HG Insights|Phoenix

Skill: HG Firmographic

No more 'we ran the numbers on the subsidiary, not the parent' surprises in the deal review.

Overview

Make sure 'research Siemens' returns the 300,000-employee parent, not a 60,000-employee subsidiary. Claude learns to disambiguate parent vs. subsidiary records, hedge low-confidence revenue numbers instead of quoting them as fact, and fall back to name-based search when a domain doesn't resolve cleanly.

Use cases

  • Account briefs sized for the right entity

    Claude detects when a domain resolves to a subsidiary, follows the parent link if the workflow expected parent-level scope, and logs the choice so the AE can audit it. The brief's employee count matches the deal review's expectation.

  • Revenue numbers hedged when HG's confidence is low

    A medium-confidence reading shows up as 'approximately $42M', a low-confidence one as 'mid-market by revenue'. Your CRO never sees a precise-looking number that turns out to be a wide band.

View full skill

HG Firmographic

When to use

  • A workflow takes a domain and needs the canonical HG company record (employees, revenue, industry, geography).
  • A prompt is resolving a name to a domain — fall back from company_firmographic to search_companies for ambiguous matches.
  • An author hits a 31-character hg_id and isn't sure what it means.

Tools you'll touch

  • company_firmographic — direct lookup by domain
  • search_companies — name/keyword search returning candidate matches

What HG actually returns

company_firmographic takes a primary identifier (typically a domain) and returns the canonical HG company record:

  • hg_id — HG's internal company identifier. Important: a 31-character hg_id is a subsidiary identifier; canonical-parent companies have shorter ids. See ADR-007 for the full reasoning.
  • name, domain, employees, revenue, industry (NAICS), geography
  • parent_hg_id — present when this company is a subsidiary; null for canonical parents
  • Employee + revenue confidence bands: HG returns a point estimate plus a confidence label (high/medium/low). Low-confidence numbers should be hedged in prose.

The Phoenix-side credit cost is 1 credit per call (pricing.ts).

Parent/subsidiary handling (ADR-007)

The 31-character hg_id case (issue #1138 root cause):

A workflow that expects "Siemens" data may receive a 31-char hg_id pointing to Siemens Healthineers (subsidiary) rather than Siemens AG (parent). Without disambiguation, downstream tools (technographic, spend, contracts) return the subsidiary's data — a 60,000-employee company instead of a 300,000-employee one — and the workflow output is wrong.

Resolution rule (per ADR-007):

  1. If the hg_id is 31 characters AND the workflow expects parent-level data, follow the parent_hg_id to the canonical record.
  2. If the workflow explicitly wants the subsidiary (rare; usually only when the user named the subsidiary by domain), keep the 31-char id.
  3. Always log which one the workflow chose so audits can trace decisions.

The composition skill hg-subsidiary-handling covers the full disambiguation flow.

Fallback to search_companies

company_firmographic requires a domain. When the input is a name without a clean domain (typed by a user, parsed from prose, or matched against a CRM with stale data), call search_companies first, pick the highest-ranked candidate, then call company_firmographic with its domain.

search_companies is metered dynamically — the static cost is 0 credits, but each company returned costs 1 credit. Cap max_results to 5 unless you actually need a long list.

How to read it

Confidence bands on employees + revenue:

  • High → quote the number directly.
  • Medium → "approximately N employees".
  • Low → use a band: "in the X–Y employee range" or "mid-market by revenue".

NAICS at the right depth. The 6-digit NAICS is precise but often too narrow. NAICS-4 (industry group) is the right grain for peer cohorts; NAICS-2 (sector) is too coarse. See hg-peer-cohort-selection.

Common pitfalls

  1. Trusting a 31-char hg_id for parent-level workflows. This is exactly the #1138 failure mode.
  2. Quoting low-confidence revenue without hedging. "Their revenue is $42.7M" reads like fact; HG may have ±30% uncertainty.
  3. Calling search_companies with default max_results and burning credits. Set the cap.

Citation rules

For domain-resolved company_firmographic results, cite at the source boundary ("HG, May 2026") once at the top of the section. For search_companies candidates, name the search term that produced the match.

Reference