01-projects/hcls

Mercor lessons, and what selling an HCLS dataset on Snowflake Marketplace actually takes

2026-09-30·research-brief·status: draft

Mercor lessons, and what selling an HCLS dataset on Snowflake takes

Ask (founder, iMessage 2026-09-30 21:37 ET): study Mercor's model for lessons we can repeat; "I would love to monetize a data set on Snowflake."

Bottom line: Raw public CMS data is already a commodity on the Marketplace. Snowflake and small LLCs list it for free. The sellable part is Mercor's: expert judgment packaged as data, sold to teams building AI. Ray's pick, which awaits the founder's read, is a Cortex Knowledge Extension (CKE) over Medicare coverage guidance. Confidence: low. There is no demand signal yet, and the source text is licensed (§4).

1. Mercor

Model. Mercor sells hourly expert labor (doctors, lawyers, PhDs) to AI labs and keeps the spread. Contractors receive 60-70% of gross revenue (The Information via AI Weekly, 2026-07-21). ValueAdd VC puts the take rate at "roughly 35%", within a 30-40% range (2026-07-19, estimate).

Growth:

Lessons a solo builder with a day job can repeat:

  1. Sell the expert layer, not the pipe. The margin is in the rules a domain expert writes, not in parsing CMS files (assumption for HCLS data).
  2. Build controls early. Mercor credits an immutable ledger and observability (blog, 2026-05-08). For us: lineage, freshness checks, a source URL on every row.
  3. Concentration is a risk. A reported ~91% of H1 2026 gross revenue came from two labs. Design for many small buyers.

Not repeatable: VC capital, lab relationships, a network of 30,000 experts, 2025 timing.

2. Snowflake Marketplace mechanics (docs fetched 2026-09-30)

3. HCLS dataset ideas, ranked

Snowflake's free listing already bundles CMS and NPPES data with a 3-month lag (GZTSZ290BV255, 2026-09-30).

# Idea Source data (no PHI; licenses in §4) Expert layer Buyer Marketplace competitors (2026-09-30) Effort (estimate) Kill criterion
1 Medicare coverage CKE Medicare Coverage Database (MCD): local and national coverage determinations (LCDs, NCDs); MLN articles Tags by service, specialty and jurisdiction; effective dates; citations Revenue cycle (RCM) vendors, payers and health systems building Cortex agents (assumption) None among the 17 CKEs 3-5 weekends A code-stripped prototype cannot answer the 10 test questions, or fewer than 3 access requests in 60 days
2 Independent-practice segment map NPPES, PECOS, Care Compare Independent vs system-owned, plus ownership-change signals Vendors selling to practices; PE Dataplex "NPPES Golden Record" (paid) 4-6 weekends Dataplex adds an ownership flag, or buyers won't pay over it
3 Denial-to-fix playbook Denial-code lists, CMS manuals Root cause and fix per code × specialty for independent practices. The founder stepped back from RCM as "the" idea on 09-16 (problem log) RCM vendors, billing companies No playbook found; FinThrive's claims and remits data (free) is adjacent 4-6 weekends X12 license cost exceeds year-one revenue

Re-ranked after the licensing check: the segment map moves up to #2 and the playbook drops to #3. The map's NPPES and PECOS sources carry no code licenses. The playbook depends on code lists that X12 copyrights.

Searched and dropped as crowded: hospital cost reports and 990s; Open Payments × prescribing; price transparency. Each already has 3-6 sellers.

4. Risks and flags (not resolved here)

5. Suggested first step

Before building anything, the founder reads the five clauses. This brief does not judge whether a prototype is covered. If they are clear: build a private CKE prototype over 20 LCDs with code content stripped, and test it on 10 real billing questions. No public post until the agreement and license questions are answered.