06-reference

mostlymetrics token cost budgeting frontier convergence

2026-09-29·reference·source: Mostly Metrics·by CJ Gustafson
mostly-metricstoken-costsannual-planningagent-cost-attributioncfo-roleharness-thesis

How to Budget for Token Costs Internally — Mostly Metrics

Part 2 of CJ Gustafson's two-part token-cost series (Part 1 audited the outbound AI bill; this one budgets internal seat + usage spend on Claude/OpenAI/Cursor for next year's annual plan). Method: join the Claude Spend report to the HR roster on email, look at spend per head (not totals, which just restate department size), then forecast by role rather than by named person using three escalating methods — $/head average, AI-as-percent-of-comp (healthcare-style benefits loading), and "frontier convergence" (find each role's top-decile spender, treat that as what the role looks like when someone is good at the tools, ramp everyone else up to it over a few quarters with a margin of error). Attributed the frontier-convergence method to Meredith Finn, CFO/COO of Front. Video walkthrough plus downloadable source files and an 8-prompt sequence for running the same analysis on a reader's own data.

Why this is in the vault

A second reusable prompt-driven method (alongside Part 1's bill-audit sequence) for turning raw LLM spend data into a forecastable internal budget line — and a clean example of CFO-facing content treating AI seats as a comp-like cost category, directly on-thesis for RDCO's credibility-building lane.

Mapping against Ray Data Co

The founder's own disclosed constraint — a $100/mo Claude token budget at phData with a ~$200 hard cap, "nobody gets more" (surfaced in [[2026-06-01-jaya-gupta-token-budget-wars]]) — is exactly the per-head allocation problem CJ is diagramming from the other side of the table: a company decided phData's $/head number without ever running a frontier-convergence pass, so the cap likely reflects an average-user assumption rather than what a frontier user (someone dispatching Fable at high/xhigh per the delegation-model-effort-pairing memory) actually needs. More directly, this issue extends Part 1's audit method ([[2026-09-01-mostlymetrics-audit-ai-bill-for-savings]]) from "where is spend leaking" to "how do we forecast it going into next year" — and RDCO still has neither report to run this against, since Max-subscription flat pricing hides the per-call number (per [[feedback_api_cost_budget_controlled]]). The strategic read is less "build this for RDCO" and more "bank the framework for phData sales credibility": per the founder's 2026-09-27 direction that credibility work should target agents-in-production domain expertise (not subs), a CFO-facing, doc-verifiable explainer of how a finance org should budget AI-tool seats is a legible artifact type in that lane — CJ has already built the audience-prep (the "frontier convergence" vocabulary, the per-head-not-total framing) that a phData-credibility piece on agent deployment economics could cite or riff against rather than re-derive.

The core argument

Three methods, escalating in rigor: (1) $/head average — clean top-down sanity check, useless for allocation because AI usage is lumpy and a few heavy users drag the mean; (2) AI-as-percent-of-comp — good board communication ("we spend 20% of comp per employee on tokens") but not an operator-usable lever; (3) frontier convergence — find each role's top-decile (not single-max, to avoid one outlier setting the ceiling) monthly spend, treat that as the target, and ramp laggards toward it over multiple quarters rather than assuming instant adoption. CJ converts this into a model: enter next year's headcount by role, price each head at the frontier, apply an adoption ramp, add a margin of error, get monthly/annual totals. Explicit caveats: single-person roles (e.g., a company's one CFO) show no internal gap and need cross-role judgment; a reader is almost certainly only budgeting one vendor of several (Claude report only, when most companies run 2+ model vendors per Brex's own cross-platform data cited in the piece); don't claw back "silent" unused seats without first checking whether the person was ever shown how to use the tool.

⚠️ Sponsorship

Brex-sponsored issue — standing recurring sponsor for this newsletter (confirmed sponsor of the 2026-08-04 CFO-token-cost-gate and 2026-09-01 Part 1 issues from this same sender). Opening promo block pitches Brex corporate cards + AI-native expense workflows (brex.com/metrics partner link). Brex's own cross-platform AI-spend data (median 2-3 AI tools paid per company per month, 61% paying 2+ model vendors) is cited as supporting evidence inside the tutorial itself — the sponsor is also the data source, which is a tighter coupling than a standard ad block; CJ's argument doesn't depend on Brex's specific numbers holding up, but the framing that "AI tool sprawl is now measurable and rising" serves Brex's pitch (buy Brex to see your own sprawl) as much as it serves the budgeting thesis.

Related