → P3 (intent) — net-new sourcing + the living system + the agent's hands, all in the 🟢 zone, before touching a single authenticated surface. Two force-multipliers: Apify MCP (45k actors as agent tools) and Crawlee (proprietary AU pipelines).
§8 · Open numbers to close

What makes this decision-grade