Cost Matrix — Four Business Models

§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.

Data licence (Cotality/RP Data) — irreducible everywhere. The redistribution licence gates Models 1, 2, 4 — not Model 3. That asymmetry is the whole story.
Inference + voice — a pure CapEx↔OpEx swap. The Framework Desktop converts per-minute cloud spend into a fixed box.
Labour — SaaS amortises it across tenants; in-house is labour-heavy; franchise pushes it onto the franchisee.

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.

§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 box	2 · Hosted SaaS	3 · In-house	4 · Franchise
Hardware CapEx	box / ~5 seats	~$0	box (own use)	box/node — franchisee
Setup labour	High (per site)	Low	Low (just you)	Med
Fixed OpEx/mo	~$0 + power	CLOUD_INFRA	CLOUD_INFRA	central clearing only
Data base/mo	API_BASE ⚠	API_BASE ⚠	portal — see §4	API_BASE ⚠
Data per seat/mo	API_SEAT ⚠	API_SEAT ⚠	RPP_SEAT ⚠	API_SEAT ⚠
Inference + voice	~$0 local	per-min unless relay	~$0 local	~$0 local
Sales / closing labour	buyer's	buyer's	HIGH — your closers	franchisee's
Redistribution licence?	if leads leave site	required (worst)	not required ✅	required (existential)
Marginal / +1 seat	step + data	~flat + data	new closer $$	~$0 to you
Marginal / +1 lead	~$0	cents	data + commission	~cents
Revenue capture	box + licence/support	per-seat sub	full commission/deal	rake % + franchise fee
Cost-scaling shape	stepwise	linear	linear in closers	sub-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.

§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.

Revised Model 3 cost line

Line	API path	Portal path
Data UAT one-off	API_UAT ($2.5k)	$0
Data base/mo	API_BASE ($500)	$0
Per seat/mo	API_SEAT ⚠	RPP_SEAT ⚠ (your heads)
Redistribution licence	required at scale	not required ✅
Throughput	high (sanctioned)	capped by posture (§5)
Cash delta (portal saves)	—	~$2.5k once + $500/mo ≈ $6k+/yr

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.

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 ~$6k/yr). But API carries ~zero contingent liability. The honest comparison:

true cost(portal) = cash − ~$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.

Posture	P(discovery/yr)	Severity if caught	E[contingent]/yr	Verdict vs API
CLEAN — L1–L2, no stealth, logged, human-rate	~2%	~$20k (seat pause + rebuild)	~$400	portal wins — save $6k, risk $400
STEALTH-HEAVY — L3–L4 + concealment, in negotiation	~15%	~$300k (licence talks collapse)	~$45k	API 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.

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.

§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.

1 · In-house first (3), portal-only, CLEAN posture. Zero licensing risk; generates the closed-deal signal the price-discovery agent needs.
2 · Productise your own box as on-prem (1). Already built in step 1 — repackaging, not new R&D. First point you need the licence → time it to the negotiation.
3 · Franchise the proven node + clearing layer (4) once the licence lands and you have closed-deal history for real price discovery.
4 · SaaS (2) only if a segment wants cloud and the licence already covers redistribution. Never the lead model.

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.

§9 · Social enrichment

Buy clean. Don't scrape.

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. DIY scraping makes you a miniature Proxycurl; the cheap Apify actors are those farms, one layer removed — provenance-tainted, and able to vanish overnight.

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.

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 $2,400 to acquire, viable only against SMSF lifetime value — and front-loaded, so it's working-capital heavy.

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 runs the data plane on the box, next to the voice + inference stack. The CapEx↔OpEx swap now covers the whole data plane. But the cost saving isn't the headline — an AU-located box means PII never leaves the country. 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.

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); cross-node federation stays bespoke (the clearing-saga atom).

§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.

Mechanics (confirmed direction)

Bulk enrich = clean APIs (ABN/ASIC, Roy Morgan, Apollo/PDL) — not Camoufox.
Camoufox = on-demand, human-triggered, per-lead research + engagement ("research this prospect I'm working now") — not bulk harvest.
Close = scoped. The agent enriches, warms, tees up; a licensed human closes anything that's a financial product (AFSL + anti-hawking s992A line).
Scale = distributed human scale. The franchise is this — many real humans, each amplified, never an account farm.

The differentiator, and it's true: "while everyone else's automation gets shut off."

§8 · Open numbers to close

What makes this decision-grade

RPP_SEAT — actual RP Data portal seat price (biggest swing).
API_SEAT — actual API per-seat entitlement.
RP Data portal ToU clause wording on automated access → resolves CLEAN-vs-STEALTH on real text.
State land-registry codes on owner-data-for-marketing → counsel line for Model 3 outreach.
Headcount horizon (5 / 50 / 500) → sets where on-prem step-cost crosses SaaS linear-cost.