06-reference

mostly metrics when to allocate overhead

2026-05-19·reference·source: Mostly Metrics·by CJ Gustafson

Why this is in the vault

CJ's clearest practitioner take on WHEN and WHAT to allocate from G&A out to departments. Not actionable for solo-founder RDCO today, but it's the playbook for when multi-bet attribution (MAC, Squarely, Sanity Check, investing) crosses from "one Ray, one credit card" to "which bet owns which slice of Claude API + compute + tooling." Notable: he carves out "AI Token Costs" as the new Cloud Hosting line with the same Internal/External/Admin split and warns that companies who don't separate them "risk demolishing their GMs down the line."

Sponsorship

Brex is the top-of-newsletter sponsor and CJ is also a customer ("That's why I use Brex"). Body is independent — sponsor block is structurally separate from the analysis. Standard Mostly Metrics shape; no editorial entanglement in the substantive piece. Second-house plug at the bottom for Mostly Talent (CJ's own recruiting business).

The core argument

When to start: Honest answer is "when your investors make you" — Series B/C, when opex-as-% benchmarks start mattering. Not a fundraise-blocker, but you should at least be able to swag it. Gating function: you can delay as long as you can reasonably estimate.

Stack rank of what to allocate, soonest → latest:

  1. Travel — follows the employee, split internal vs external
  2. Departmental software — Salesforce → Sales, Github → Eng, Figma → Product (allocate where usage is; dept head owns renewal)
  3. Corporate software (Zoom, Slack, Notion, 1Password) — IT provisions, allocate by headcount %
  4. Rent — by headcount %, only once you have multiple real offices
  5. IT — hold as long as possible; push out if IT exceeds ~1% of revenue
  6. Cloud hosting — three buckets: Internal (engineers building), External (goes in COGS, exposes true GM), Admin
  7. AI token costs — "the NEW cloud costs," same Internal/External/Admin split; "this line is going to get real chunky real fast"
  8. DevOps — ~50/50 cost-to-serve and R&D; CJ's punt: just park in R&D
  9. Recruiting — direct/dedicated recruiters allocate to function, generalist stays in G&A
  10. SBC — follows the employee; the dreaded investor question
  11. Depreciation — by headcount %; non-cash, "no one wants to be off because of D&A"
  12. CEO time — don't try to split; park in exec/G&A, accept the optics hit

Budget vs allocation discipline: Hold managers to a budget they can directly impact. Allocate for management reporting, but run monthly check-ins on the unallocated (controllable) number. "There's a difference between what you track and what you discuss." Quoted FP&A director allocates only Rent, Office Expenses, IT, Depreciation on ERP schedules and steers conversations toward HC, contractors, T&E, marketing.

Resist over-allocation: Marginal value of allocating the last 7% is ~zero; marginal value of telling a CMO "this is what you own" is huge. A strategic acquirer wants the P&L without allocations — build reporting both ways.

Mapping against Ray Data Co

Forward-looking mapping, weak immediate mapping. Today RDCO is single-founder, single-credit-card; allocation is functionally moot. Three places this becomes load-bearing:

  1. Cross-bet attribution of Claude API + compute spend. MAC, Squarely, Sanity Check, investing all share one Anthropic API key, one Mac Mini, one Cloudflare account — currently pooled in "tools." When any single bet starts generating attributable revenue, CJ's AI-tokens Internal/External/Admin split is the right shape: Internal (Claude time spent building each bet), External (Claude tokens consumed inside a customer-facing product), Admin (the COO agent running back-office). Worth instrumenting at the prompt-tag level before it's load-bearing, per CJ's "comparability" warning about not having to backfill 19 quarters later.

  2. Investing cohort cost attribution. _attribution-ledger.csv already tags research spend → thesis ID → backtest cohort. CJ's "allocate where usage is, dept head owns renewal" is the same shape with s/department/thesis/. Open: do we cleanly tag Claude API spend to a specific thesis when a deep-research run fires?

  3. Acquirer-readiness, eventually. Strategic acquirer wants the P&L without allocations — inverse of why investors want them in. If any bet becomes acquireable, build reporting both ways from the start.

Skip for now: Rent, depreciation, recruiting, SBC, CEO time. Revisit when headcount > 1 or a real second workspace exists.

Related