project_ai_vendor_commercial_model

2026-04-25 REFRAMING — catala-explain as commercial differentiator

Discovered via library+web research (11th session auto-deep-dive win pattern) that github.com/CatalaLang/catala-explain is a production Catala ecosystem module producing .docx explanation documents from Catala execution traces.

Commercial reframing: value proposition shifts from “AI-vendors pay for access to INHERIT schema + InheritKit rules” (prior framing) to “AI-vendors pay for Catala-backed legal reasoning + per-query .docx explanation artifacts that satisfy AI-safety requirements for legal-reasoning transparency” (new framing).

New tiered commercial model:

• Free: raw CatalaResult (JSON values only)
• Pro: CatalaResult + catala-explain .docx per query
• Enterprise: above + custom InheritKit-branded .docx templates + partner-crate jurisdictional-localisation + SLA
• AI-Vendor: above + programmatic .docx streaming API + audit-pack for AI-safety-review + white-labelling (£500k+/year custom terms; matches iText precedent)

Why reframing is structurally stronger:

1. Aligns with AI-safety explainability requirements at OpenAI/Anthropic/Google/Meta/xAI
2. Moat deepens (Catala+catala-explain pipeline harder to replicate than schema docs)
3. Regulatory tailwind (EU AI Act + UK AI White Paper + US EO 14110 all create compliance-driven demand for explainable AI legal reasoning)
4. Partner-crate distribution channel enables jurisdictional .docx localisation
5. Per-query value scales with AI-vendor query volume

NEW Route 4 — Audit-as-a-service: emerges from .docx capability. AI-vendor publishes InheritKit-generated .docx explanations as audit trail; TT charges subscription for audit-pack assembly across queries.

How to apply going forward:

• Phase-C Step 4 Commercial-model spec frames tiers around .docx artifact delivery
• AI-vendor first-contact pitch leads with explainability value, not schema access
• Bruno Lowagie / iText commercial-terms research expanded to cover explanation-document-per-query pricing patterns

See project_catala_explain_as_commercial_artifact.md for full reframing detail.

---

Original framing (pre-2026-04-25; RETAINED FOR CONTEXT)

Rich’s stated view (Friday 17 April 2026)

“I think this commercial model has a very high likelihood of becoming huge for us.”

Context: Rich asked how LLMs / AI agents would access INHERIT and InheritKit in future, and whether MCP is the answer. He explicitly stated he wants to charge OpenAI and Anthropic for InheritKit use. v3.4 foundation-architecture prompt did not address this; v3.5 adds it as a research subject and deliverable.

Three access patterns for LLMs consuming INHERIT + InheritKit

Pattern A — pre-training ingestion. LLMs learn INHERIT from public docs during training. Free marketing, not revenue. Unavoidable.

Pattern B — inference-time reference-data retrieval. LLM calls API / MCP server; returns authoritative current rule values with citations. Addresses hallucination on dated rule values.

Pattern C — computational invocation of InheritKit. LLM delegates faraid / IHT / intestacy math to the authoritative engine rather than reimplementing. Highest InheritKit commercial value.

Three revenue routes

Route 1 — Commercial licence for embedded use. AI vendor wants InheritKit running in their infrastructure → AGPL triggers → commercial licence required. iText precedent. Pricing: probably £500k+/year custom terms. Clean revenue path under existing licence stack.

Route 2 — Hosted metered API service. TT (possibly jointly with Sourcemeta) operates hosted InheritKit API; AI vendor calls per user query. Per-call or per-user-per-month pricing. Scales with LLM-scale traffic. Potentially very large revenue — 100M+ users × £0.001/query × estate-related queries could be material.

Route 3 — Training-data-access charging. Precedents: Reddit $60M/y from Google; StackOverflow; News Corp; Financial Times. Works when content is proprietary. Apache 2.0 foundation forecloses directly — requires bifurcated pattern (Apache spec + proprietary partner-validated content layer) to preserve.

Why this could be large

• LLMs are bad at legal reasoning. They hallucinate tax rates, miss jurisdictional differences, get faraid fractions wrong, can’t compose cross-layer (succession × tax × faith) consistently.
• Delegating to InheritKit is the correct engineering pattern. Deterministic, auditable, reproducible. Partner-validated.
• The commercial pitch writes itself. “LLMs are unreliable at legal reasoning. InheritKit does it correctly. Licence it or hit production legal errors with real consumer consequences.”
• The customer is every hyperscaler building legal / financial / estate-planning AI agents. That’s OpenAI, Anthropic, Google, Meta, xAI — and sector-specialists (Harvey, Pilot Legal, dozens more).
• Defensive moat is real: partner-validated rules (what Sharia Councils, STEP, Law Society endorse), 20-year-evolving reference data accumulation, conformance test track record, institutional endorsements.

Key dependencies

• Licence strategy (v3.4 Phase 1 subject 32) determines Route 3 viability
• Bruno Lowagie research (v3.4 Phase 1 subject 33) extended to cover iText’s actual AI-vendor commercial terms
• MCP / OpenAPI / agent-readiness (v3.2–v3.3 layers) is the technical enabler
• Sourcemeta One auto-MCP (v3.3) might be the hosted-service delivery mechanism for Route 2
• Jentic AI-Readiness Framework scoring (v3.2) validates the interface quality AI vendors would pay for

What the research session must produce

• Recommended revenue-route combination (likely Route 1 + Route 2 primary; Route 3 contingent)
• Pricing models for each route with sensitivity analysis
• First-contact strategy per hyperscaler (OpenAI, Anthropic, Google)
• Competitive positioning (what prevents hyperscalers building their own equivalent)
• Pricing exemplars for specific tiers

Strategic priority

High. Rich’s explicit conviction. The licence stack decision (v3.4) and the MCP / OpenAPI / agent-readiness layer (v3.2–v3.3) both need to compose with this commercial model.

What this means for the v4.0.0 release

INHERIT v4.0.0 + InheritKit v1.0 should ship with AI-vendor commercial engagement ready to start. Specifically:

• Hosted InheritKit MCP server running at a production-grade URL
• OpenAPI 3.1+ spec with Jentic-AI-Readiness-compliant structure
• Rate-limiting + metering infrastructure on the hosted service
• Commercial-licence templates ready for AI-vendor first-contact conversations
• Bruno-Lowagie-informed pricing tiers documented
• First-contact scripts + pitch materials for OpenAI / Anthropic / Google

Without these, “we’ll start charging OpenAI after v4.0.0 ships” becomes “we’ll start charging OpenAI in 2029 after 12 more months of product work” — slipping on the high-conviction revenue thread.